diff --git a/configs/det/det_mv3_db.yml b/configs/det/det_mv3_db.yml index b55ff99be91bfea795d4b023edcc57e95b97ecc2..640f3a205b1fd3ec7fe19d5c6b6e3aef9ddf3968 100644 --- a/configs/det/det_mv3_db.yml +++ b/configs/det/det_mv3_db.yml @@ -45,9 +45,7 @@ Optimizer: beta1: 0.9 beta2: 0.999 lr: -# name: Cosine learning_rate: 0.001 -# warmup_epoch: 0 regularizer: name: 'L2' factor: 0 diff --git a/configs/det/det_mv3_east.yml b/configs/det/det_mv3_east.yml new file mode 100644 index 0000000000000000000000000000000000000000..05581a761cc91abce6d1e83bcdd7aacef1349148 --- /dev/null +++ b/configs/det/det_mv3_east.yml @@ -0,0 +1,111 @@ +Global: + use_gpu: true + epoch_num: 10000 + log_smooth_window: 20 + print_batch_step: 2 + save_model_dir: ./output/east_mv3/ + save_epoch_step: 1000 + # evaluation is run every 5000 iterations after the 4000th iteration + eval_batch_step: [4000, 5000] + # if pretrained_model is saved in static mode, load_static_weights must set to True + load_static_weights: True + cal_metric_during_train: False + pretrained_model: ./pretrain_models/MobileNetV3_large_x0_5_pretrained + checkpoints: + save_inference_dir: + use_visualdl: False + infer_img: + save_res_path: ./output/det_east/predicts_east.txt + +Architecture: + model_type: det + algorithm: EAST + Transform: + Backbone: + name: MobileNetV3 + scale: 0.5 + model_name: large + Neck: + name: EASTFPN + model_name: small + Head: + name: EASTHead + model_name: small + +Loss: + name: EASTLoss + +Optimizer: + name: Adam + beta1: 0.9 + beta2: 0.999 + lr: + # name: Cosine + learning_rate: 0.001 + # warmup_epoch: 0 + regularizer: + name: 'L2' + factor: 0 + +PostProcess: + name: EASTPostProcess + score_thresh: 0.8 + cover_thresh: 0.1 + nms_thresh: 0.2 + +Metric: + name: DetMetric + main_indicator: hmean + +Train: + dataset: + name: SimpleDataSet + data_dir: ./train_data/icdar2015/text_localization/ + label_file_list: + - ./train_data/icdar2015/text_localization/train_icdar2015_label.txt + ratio_list: [1.0] + transforms: + - DecodeImage: # load image + img_mode: BGR + channel_first: False + - DetLabelEncode: # Class handling label + - EASTProcessTrain: + image_shape: [512, 512] + background_ratio: 0.125 + min_crop_side_ratio: 0.1 + min_text_size: 10 + - KeepKeys: + keep_keys: ['image', 'score_map', 'geo_map', 'training_mask'] # dataloader will return list in this order + loader: + shuffle: True + drop_last: False + batch_size_per_card: 16 + num_workers: 8 + +Eval: + dataset: + name: SimpleDataSet + data_dir: ./train_data/icdar2015/text_localization/ + label_file_list: + - ./train_data/icdar2015/text_localization/test_icdar2015_label.txt + transforms: + - DecodeImage: # load image + img_mode: BGR + channel_first: False + - DetLabelEncode: # Class handling label + - DetResizeForTest: + limit_side_len: 2400 + limit_type: max + - NormalizeImage: + scale: 1./255. + mean: [0.485, 0.456, 0.406] + std: [0.229, 0.224, 0.225] + order: 'hwc' + - ToCHWImage: + - KeepKeys: + keep_keys: ['image', 'shape', 'polys', 'ignore_tags'] + loader: + shuffle: False + drop_last: False + batch_size_per_card: 1 # must be 1 + num_workers: 2 \ No newline at end of file diff --git a/configs/det/det_r50_vd_db.yml b/configs/det/det_r50_vd_db.yml new file mode 100644 index 0000000000000000000000000000000000000000..491983f57a59f0a4105712743a69a56c8212a0e9 --- /dev/null +++ b/configs/det/det_r50_vd_db.yml @@ -0,0 +1,130 @@ +Global: + use_gpu: true + epoch_num: 1200 + log_smooth_window: 20 + print_batch_step: 10 + save_model_dir: ./output/det_rc/det_r50_vd/ + save_epoch_step: 1200 + # evaluation is run every 5000 iterations after the 4000th iteration + eval_batch_step: [5000,4000] + # if pretrained_model is saved in static mode, load_static_weights must set to True + load_static_weights: True + cal_metric_during_train: False + pretrained_model: ./pretrain_models/ResNet50_vd_ssld_pretrained + checkpoints: + save_inference_dir: + use_visualdl: False + infer_img: doc/imgs_en/img_10.jpg + save_res_path: ./output/det_db/predicts_db.txt + +Architecture: + model_type: det + algorithm: DB + Transform: + Backbone: + name: ResNet + layers: 50 + Neck: + name: DBFPN + out_channels: 256 + Head: + name: DBHead + k: 50 + +Loss: + name: DBLoss + balance_loss: true + main_loss_type: DiceLoss + alpha: 5 + beta: 10 + ohem_ratio: 3 + +Optimizer: + name: Adam + beta1: 0.9 + beta2: 0.999 + lr: + learning_rate: 0.001 + regularizer: + name: 'L2' + factor: 0 + +PostProcess: + name: DBPostProcess + thresh: 0.3 + box_thresh: 0.7 + max_candidates: 1000 + unclip_ratio: 1.5 + +Metric: + name: DetMetric + main_indicator: hmean + +Train: + dataset: + name: SimpleDataSet + data_dir: ./train_data/icdar2015/text_localization/ + label_file_list: + - ./train_data/icdar2015/text_localization/train_icdar2015_label.txt + ratio_list: [0.5] + transforms: + - DecodeImage: # load image + img_mode: BGR + channel_first: False + - DetLabelEncode: # Class handling label + - IaaAugment: + augmenter_args: + - { 'type': Fliplr, 'args': { 'p': 0.5 } } + - { 'type': Affine, 'args': { 'rotate': [-10, 10] } } + - { 'type': Resize, 'args': { 'size': [0.5, 3] } } + - EastRandomCropData: + size: [640, 640] + max_tries: 50 + keep_ratio: true + - MakeBorderMap: + shrink_ratio: 0.4 + thresh_min: 0.3 + thresh_max: 0.7 + - MakeShrinkMap: + shrink_ratio: 0.4 + min_text_size: 8 + - NormalizeImage: + scale: 1./255. + mean: [0.485, 0.456, 0.406] + std: [0.229, 0.224, 0.225] + order: 'hwc' + - ToCHWImage: + - KeepKeys: + keep_keys: ['image', 'threshold_map', 'threshold_mask', 'shrink_map', 'shrink_mask'] # the order of the dataloader list + loader: + shuffle: True + drop_last: False + batch_size_per_card: 16 + num_workers: 8 + +Eval: + dataset: + name: SimpleDataSet + data_dir: ./train_data/icdar2015/text_localization/ + label_file_list: + - ./train_data/icdar2015/text_localization/test_icdar2015_label.txt + transforms: + - DecodeImage: # load image + img_mode: BGR + channel_first: False + - DetLabelEncode: # Class handling label + - DetResizeForTest: + image_shape: [736, 1280] + - NormalizeImage: + scale: 1./255. + mean: [0.485, 0.456, 0.406] + std: [0.229, 0.224, 0.225] + order: 'hwc' + - ToCHWImage: + - KeepKeys: + keep_keys: ['image', 'shape', 'polys', 'ignore_tags'] + loader: + shuffle: False + drop_last: False + batch_size_per_card: 1 # must be 1 + num_workers: 8 \ No newline at end of file diff --git a/configs/det/det_r50_vd_east.yml b/configs/det/det_r50_vd_east.yml new file mode 100644 index 0000000000000000000000000000000000000000..b8fe55d4ac5473f6c0392820f5b1f651448bdddd --- /dev/null +++ b/configs/det/det_r50_vd_east.yml @@ -0,0 +1,110 @@ +Global: + use_gpu: true + epoch_num: 10000 + log_smooth_window: 20 + print_batch_step: 2 + save_model_dir: ./output/east_r50_vd/ + save_epoch_step: 1000 + # evaluation is run every 5000 iterations after the 4000th iteration + eval_batch_step: [4000, 5000] + # if pretrained_model is saved in static mode, load_static_weights must set to True + load_static_weights: True + cal_metric_during_train: False + pretrained_model: ./pretrain_models/ResNet50_vd_pretrained/ + checkpoints: + save_inference_dir: + use_visualdl: False + infer_img: + save_res_path: ./output/det_east/predicts_east.txt + +Architecture: + model_type: det + algorithm: EAST + Transform: + Backbone: + name: ResNet + layers: 50 + Neck: + name: EASTFPN + model_name: large + Head: + name: EASTHead + model_name: large + +Loss: + name: EASTLoss + +Optimizer: + name: Adam + beta1: 0.9 + beta2: 0.999 + lr: + # name: Cosine + learning_rate: 0.001 + # warmup_epoch: 0 + regularizer: + name: 'L2' + factor: 0 + +PostProcess: + name: EASTPostProcess + score_thresh: 0.8 + cover_thresh: 0.1 + nms_thresh: 0.2 + +Metric: + name: DetMetric + main_indicator: hmean + +Train: + dataset: + name: SimpleDataSet + data_dir: ./train_data/icdar2015/text_localization/ + label_file_list: + - ./train_data/icdar2015/text_localization/train_icdar2015_label.txt + ratio_list: [1.0] + transforms: + - DecodeImage: # load image + img_mode: BGR + channel_first: False + - DetLabelEncode: # Class handling label + - EASTProcessTrain: + image_shape: [512, 512] + background_ratio: 0.125 + min_crop_side_ratio: 0.1 + min_text_size: 10 + - KeepKeys: + keep_keys: ['image', 'score_map', 'geo_map', 'training_mask'] # dataloader will return list in this order + loader: + shuffle: True + drop_last: False + batch_size_per_card: 8 + num_workers: 8 + +Eval: + dataset: + name: SimpleDataSet + data_dir: ./train_data/icdar2015/text_localization/ + label_file_list: + - ./train_data/icdar2015/text_localization/test_icdar2015_label.txt + transforms: + - DecodeImage: # load image + img_mode: BGR + channel_first: False + - DetLabelEncode: # Class handling label + - DetResizeForTest: + limit_side_len: 2400 + limit_type: max + - NormalizeImage: + scale: 1./255. + mean: [0.485, 0.456, 0.406] + std: [0.229, 0.224, 0.225] + order: 'hwc' + - ToCHWImage: + - KeepKeys: + keep_keys: ['image', 'shape', 'polys', 'ignore_tags'] + loader: + shuffle: False + drop_last: False + batch_size_per_card: 1 # must be 1 + num_workers: 2 \ No newline at end of file diff --git a/configs/det/det_r50_vd_sast_icdar15.yml b/configs/det/det_r50_vd_sast_icdar15.yml new file mode 100644 index 0000000000000000000000000000000000000000..7ca93cecec3dbf7152c1d509c4d8ca614dec388f --- /dev/null +++ b/configs/det/det_r50_vd_sast_icdar15.yml @@ -0,0 +1,110 @@ +Global: + use_gpu: true + epoch_num: 5000 + log_smooth_window: 20 + print_batch_step: 2 + save_model_dir: ./output/sast_r50_vd_ic15/ + save_epoch_step: 1000 + # evaluation is run every 5000 iterations after the 4000th iteration + eval_batch_step: [4000, 5000] + # if pretrained_model is saved in static mode, load_static_weights must set to True + load_static_weights: True + cal_metric_during_train: False + pretrained_model: ./pretrain_models/ResNet50_vd_ssld_pretrained/ + checkpoints: + save_inference_dir: + use_visualdl: False + infer_img: + save_res_path: ./output/sast_r50_vd_ic15/predicts_sast.txt + +Architecture: + model_type: det + algorithm: SAST + Transform: + Backbone: + name: ResNet_SAST + layers: 50 + Neck: + name: SASTFPN + with_cab: True + Head: + name: SASTHead + +Loss: + name: SASTLoss + +Optimizer: + name: Adam + beta1: 0.9 + beta2: 0.999 + lr: + # name: Cosine + learning_rate: 0.001 + # warmup_epoch: 0 + regularizer: + name: 'L2' + factor: 0 + +PostProcess: + name: SASTPostProcess + score_thresh: 0.5 + sample_pts_num: 2 + nms_thresh: 0.2 + expand_scale: 1.0 + shrink_ratio_of_width: 0.3 + +Metric: + name: DetMetric + main_indicator: hmean + +Train: + dataset: + name: SimpleDataSet + data_dir: ./train_data/ + label_file_path: [./train_data/art_latin_icdar_14pt/train_no_tt_test/train_label_json.txt, ./train_data/total_text_icdar_14pt/train_label_json.txt] + data_ratio_list: [0.5, 0.5] + transforms: + - DecodeImage: # load image + img_mode: BGR + channel_first: False + - DetLabelEncode: # Class handling label + - SASTProcessTrain: + image_shape: [512, 512] + min_crop_side_ratio: 0.3 + min_crop_size: 24 + min_text_size: 4 + max_text_size: 512 + - KeepKeys: + keep_keys: ['image', 'score_map', 'border_map', 'training_mask', 'tvo_map', 'tco_map'] # dataloader will return list in this order + loader: + shuffle: True + drop_last: False + batch_size_per_card: 4 + num_workers: 4 + +Eval: + dataset: + name: SimpleDataSet + data_dir: ./train_data/icdar2015/text_localization/ + label_file_list: + - ./train_data/icdar2015/text_localization/test_icdar2015_label.txt + transforms: + - DecodeImage: # load image + img_mode: BGR + channel_first: False + - DetLabelEncode: # Class handling label + - DetResizeForTest: + resize_long: 1536 + - NormalizeImage: + scale: 1./255. + mean: [0.485, 0.456, 0.406] + std: [0.229, 0.224, 0.225] + order: 'hwc' + - ToCHWImage: + - KeepKeys: + keep_keys: ['image', 'shape', 'polys', 'ignore_tags'] + loader: + shuffle: False + drop_last: False + batch_size_per_card: 1 # must be 1 + num_workers: 2 \ No newline at end of file diff --git a/configs/det/det_r50_vd_sast_totaltext.yml b/configs/det/det_r50_vd_sast_totaltext.yml new file mode 100644 index 0000000000000000000000000000000000000000..a9a037c8bd940d04d08d0dae3a90139ead68cdba --- /dev/null +++ b/configs/det/det_r50_vd_sast_totaltext.yml @@ -0,0 +1,109 @@ +Global: + use_gpu: true + epoch_num: 5000 + log_smooth_window: 20 + print_batch_step: 2 + save_model_dir: ./output/sast_r50_vd_tt/ + save_epoch_step: 1000 + # evaluation is run every 5000 iterations after the 4000th iteration + eval_batch_step: [4000, 5000] + # if pretrained_model is saved in static mode, load_static_weights must set to True + load_static_weights: True + cal_metric_during_train: False + pretrained_model: ./pretrain_models/ResNet50_vd_ssld_pretrained/ + checkpoints: + save_inference_dir: + use_visualdl: False + infer_img: + save_res_path: ./output/sast_r50_vd_tt/predicts_sast.txt + +Architecture: + model_type: det + algorithm: SAST + Transform: + Backbone: + name: ResNet_SAST + layers: 50 + Neck: + name: SASTFPN + with_cab: True + Head: + name: SASTHead + +Loss: + name: SASTLoss + +Optimizer: + name: Adam + beta1: 0.9 + beta2: 0.999 + lr: + # name: Cosine + learning_rate: 0.001 + # warmup_epoch: 0 + regularizer: + name: 'L2' + factor: 0 + +PostProcess: + name: SASTPostProcess + score_thresh: 0.5 + sample_pts_num: 6 + nms_thresh: 0.2 + expand_scale: 1.2 + shrink_ratio_of_width: 0.2 + +Metric: + name: DetMetric + main_indicator: hmean + +Train: + dataset: + name: SimpleDataSet + label_file_list: [./train_data/icdar2013/train_label_json.txt, ./train_data/icdar2015/train_label_json.txt, ./train_data/icdar17_mlt_latin/train_label_json.txt, ./train_data/coco_text_icdar_4pts/train_label_json.txt] + ratio_list: [0.1, 0.45, 0.3, 0.15] + transforms: + - DecodeImage: # load image + img_mode: BGR + channel_first: False + - DetLabelEncode: # Class handling label + - SASTProcessTrain: + image_shape: [512, 512] + min_crop_side_ratio: 0.3 + min_crop_size: 24 + min_text_size: 4 + max_text_size: 512 + - KeepKeys: + keep_keys: ['image', 'score_map', 'border_map', 'training_mask', 'tvo_map', 'tco_map'] # dataloader will return list in this order + loader: + shuffle: True + drop_last: False + batch_size_per_card: 4 + num_workers: 4 + +Eval: + dataset: + name: SimpleDataSet + data_dir: ./train_data/ + label_file_list: + - ./train_data/total_text_icdar_14pt/test_label_json.txt + transforms: + - DecodeImage: # load image + img_mode: BGR + channel_first: False + - DetLabelEncode: # Class handling label + - DetResizeForTest: + resize_long: 768 + - NormalizeImage: + scale: 1./255. + mean: [0.485, 0.456, 0.406] + std: [0.229, 0.224, 0.225] + order: 'hwc' + - ToCHWImage: + - KeepKeys: + keep_keys: ['image', 'shape', 'polys', 'ignore_tags'] + loader: + shuffle: False + drop_last: False + batch_size_per_card: 1 # must be 1 + num_workers: 2 \ No newline at end of file diff --git a/configs/rec/rec_mv3_none_bilstm_ctc.yml b/configs/rec/rec_mv3_none_bilstm_ctc.yml index def72375142ccf9f0988c0821d041b837442f3d0..38f1e8691e6056ada01a2d5c19f70955e8117498 100644 --- a/configs/rec/rec_mv3_none_bilstm_ctc.yml +++ b/configs/rec/rec_mv3_none_bilstm_ctc.yml @@ -5,7 +5,7 @@ Global: print_batch_step: 10 save_model_dir: ./output/rec/mv3_none_bilstm_ctc/ save_epoch_step: 3 - # evaluation is run every 5000 iterations after the 4000th iteration + # evaluation is run every 2000 iterations eval_batch_step: [0, 2000] # if pretrained_model is saved in static mode, load_static_weights must set to True cal_metric_during_train: True @@ -13,7 +13,7 @@ Global: checkpoints: save_inference_dir: use_visualdl: False - infer_img: doc/imgs_words/ch/word_1.jpg + infer_img: doc/imgs_words_en/word_10.png # for data or label process character_dict_path: character_type: en @@ -21,7 +21,6 @@ Global: infer_mode: False use_space_char: False - Optimizer: name: Adam beta1: 0.9 diff --git a/configs/rec/rec_mv3_none_none_ctc.yml b/configs/rec/rec_mv3_none_none_ctc.yml new file mode 100644 index 0000000000000000000000000000000000000000..33079ad48c94c217ef86ef3f245492a540559350 --- /dev/null +++ b/configs/rec/rec_mv3_none_none_ctc.yml @@ -0,0 +1,95 @@ +Global: + use_gpu: True + epoch_num: 72 + log_smooth_window: 20 + print_batch_step: 10 + save_model_dir: ./output/rec/mv3_none_none_ctc/ + save_epoch_step: 3 + # evaluation is run every 2000 iterations + eval_batch_step: [0, 2000] + # if pretrained_model is saved in static mode, load_static_weights must set to True + cal_metric_during_train: True + pretrained_model: + checkpoints: + save_inference_dir: + use_visualdl: False + infer_img: doc/imgs_words_en/word_10.png + # for data or label process + character_dict_path: + character_type: en + max_text_length: 25 + infer_mode: False + use_space_char: False + +Optimizer: + name: Adam + beta1: 0.9 + beta2: 0.999 + lr: + learning_rate: 0.0005 + regularizer: + name: 'L2' + factor: 0 + +Architecture: + model_type: rec + algorithm: Rosetta + Transform: + Backbone: + name: MobileNetV3 + scale: 0.5 + model_name: large + Neck: + name: SequenceEncoder + encoder_type: reshape + Head: + name: CTCHead + fc_decay: 0.0004 + +Loss: + name: CTCLoss + +PostProcess: + name: CTCLabelDecode + +Metric: + name: RecMetric + main_indicator: acc + +Train: + dataset: + name: LMDBDateSet + data_dir: ./train_data/data_lmdb_release/training/ + transforms: + - DecodeImage: # load image + img_mode: BGR + channel_first: False + - CTCLabelEncode: # Class handling label + - RecResizeImg: + image_shape: [3, 32, 100] + - KeepKeys: + keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order + loader: + shuffle: False + batch_size_per_card: 256 + drop_last: True + num_workers: 8 + +Eval: + dataset: + name: LMDBDateSet + data_dir: ./train_data/data_lmdb_release/validation/ + transforms: + - DecodeImage: # load image + img_mode: BGR + channel_first: False + - CTCLabelEncode: # Class handling label + - RecResizeImg: + image_shape: [3, 32, 100] + - KeepKeys: + keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order + loader: + shuffle: False + drop_last: False + batch_size_per_card: 256 + num_workers: 8 diff --git a/configs/rec/rec_mv3_tps_bilstm_ctc.yml b/configs/rec/rec_mv3_tps_bilstm_ctc.yml new file mode 100644 index 0000000000000000000000000000000000000000..08f68939d4f1e6de1c3688652bd86f6556a43384 --- /dev/null +++ b/configs/rec/rec_mv3_tps_bilstm_ctc.yml @@ -0,0 +1,100 @@ +Global: + use_gpu: true + epoch_num: 72 + log_smooth_window: 20 + print_batch_step: 10 + save_model_dir: ./output/rec/mv3_tps_bilstm_ctc/ + save_epoch_step: 3 + # evaluation is run every 2000 iterations + eval_batch_step: [0, 2000] + # if pretrained_model is saved in static mode, load_static_weights must set to True + cal_metric_during_train: True + pretrained_model: + checkpoints: + save_inference_dir: + use_visualdl: False + infer_img: doc/imgs_words_en/word_10.png + # for data or label process + character_dict_path: + character_type: en + max_text_length: 25 + infer_mode: False + use_space_char: False + +Optimizer: + name: Adam + beta1: 0.9 + beta2: 0.999 + lr: + learning_rate: 0.0005 + regularizer: + name: 'L2' + factor: 0 + +Architecture: + model_type: rec + algorithm: STARNet + Transform: + name: TPS + num_fiducial: 20 + loc_lr: 0.1 + model_name: small + Backbone: + name: MobileNetV3 + scale: 0.5 + model_name: large + Neck: + name: SequenceEncoder + encoder_type: rnn + hidden_size: 96 + Head: + name: CTCHead + fc_decay: 0.0004 + +Loss: + name: CTCLoss + +PostProcess: + name: CTCLabelDecode + +Metric: + name: RecMetric + main_indicator: acc + +Train: + dataset: + name: LMDBDateSet + data_dir: ./train_data/data_lmdb_release/training/ + transforms: + - DecodeImage: # load image + img_mode: BGR + channel_first: False + - CTCLabelEncode: # Class handling label + - RecResizeImg: + image_shape: [3, 32, 100] + - KeepKeys: + keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order + loader: + shuffle: False + batch_size_per_card: 256 + drop_last: True + num_workers: 8 + +Eval: + dataset: + name: LMDBDateSet + data_dir: ./train_data/data_lmdb_release/validation/ + transforms: + - DecodeImage: # load image + img_mode: BGR + channel_first: False + - CTCLabelEncode: # Class handling label + - RecResizeImg: + image_shape: [3, 32, 100] + - KeepKeys: + keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order + loader: + shuffle: False + drop_last: False + batch_size_per_card: 256 + num_workers: 4 diff --git a/configs/rec/rec_r34_vd_none_bilstm_ctc.yml b/configs/rec/rec_r34_vd_none_bilstm_ctc.yml index 74937138aac1ae155ce8754d861d21a85b35e031..4ad2ff89ef1e72c58c426670742bc2ada27cfc4a 100644 --- a/configs/rec/rec_r34_vd_none_bilstm_ctc.yml +++ b/configs/rec/rec_r34_vd_none_bilstm_ctc.yml @@ -5,7 +5,7 @@ Global: print_batch_step: 10 save_model_dir: ./output/rec/r34_vd_none_bilstm_ctc/ save_epoch_step: 3 - # evaluation is run every 5000 iterations after the 4000th iteration + # evaluation is run every 2000 iterations eval_batch_step: [0, 2000] # if pretrained_model is saved in static mode, load_static_weights must set to True cal_metric_during_train: True @@ -13,7 +13,7 @@ Global: checkpoints: save_inference_dir: use_visualdl: False - infer_img: doc/imgs_words/ch/word_1.jpg + infer_img: doc/imgs_words_en/word_10.png # for data or label process character_dict_path: character_type: en @@ -21,7 +21,6 @@ Global: infer_mode: False use_space_char: False - Optimizer: name: Adam beta1: 0.9 @@ -71,7 +70,7 @@ Train: - KeepKeys: keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order loader: - shuffle: False + shuffle: True batch_size_per_card: 256 drop_last: True num_workers: 8 diff --git a/configs/rec/rec_r34_vd_none_none_ctc.yml b/configs/rec/rec_r34_vd_none_none_ctc.yml new file mode 100644 index 0000000000000000000000000000000000000000..9c1eeb304f41d46e49cee350e5d659dd1e0c8b0e --- /dev/null +++ b/configs/rec/rec_r34_vd_none_none_ctc.yml @@ -0,0 +1,93 @@ +Global: + use_gpu: true + epoch_num: 72 + log_smooth_window: 20 + print_batch_step: 10 + save_model_dir: ./output/rec/r34_vd_none_none_ctc/ + save_epoch_step: 3 + # evaluation is run every 2000 iterations + eval_batch_step: [0, 2000] + # if pretrained_model is saved in static mode, load_static_weights must set to True + cal_metric_during_train: True + pretrained_model: + checkpoints: + save_inference_dir: + use_visualdl: False + infer_img: doc/imgs_words_en/word_10.png + # for data or label process + character_dict_path: + character_type: en + max_text_length: 25 + infer_mode: False + use_space_char: False + +Optimizer: + name: Adam + beta1: 0.9 + beta2: 0.999 + lr: + learning_rate: 0.0005 + regularizer: + name: 'L2' + factor: 0 + +Architecture: + model_type: rec + algorithm: Rosetta + Backbone: + name: ResNet + layers: 34 + Neck: + name: SequenceEncoder + encoder_type: reshape + Head: + name: CTCHead + fc_decay: 0.0004 + +Loss: + name: CTCLoss + +PostProcess: + name: CTCLabelDecode + +Metric: + name: RecMetric + main_indicator: acc + +Train: + dataset: + name: LMDBDateSet + data_dir: ./train_data/data_lmdb_release/training/ + transforms: + - DecodeImage: # load image + img_mode: BGR + channel_first: False + - CTCLabelEncode: # Class handling label + - RecResizeImg: + image_shape: [3, 32, 100] + - KeepKeys: + keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order + loader: + shuffle: True + batch_size_per_card: 256 + drop_last: True + num_workers: 8 + +Eval: + dataset: + name: LMDBDateSet + data_dir: ./train_data/data_lmdb_release/validation/ + transforms: + - DecodeImage: # load image + img_mode: BGR + channel_first: False + - CTCLabelEncode: # Class handling label + - RecResizeImg: + image_shape: [3, 32, 100] + - KeepKeys: + keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order + loader: + shuffle: False + drop_last: False + batch_size_per_card: 256 + num_workers: 4 diff --git a/configs/rec/rec_r34_vd_tps_bilstm_ctc.yml b/configs/rec/rec_r34_vd_tps_bilstm_ctc.yml index 269f1e411776d591610082cff2900ca6fc621752..aeded4926a6d09cf30210f2d348d2933461a06b1 100644 --- a/configs/rec/rec_r34_vd_tps_bilstm_ctc.yml +++ b/configs/rec/rec_r34_vd_tps_bilstm_ctc.yml @@ -5,7 +5,7 @@ Global: print_batch_step: 10 save_model_dir: ./output/rec/r34_vd_tps_bilstm_ctc/ save_epoch_step: 3 - # evaluation is run every 5000 iterations after the 4000th iteration + # evaluation is run every 2000 iterations eval_batch_step: [0, 2000] # if pretrained_model is saved in static mode, load_static_weights must set to True cal_metric_during_train: True @@ -13,7 +13,7 @@ Global: checkpoints: save_inference_dir: use_visualdl: False - infer_img: doc/imgs_words/ch/word_1.jpg + infer_img: doc/imgs_words_en/word_10.png # for data or label process character_dict_path: character_type: en @@ -21,7 +21,6 @@ Global: infer_mode: False use_space_char: False - Optimizer: name: Adam beta1: 0.9 @@ -34,7 +33,7 @@ Optimizer: Architecture: model_type: rec - algorithm: CRNN + algorithm: STARNet Transform: name: TPS num_fiducial: 20 diff --git a/deploy/hubserving/ocr_cls/__init__.py b/deploy/hubserving/ocr_cls/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/deploy/hubserving/ocr_cls/config.json b/deploy/hubserving/ocr_cls/config.json new file mode 100644 index 0000000000000000000000000000000000000000..2ced861b57efe3b9a426c129ff8a3f56e9d83a60 --- /dev/null +++ b/deploy/hubserving/ocr_cls/config.json @@ -0,0 +1,15 @@ +{ + "modules_info": { + "ocr_cls": { + "init_args": { + "version": "1.0.0", + "use_gpu": true + }, + "predict_args": { + } + } + }, + "port": 8866, + "use_multiprocess": false, + "workers": 2 +} diff --git a/deploy/hubserving/ocr_cls/module.py b/deploy/hubserving/ocr_cls/module.py new file mode 100644 index 0000000000000000000000000000000000000000..1b91580ca37027e512a73138270d051497af5b89 --- /dev/null +++ b/deploy/hubserving/ocr_cls/module.py @@ -0,0 +1,121 @@ +# -*- coding:utf-8 -*- +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import os +import sys +sys.path.insert(0, ".") + +from paddlehub.common.logger import logger +from paddlehub.module.module import moduleinfo, runnable, serving +import cv2 +import paddlehub as hub + +from tools.infer.utility import base64_to_cv2 +from tools.infer.predict_cls import TextClassifier + + +@moduleinfo( + name="ocr_cls", + version="1.0.0", + summary="ocr recognition service", + author="paddle-dev", + author_email="paddle-dev@baidu.com", + type="cv/text_recognition") +class OCRCls(hub.Module): + def _initialize(self, use_gpu=False, enable_mkldnn=False): + """ + initialize with the necessary elements + """ + from ocr_cls.params import read_params + cfg = read_params() + + cfg.use_gpu = use_gpu + if use_gpu: + try: + _places = os.environ["CUDA_VISIBLE_DEVICES"] + int(_places[0]) + print("use gpu: ", use_gpu) + print("CUDA_VISIBLE_DEVICES: ", _places) + cfg.gpu_mem = 8000 + except: + raise RuntimeError( + "Environment Variable CUDA_VISIBLE_DEVICES is not set correctly. If you wanna use gpu, please set CUDA_VISIBLE_DEVICES via export CUDA_VISIBLE_DEVICES=cuda_device_id." + ) + cfg.ir_optim = True + cfg.enable_mkldnn = enable_mkldnn + + self.text_classifier = TextClassifier(cfg) + + def read_images(self, paths=[]): + images = [] + for img_path in paths: + assert os.path.isfile( + img_path), "The {} isn't a valid file.".format(img_path) + img = cv2.imread(img_path) + if img is None: + logger.info("error in loading image:{}".format(img_path)) + continue + images.append(img) + return images + + def predict(self, images=[], paths=[]): + """ + Get the text angle in the predicted images. + Args: + images (list(numpy.ndarray)): images data, shape of each is [H, W, C]. If images not paths + paths (list[str]): The paths of images. If paths not images + Returns: + res (list): The result of text detection box and save path of images. + """ + + if images != [] and isinstance(images, list) and paths == []: + predicted_data = images + elif images == [] and isinstance(paths, list) and paths != []: + predicted_data = self.read_images(paths) + else: + raise TypeError("The input data is inconsistent with expectations.") + + assert predicted_data != [], "There is not any image to be predicted. Please check the input data." + + img_list = [] + for img in predicted_data: + if img is None: + continue + img_list.append(img) + + rec_res_final = [] + try: + img_list, cls_res, predict_time = self.text_classifier(img_list) + for dno in range(len(cls_res)): + angle, score = cls_res[dno] + rec_res_final.append({ + 'angle': angle, + 'confidence': float(score), + }) + except Exception as e: + print(e) + return [[]] + + return [rec_res_final] + + @serving + def serving_method(self, images, **kwargs): + """ + Run as a service. + """ + images_decode = [base64_to_cv2(image) for image in images] + results = self.predict(images_decode, **kwargs) + return results + + +if __name__ == '__main__': + ocr = OCRCls() + image_path = [ + './doc/imgs_words/ch/word_1.jpg', + './doc/imgs_words/ch/word_2.jpg', + './doc/imgs_words/ch/word_3.jpg', + ] + res = ocr.predict(paths=image_path) + print(res) diff --git a/deploy/hubserving/ocr_cls/params.py b/deploy/hubserving/ocr_cls/params.py new file mode 100644 index 0000000000000000000000000000000000000000..bcdb2d6e3800c0ba7897b71f0b0999cafdc223af --- /dev/null +++ b/deploy/hubserving/ocr_cls/params.py @@ -0,0 +1,24 @@ +# -*- coding:utf-8 -*- +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + + +class Config(object): + pass + + +def read_params(): + cfg = Config() + + #params for text classifier + cfg.cls_model_dir = "./inference/ch_ppocr_mobile_v1.1_cls_infer/" + cfg.cls_image_shape = "3, 48, 192" + cfg.label_list = ['0', '180'] + cfg.cls_batch_num = 30 + cfg.cls_thresh = 0.9 + + cfg.use_zero_copy_run = False + cfg.use_pdserving = False + + return cfg diff --git a/deploy/hubserving/ocr_det/config.json b/deploy/hubserving/ocr_det/config.json index c8ef055e05470b8011db7a59782d2edc8c123782..6080d1c53ad065f47739a77e95bd15f792c63da0 100644 --- a/deploy/hubserving/ocr_det/config.json +++ b/deploy/hubserving/ocr_det/config.json @@ -9,7 +9,7 @@ } } }, - "port": 8866, + "port": 8865, "use_multiprocess": false, "workers": 2 } diff --git a/deploy/hubserving/ocr_det/module.py b/deploy/hubserving/ocr_det/module.py index be74306dacf4a3648e3227f11227e6399e6ed2eb..5f7bd6c473884b80fbe4d24088808444633f100d 100644 --- a/deploy/hubserving/ocr_det/module.py +++ b/deploy/hubserving/ocr_det/module.py @@ -3,20 +3,14 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function -import argparse -import ast -import copy -import math import os -import time +import sys +sys.path.insert(0, ".") -from paddle.fluid.core import AnalysisConfig, create_paddle_predictor, PaddleTensor from paddlehub.common.logger import logger from paddlehub.module.module import moduleinfo, runnable, serving -from PIL import Image import cv2 import numpy as np -import paddle.fluid as fluid import paddlehub as hub from tools.infer.utility import base64_to_cv2 @@ -67,9 +61,7 @@ class OCRDet(hub.Module): images.append(img) return images - def predict(self, - images=[], - paths=[]): + def predict(self, images=[], paths=[]): """ Get the text box in the predicted images. Args: @@ -87,7 +79,7 @@ class OCRDet(hub.Module): raise TypeError("The input data is inconsistent with expectations.") assert predicted_data != [], "There is not any image to be predicted. Please check the input data." - + all_results = [] for img in predicted_data: if img is None: @@ -99,11 +91,9 @@ class OCRDet(hub.Module): rec_res_final = [] for dno in range(len(dt_boxes)): - rec_res_final.append( - { - 'text_region': dt_boxes[dno].astype(np.int).tolist() - } - ) + rec_res_final.append({ + 'text_region': dt_boxes[dno].astype(np.int).tolist() + }) all_results.append(rec_res_final) return all_results @@ -116,7 +106,7 @@ class OCRDet(hub.Module): results = self.predict(images_decode, **kwargs) return results - + if __name__ == '__main__': ocr = OCRDet() image_path = [ @@ -124,4 +114,4 @@ if __name__ == '__main__': './doc/imgs/12.jpg', ] res = ocr.predict(paths=image_path) - print(res) \ No newline at end of file + print(res) diff --git a/deploy/hubserving/ocr_det/params.py b/deploy/hubserving/ocr_det/params.py index e88ab45c7bb548ef971465d4aaefb30d247ab17f..4d4a9fc27b727034d8185c82dad3e542659fd463 100644 --- a/deploy/hubserving/ocr_det/params.py +++ b/deploy/hubserving/ocr_det/params.py @@ -10,16 +10,17 @@ class Config(object): def read_params(): cfg = Config() - + #params for text detector cfg.det_algorithm = "DB" - cfg.det_model_dir = "./inference/ch_det_mv3_db/" - cfg.det_max_side_len = 960 + cfg.det_model_dir = "./inference/ch_ppocr_mobile_v1.1_det_infer/" + cfg.det_limit_side_len = 960 + cfg.det_limit_type = 'max' #DB parmas - cfg.det_db_thresh =0.3 - cfg.det_db_box_thresh =0.5 - cfg.det_db_unclip_ratio =2.0 + cfg.det_db_thresh = 0.3 + cfg.det_db_box_thresh = 0.5 + cfg.det_db_unclip_ratio = 2.0 # #EAST parmas # cfg.det_east_score_thresh = 0.8 @@ -37,5 +38,6 @@ def read_params(): # cfg.use_space_char = True cfg.use_zero_copy_run = False + cfg.use_pdserving = False return cfg diff --git a/deploy/hubserving/ocr_rec/module.py b/deploy/hubserving/ocr_rec/module.py index 846f5437fe3b0a5136bff6c902481f888558d594..41a42104a81b736eeca346737d2dcefe3d728ef8 100644 --- a/deploy/hubserving/ocr_rec/module.py +++ b/deploy/hubserving/ocr_rec/module.py @@ -3,20 +3,13 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function -import argparse -import ast -import copy -import math import os -import time +import sys +sys.path.insert(0, ".") -from paddle.fluid.core import AnalysisConfig, create_paddle_predictor, PaddleTensor from paddlehub.common.logger import logger from paddlehub.module.module import moduleinfo, runnable, serving -from PIL import Image import cv2 -import numpy as np -import paddle.fluid as fluid import paddlehub as hub from tools.infer.utility import base64_to_cv2 @@ -67,9 +60,7 @@ class OCRRec(hub.Module): images.append(img) return images - def predict(self, - images=[], - paths=[]): + def predict(self, images=[], paths=[]): """ Get the text box in the predicted images. Args: @@ -87,31 +78,28 @@ class OCRRec(hub.Module): raise TypeError("The input data is inconsistent with expectations.") assert predicted_data != [], "There is not any image to be predicted. Please check the input data." - + img_list = [] for img in predicted_data: if img is None: continue img_list.append(img) - + rec_res_final = [] try: rec_res, predict_time = self.text_recognizer(img_list) for dno in range(len(rec_res)): text, score = rec_res[dno] - rec_res_final.append( - { - 'text': text, - 'confidence': float(score), - } - ) + rec_res_final.append({ + 'text': text, + 'confidence': float(score), + }) except Exception as e: print(e) return [[]] return [rec_res_final] - @serving def serving_method(self, images, **kwargs): """ @@ -121,7 +109,7 @@ class OCRRec(hub.Module): results = self.predict(images_decode, **kwargs) return results - + if __name__ == '__main__': ocr = OCRRec() image_path = [ @@ -130,4 +118,4 @@ if __name__ == '__main__': './doc/imgs_words/ch/word_3.jpg', ] res = ocr.predict(paths=image_path) - print(res) \ No newline at end of file + print(res) diff --git a/deploy/hubserving/ocr_rec/params.py b/deploy/hubserving/ocr_rec/params.py index 59772e2163d1d5f8279dee85432b5bf93502914e..6f428ecb2686afa5ff66b84d963d1c2175b9cee2 100644 --- a/deploy/hubserving/ocr_rec/params.py +++ b/deploy/hubserving/ocr_rec/params.py @@ -10,25 +10,10 @@ class Config(object): def read_params(): cfg = Config() - - # #params for text detector - # cfg.det_algorithm = "DB" - # cfg.det_model_dir = "./inference/ch_det_mv3_db/" - # cfg.det_max_side_len = 960 - - # #DB parmas - # cfg.det_db_thresh =0.3 - # cfg.det_db_box_thresh =0.5 - # cfg.det_db_unclip_ratio =2.0 - - # #EAST parmas - # cfg.det_east_score_thresh = 0.8 - # cfg.det_east_cover_thresh = 0.1 - # cfg.det_east_nms_thresh = 0.2 #params for text recognizer cfg.rec_algorithm = "CRNN" - cfg.rec_model_dir = "./inference/ch_rec_mv3_crnn/" + cfg.rec_model_dir = "./inference/ch_ppocr_mobile_v1.1_rec_infer/" cfg.rec_image_shape = "3, 32, 320" cfg.rec_char_type = 'ch' @@ -39,5 +24,6 @@ def read_params(): cfg.use_space_char = True cfg.use_zero_copy_run = False + cfg.use_pdserving = False return cfg diff --git a/deploy/hubserving/ocr_system/module.py b/deploy/hubserving/ocr_system/module.py index cb526e1185d8eb623af84ba2451ff38523bbf642..7f3617330bc87ac77a150af9e40dc125d3bfd7de 100644 --- a/deploy/hubserving/ocr_system/module.py +++ b/deploy/hubserving/ocr_system/module.py @@ -3,20 +3,16 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function -import argparse -import ast -import copy -import math import os +import sys +sys.path.insert(0, ".") + import time -from paddle.fluid.core import AnalysisConfig, create_paddle_predictor, PaddleTensor from paddlehub.common.logger import logger from paddlehub.module.module import moduleinfo, runnable, serving -from PIL import Image import cv2 import numpy as np -import paddle.fluid as fluid import paddlehub as hub from tools.infer.utility import base64_to_cv2 @@ -52,7 +48,7 @@ class OCRSystem(hub.Module): ) cfg.ir_optim = True cfg.enable_mkldnn = enable_mkldnn - + self.text_sys = TextSystem(cfg) def read_images(self, paths=[]): @@ -67,9 +63,7 @@ class OCRSystem(hub.Module): images.append(img) return images - def predict(self, - images=[], - paths=[]): + def predict(self, images=[], paths=[]): """ Get the chinese texts in the predicted images. Args: @@ -104,13 +98,11 @@ class OCRSystem(hub.Module): for dno in range(dt_num): text, score = rec_res[dno] - rec_res_final.append( - { - 'text': text, - 'confidence': float(score), - 'text_region': dt_boxes[dno].astype(np.int).tolist() - } - ) + rec_res_final.append({ + 'text': text, + 'confidence': float(score), + 'text_region': dt_boxes[dno].astype(np.int).tolist() + }) all_results.append(rec_res_final) return all_results @@ -123,7 +115,7 @@ class OCRSystem(hub.Module): results = self.predict(images_decode, **kwargs) return results - + if __name__ == '__main__': ocr = OCRSystem() image_path = [ @@ -131,4 +123,4 @@ if __name__ == '__main__': './doc/imgs/12.jpg', ] res = ocr.predict(paths=image_path) - print(res) \ No newline at end of file + print(res) diff --git a/deploy/hubserving/ocr_system/params.py b/deploy/hubserving/ocr_system/params.py index 0ff56d37d50b30b09bb13b529a48a260dfe8f84a..1f6a07bcc0167e90564edab9c4719b9192233b4c 100644 --- a/deploy/hubserving/ocr_system/params.py +++ b/deploy/hubserving/ocr_system/params.py @@ -10,16 +10,17 @@ class Config(object): def read_params(): cfg = Config() - + #params for text detector cfg.det_algorithm = "DB" - cfg.det_model_dir = "./inference/ch_det_mv3_db/" - cfg.det_max_side_len = 960 + cfg.det_model_dir = "./inference/ch_ppocr_mobile_v1.1_det_infer/" + cfg.det_limit_side_len = 960 + cfg.det_limit_type = 'max' #DB parmas - cfg.det_db_thresh =0.3 - cfg.det_db_box_thresh =0.5 - cfg.det_db_unclip_ratio =2.0 + cfg.det_db_thresh = 0.3 + cfg.det_db_box_thresh = 0.5 + cfg.det_db_unclip_ratio = 2.0 #EAST parmas cfg.det_east_score_thresh = 0.8 @@ -28,7 +29,7 @@ def read_params(): #params for text recognizer cfg.rec_algorithm = "CRNN" - cfg.rec_model_dir = "./inference/ch_rec_mv3_crnn/" + cfg.rec_model_dir = "./inference/ch_ppocr_mobile_v1.1_rec_infer/" cfg.rec_image_shape = "3, 32, 320" cfg.rec_char_type = 'ch' @@ -38,6 +39,15 @@ def read_params(): cfg.rec_char_dict_path = "./ppocr/utils/ppocr_keys_v1.txt" cfg.use_space_char = True + #params for text classifier + cfg.use_angle_cls = True + cfg.cls_model_dir = "./inference/ch_ppocr_mobile_v1.1_cls_infer/" + cfg.cls_image_shape = "3, 48, 192" + cfg.label_list = ['0', '180'] + cfg.cls_batch_num = 30 + cfg.cls_thresh = 0.9 + cfg.use_zero_copy_run = False + cfg.use_pdserving = False return cfg diff --git a/deploy/hubserving/readme.md b/deploy/hubserving/readme.md new file mode 100644 index 0000000000000000000000000000000000000000..f64bd372569f12ea52214e3e89927df0c859a17f --- /dev/null +++ b/deploy/hubserving/readme.md @@ -0,0 +1,193 @@ +[English](readme_en.md) | 简体中文 + +PaddleOCR提供2种服务部署方式： +- 基于PaddleHub Serving的部署：代码路径为"`./deploy/hubserving`"，按照本教程使用； +- 基于PaddleServing的部署：代码路径为"`./deploy/pdserving`"，使用方法参考[文档](../../deploy/pdserving/readme.md)。 + +# 基于PaddleHub Serving的服务部署 + +hubserving服务部署目录下包括检测、识别、2阶段串联三种服务包，请根据需求选择相应的服务包进行安装和启动。目录结构如下： +``` +deploy/hubserving/ + └─ ocr_cls 分类模块服务包 + └─ ocr_det 检测模块服务包 + └─ ocr_rec 识别模块服务包 + └─ ocr_system 检测+识别串联服务包 +``` + +每个服务包下包含3个文件。以2阶段串联服务包为例，目录如下： +``` +deploy/hubserving/ocr_system/ + └─ __init__.py 空文件，必选 + └─ config.json 配置文件，可选，使用配置启动服务时作为参数传入 + └─ module.py 主模块，必选，包含服务的完整逻辑 + └─ params.py 参数文件，必选，包含模型路径、前后处理参数等参数 +``` + +## 快速启动服务 +以下步骤以检测+识别2阶段串联服务为例，如果只需要检测服务或识别服务，替换相应文件路径即可。 +### 1. 准备环境 +```shell +# 安装paddlehub +pip3 install paddlehub --upgrade -i https://pypi.tuna.tsinghua.edu.cn/simple +``` + +### 2. 下载推理模型 +安装服务模块前，需要准备推理模型并放到正确路径。默认使用的是v1.1版的超轻量模型，默认模型路径为： +``` +检测模型：./inference/ch_ppocr_mobile_v1.1_det_infer/ +识别模型：./inference/ch_ppocr_mobile_v1.1_rec_infer/ +方向分类器：./inference/ch_ppocr_mobile_v1.1_cls_infer/ +``` + +**模型路径可在`params.py`中查看和修改。** 更多模型可以从PaddleOCR提供的[模型库](../../doc/doc_ch/models_list.md)下载，也可以替换成自己训练转换好的模型。 + +### 3. 安装服务模块 +PaddleOCR提供3种服务模块，根据需要安装所需模块。 + +* 在Linux环境下，安装示例如下： +```shell +# 安装检测服务模块： +hub install deploy/hubserving/ocr_det/ + +# 或，安装分类服务模块： +hub install deploy/hubserving/ocr_cls/ + +# 或，安装识别服务模块： +hub install deploy/hubserving/ocr_rec/ + +# 或，安装检测+识别串联服务模块： +hub install deploy/hubserving/ocr_system/ +``` + +* 在Windows环境下(文件夹的分隔符为`\`)，安装示例如下： +```shell +# 安装检测服务模块： +hub install deploy\hubserving\ocr_det\ + +# 或，安装分类服务模块： +hub install deploy\hubserving\ocr_cls\ + +# 或，安装识别服务模块： +hub install deploy\hubserving\ocr_rec\ + +# 或，安装检测+识别串联服务模块： +hub install deploy\hubserving\ocr_system\ +``` + +### 4. 启动服务 +#### 方式1. 命令行命令启动（仅支持CPU） +**启动命令：** +```shell +$ hub serving start --modules [Module1==Version1, Module2==Version2, ...] \ + --port XXXX \ + --use_multiprocess \ + --workers \ +``` + +**参数：** + +|参数|用途| +|-|-| +|--modules/-m|PaddleHub Serving预安装模型，以多个Module==Version键值对的形式列出
*`当不指定Version时，默认选择最新版本`*| +|--port/-p|服务端口，默认为8866| +|--use_multiprocess|是否启用并发方式，默认为单进程方式，推荐多核CPU机器使用此方式
*`Windows操作系统只支持单进程方式`*| +|--workers|在并发方式下指定的并发任务数，默认为`2*cpu_count-1`，其中`cpu_count`为CPU核数| + +如启动串联服务： ```hub serving start -m ocr_system``` + +这样就完成了一个服务化API的部署，使用默认端口号8866。 + +#### 方式2. 配置文件启动（支持CPU、GPU） +**启动命令：** +```hub serving start -c config.json``` + +其中，`config.json`格式如下： +```python +{ + "modules_info": { + "ocr_system": { + "init_args": { + "version": "1.0.0", + "use_gpu": true + }, + "predict_args": { + } + } + }, + "port": 8868, + "use_multiprocess": false, + "workers": 2 +} +``` + +- `init_args`中的可配参数与`module.py`中的`_initialize`函数接口一致。其中，**当`use_gpu`为`true`时，表示使用GPU启动服务**。 +- `predict_args`中的可配参数与`module.py`中的`predict`函数接口一致。 + +**注意:** +- 使用配置文件启动服务时，其他参数会被忽略。 +- 如果使用GPU预测(即，`use_gpu`置为`true`)，则需要在启动服务之前，设置CUDA_VISIBLE_DEVICES环境变量，如：```export CUDA_VISIBLE_DEVICES=0```，否则不用设置。 +- **`use_gpu`不可与`use_multiprocess`同时为`true`**。 + +如，使用GPU 3号卡启动串联服务： +```shell +export CUDA_VISIBLE_DEVICES=3 +hub serving start -c deploy/hubserving/ocr_system/config.json +``` + +## 发送预测请求 +配置好服务端，可使用以下命令发送预测请求，获取预测结果: + +```python tools/test_hubserving.py server_url image_path``` + +需要给脚本传递2个参数： +- **server_url**：服务地址，格式为 +`http://[ip_address]:[port]/predict/[module_name]` +例如，如果使用配置文件启动分类，检测、识别，检测+分类+识别3阶段服务，那么发送请求的url将分别是： +`http://127.0.0.1:8865/predict/ocr_det` +`http://127.0.0.1:8866/predict/ocr_cls` +`http://127.0.0.1:8867/predict/ocr_rec` +`http://127.0.0.1:8868/predict/ocr_system` +- **image_path**：测试图像路径，可以是单张图片路径，也可以是图像集合目录路径 + +访问示例： +```python tools/test_hubserving.py http://127.0.0.1:8868/predict/ocr_system ./doc/imgs/``` + +## 返回结果格式说明 +返回结果为列表（list），列表中的每一项为词典（dict），词典一共可能包含3种字段，信息如下： + +|字段名称|数据类型|意义| +|----|----|----| +|angle|str|文本角度| +|text|str|文本内容| +|confidence|float| 文本识别置信度或文本角度分类置信度| +|text_region|list|文本位置坐标| + +不同模块返回的字段不同，如，文本识别服务模块返回结果不含`text_region`字段，具体信息如下： + +| 字段名/模块名 | ocr_det | ocr_cls | ocr_rec | ocr_system | +| ---- | ---- | ---- | ---- | ---- | +|angle| | ✔ | | ✔ | +|text| | |✔|✔| +|confidence| |✔ |✔|✔| +|text_region| ✔| | |✔ | + +**说明：** 如果需要增加、删除、修改返回字段，可在相应模块的`module.py`文件中进行修改，完整流程参考下一节自定义修改服务模块。 + +## 自定义修改服务模块 +如果需要修改服务逻辑，你一般需要操作以下步骤（以修改`ocr_system`为例）： + +- 1、 停止服务 +```hub serving stop --port/-p XXXX``` + +- 2、 到相应的`module.py`和`params.py`等文件中根据实际需求修改代码。 +例如，如果需要替换部署服务所用模型，则需要到`params.py`中修改模型路径参数`det_model_dir`和`rec_model_dir`，如果需要关闭文本方向分类器，则将参数`use_angle_cls`置为`False`，当然，同时可能还需要修改其他相关参数，请根据实际情况修改调试。 **强烈建议修改后先直接运行`module.py`调试，能正确运行预测后再启动服务测试。** + +- 3、 卸载旧服务包 +```hub uninstall ocr_system``` + +- 4、 安装修改后的新服务包 +```hub install deploy/hubserving/ocr_system/``` + +- 5、重新启动服务 +```hub serving start -m ocr_system``` diff --git a/deploy/hubserving/readme_en.md b/deploy/hubserving/readme_en.md new file mode 100644 index 0000000000000000000000000000000000000000..c6cf53413bc3eac45f933fead66356d1491cc60c --- /dev/null +++ b/deploy/hubserving/readme_en.md @@ -0,0 +1,204 @@ +English | [简体中文](readme.md) + +PaddleOCR provides 2 service deployment methods: +- Based on **PaddleHub Serving**: Code path is "`./deploy/hubserving`". Please follow this tutorial. +- Based on **PaddleServing**: Code path is "`./deploy/pdserving`". Please refer to the [tutorial](../../deploy/pdserving/readme.md) for usage. + +# Service deployment based on PaddleHub Serving + +The hubserving service deployment directory includes three service packages: detection, recognition, and two-stage series connection. Please select the corresponding service package to install and start service according to your needs. The directory is as follows: +``` +deploy/hubserving/ + └─ ocr_det detection module service package + └─ ocr_cls angle class module service package + └─ ocr_rec recognition module service package + └─ ocr_system two-stage series connection service package +``` + +Each service pack contains 3 files. Take the 2-stage series connection service package as an example, the directory is as follows: +``` +deploy/hubserving/ocr_system/ + └─ __init__.py Empty file, required + └─ config.json Configuration file, optional, passed in as a parameter when using configuration to start the service + └─ module.py Main module file, required, contains the complete logic of the service + └─ params.py Parameter file, required, including parameters such as model path, pre- and post-processing parameters +``` + +## Quick start service +The following steps take the 2-stage series service as an example. If only the detection service or recognition service is needed, replace the corresponding file path. + +### 1. Prepare the environment +```shell +# Install paddlehub +pip3 install paddlehub --upgrade -i https://pypi.tuna.tsinghua.edu.cn/simple +``` + +### 2. Download inference model +Before installing the service module, you need to prepare the inference model and put it in the correct path. By default, the ultra lightweight model of v1.1 is used, and the default model path is: +``` +detection model: ./inference/ch_ppocr_mobile_v1.1_det_infer/ +recognition model: ./inference/ch_ppocr_mobile_v1.1_rec_infer/ +text direction classifier: ./inference/ch_ppocr_mobile_v1.1_cls_infer/ +``` + +**The model path can be found and modified in `params.py`.** More models provided by PaddleOCR can be obtained from the [model library](../../doc/doc_en/models_list_en.md). You can also use models trained by yourself. + +### 3. Install Service Module +PaddleOCR provides 3 kinds of service modules, install the required modules according to your needs. + +* On Linux platform, the examples are as follows. +```shell +# Install the detection service module: +hub install deploy/hubserving/ocr_det/ + +# Or, install the angle class service module: +hub install deploy/hubserving/ocr_cls/ + +# Or, install the recognition service module: +hub install deploy/hubserving/ocr_rec/ + +# Or, install the 2-stage series service module: +hub install deploy/hubserving/ocr_system/ +``` + +* On Windows platform, the examples are as follows. +```shell +# Install the detection service module: +hub install deploy\hubserving\ocr_det\ + +# Or, install the angle class service module: +hub install deploy\hubserving\ocr_cls\ + +# Or, install the recognition service module: +hub install deploy\hubserving\ocr_rec\ + +# Or, install the 2-stage series service module: +hub install deploy\hubserving\ocr_system\ +``` + +### 4. Start service +#### Way 1. Start with command line parameters (CPU only) + +**start command：** +```shell +$ hub serving start --modules [Module1==Version1, Module2==Version2, ...] \ + --port XXXX \ + --use_multiprocess \ + --workers \ +``` +**parameters：** + +|parameters|usage| +|-|-| +|--modules/-m|PaddleHub Serving pre-installed model, listed in the form of multiple Module==Version key-value pairs
*`When Version is not specified, the latest version is selected by default`*| +|--port/-p|Service port, default is 8866| +|--use_multiprocess|Enable concurrent mode, the default is single-process mode, this mode is recommended for multi-core CPU machines
*`Windows operating system only supports single-process mode`*| +|--workers|The number of concurrent tasks specified in concurrent mode, the default is `2*cpu_count-1`, where `cpu_count` is the number of CPU cores| + +For example, start the 2-stage series service: +```shell +hub serving start -m ocr_system +``` + +This completes the deployment of a service API, using the default port number 8866. + +#### Way 2. Start with configuration file（CPU、GPU） +**start command：** +```shell +hub serving start --config/-c config.json +``` +Wherein, the format of `config.json` is as follows: +```python +{ + "modules_info": { + "ocr_system": { + "init_args": { + "version": "1.0.0", + "use_gpu": true + }, + "predict_args": { + } + } + }, + "port": 8868, + "use_multiprocess": false, + "workers": 2 +} +``` +- The configurable parameters in `init_args` are consistent with the `_initialize` function interface in `module.py`. Among them, **when `use_gpu` is `true`, it means that the GPU is used to start the service**. +- The configurable parameters in `predict_args` are consistent with the `predict` function interface in `module.py`. + +**Note:** +- When using the configuration file to start the service, other parameters will be ignored. +- If you use GPU prediction (that is, `use_gpu` is set to `true`), you need to set the environment variable CUDA_VISIBLE_DEVICES before starting the service, such as: ```export CUDA_VISIBLE_DEVICES=0```, otherwise you do not need to set it. +- **`use_gpu` and `use_multiprocess` cannot be `true` at the same time.** + +For example, use GPU card No. 3 to start the 2-stage series service: +```shell +export CUDA_VISIBLE_DEVICES=3 +hub serving start -c deploy/hubserving/ocr_system/config.json +``` + +## Send prediction requests +After the service starts, you can use the following command to send a prediction request to obtain the prediction result: +```shell +python tools/test_hubserving.py server_url image_path +``` + +Two parameters need to be passed to the script: +- **server_url**：service address，format of which is +`http://[ip_address]:[port]/predict/[module_name]` +For example, if the detection, recognition and 2-stage serial services are started with provided configuration files, the respective `server_url` would be: +`http://127.0.0.1:8865/predict/ocr_det` +`http://127.0.0.1:8866/predict/ocr_cls` +`http://127.0.0.1:8867/predict/ocr_rec` +`http://127.0.0.1:8868/predict/ocr_system` +- **image_path**：Test image path, can be a single image path or an image directory path + +**Eg.** +```shell +python tools/test_hubserving.py http://127.0.0.1:8868/predict/ocr_system ./doc/imgs/ +``` + +## Returned result format +The returned result is a list. Each item in the list is a dict. The dict may contain three fields. The information is as follows: + +|field name|data type|description| +|----|----|----| +|angle|str|angle| +|text|str|text content| +|confidence|float|text recognition confidence| +|text_region|list|text location coordinates| + +The fields returned by different modules are different. For example, the results returned by the text recognition service module do not contain `text_region`. The details are as follows: + +| field name/module name | ocr_det | ocr_cls | ocr_rec | ocr_system | +| ---- | ---- | ---- | ---- | ---- | +|angle| | ✔ | | ✔ | +|text| | |✔|✔| +|confidence| |✔ |✔|✔| +|text_region| ✔| | |✔ | + +**Note：** If you need to add, delete or modify the returned fields, you can modify the file `module.py` of the corresponding module. For the complete process, refer to the user-defined modification service module in the next section. + +## User defined service module modification +If you need to modify the service logic, the following steps are generally required (take the modification of `ocr_system` for example): + +- 1. Stop service +```shell +hub serving stop --port/-p XXXX +``` +- 2. Modify the code in the corresponding files, like `module.py` and `params.py`, according to the actual needs. +For example, if you need to replace the model used by the deployed service, you need to modify model path parameters `det_model_dir` and `rec_model_dir` in `params.py`. If you want to turn off the text direction classifier, set the parameter `use_angle_cls` to `False`. Of course, other related parameters may need to be modified at the same time. Please modify and debug according to the actual situation. It is suggested to run `module.py` directly for debugging after modification before starting the service test. +- 3. Uninstall old service module +```shell +hub uninstall ocr_system +``` +- 4. Install modified service module +```shell +hub install deploy/hubserving/ocr_system/ +``` +- 5. Restart service +```shell +hub serving start -m ocr_system +``` diff --git a/doc/doc_ch/add_new_algorithm.md b/doc/doc_ch/add_new_algorithm.md new file mode 100644 index 0000000000000000000000000000000000000000..37159c801a5ebc83c3b855a70ea772ba641e5297 --- /dev/null +++ b/doc/doc_ch/add_new_algorithm.md @@ -0,0 +1,303 @@ +# 添加新算法 + +PaddleOCR将一个算法分解为以下几个部分，并对各部分进行模块化处理，方便快速组合出新的算法。 + +* 数据加载和处理 +* 网络 +* 后处理 +* 损失函数 +* 指标评估 +* 优化器 + +下面将分别对每个部分进行介绍，并介绍如何在该部分里添加新算法所需模块。 + +## 数据加载和处理 + +数据加载和处理由不同的模块(module)组成，其完成了图片的读取、数据增强和label的制作。这一部分在[ppocr/data](../../ppocr/data)下。 各个文件及文件夹作用说明如下: + +```bash +ppocr/data/ +├── imaug # 图片的读取、数据增强和label制作相关的文件 +│ ├── label_ops.py # 对label进行变换的modules +│ ├── operators.py # 对image进行变换的modules +│ ├──..... +├── __init__.py +├── lmdb_dataset.py # 读取lmdb的数据集的dataset +└── simple_dataset.py # 读取以`image_path\tgt`形式保存的数据集的dataset +``` + +PaddleOCR内置了大量图像操作相关模块，对于没有没有内置的模块可通过如下步骤添加: + +1. 在 [ppocr/data/imaug](../../ppocr/data/imaug) 文件夹下新建文件，如my_module.py。 +2. 在 my_module.py 文件内添加相关代码，示例代码如下: + +```python +class MyModule: + def __init__(self, *args, **kwargs): + # your init code + pass + + def __call__(self, data): + img = data['image'] + label = data['label'] + # your process code + + data['image'] = img + data['label'] = label + return data +``` + +3. 在 [ppocr/data/imaug/\__init\__.py](../../ppocr/data/imaug/__init__.py) 文件内导入添加的模块。 + +数据处理的所有处理步骤由不同的模块顺序执行而成，在config文件中按照列表的形式组合并执行。如: + +```yaml +# angle class data process +transforms: + - DecodeImage: # load image + img_mode: BGR + channel_first: False + - MyModule: + args1: args1 + args2: args2 + - KeepKeys: + keep_keys: [ 'image', 'label' ] # dataloader will return list in this order +``` + +## 网络 + +网络部分完成了网络的组网操作，PaddleOCR将网络划分为四部分，这一部分在[ppocr/modeling](../../ppocr/modeling)下。 进入网络的数据将按照顺序(transforms->backbones-> +necks->heads)依次通过这四个部分。 + +```bash +├── architectures # 网络的组网代码 +├── transforms # 网络的图像变换模块 +├── backbones # 网络的特征提取模块 +├── necks # 网络的特征增强模块 +└── heads # 网络的输出模块 +``` + +PaddleOCR内置了DB,EAST,SAST,CRNN和Attention等算法相关的常用模块，对于没有内置的模块可通过如下步骤添加，四个部分添加步骤一致，以backbones为例: + +1. 在 [ppocr/modeling/backbones](../../ppocr/modeling/backbones) 文件夹下新建文件，如my_backbone.py。 +2. 在 my_backbone.py 文件内添加相关代码，示例代码如下: + +```python +import paddle +import paddle.nn as nn +import paddle.nn.functional as F + + +class MyBackbone(nn.Layer): + def __init__(self, *args, **kwargs): + super(MyBackbone, self).__init__() + # your init code + self.conv = nn.xxxx + + def forward(self, inputs): + # your necwork forward + y = self.conv(inputs) + return y +``` + +3. 在 [ppocr/modeling/backbones/\__init\__.py](../../ppocr/modeling/backbones/__init__.py)文件内导入添加的模块。 + +在完成网络的四部分模块添加之后，只需要配置文件中进行配置即可使用，如: + +```yaml +Architecture: + model_type: rec + algorithm: CRNN + Transform: + name: MyTransform + args1: args1 + args2: args2 + Backbone: + name: MyBackbone + args1: args1 + Neck: + name: MyNeck + args1: args1 + Head: + name: MyHead + args1: args1 +``` + +## 后处理 + +后处理实现解码网络输出获得文本框或者识别到的文字。这一部分在[ppocr/postprocess](../../ppocr/postprocess)下。 +PaddleOCR内置了DB,EAST,SAST,CRNN和Attention等算法相关的后处理模块，对于没有内置的组件可通过如下步骤添加: + +1. 在 [ppocr/postprocess](../../ppocr/postprocess) 文件夹下新建文件，如 my_postprocess.py。 +2. 在 my_postprocess.py 文件内添加相关代码，示例代码如下: + +```python +import paddle + + +class MyPostProcess: + def __init__(self, *args, **kwargs): + # your init code + pass + + def __call__(self, preds, label=None, *args, **kwargs): + if isinstance(preds, paddle.Tensor): + preds = preds.numpy() + # you preds decode code + preds = self.decode_preds(preds) + if label is None: + return preds + # you label decode code + label = self.decode_label(label) + return preds, label + + def decode_preds(self, preds): + # you preds decode code + pass + + def decode_label(self, preds): + # you label decode code + pass +``` + +3. 在 [ppocr/postprocess/\__init\__.py](../../ppocr/postprocess/__init__.py)文件内导入添加的模块。 + +在后处理模块添加之后，只需要配置文件中进行配置即可使用，如: + +```yaml +PostProcess: + name: MyPostProcess + args1: args1 + args2: args2 +``` + +## 损失函数 + +损失函数用于计算网络输出和label之间的距离。这一部分在[ppocr/losses](../../ppocr/losses)下。 +PaddleOCR内置了DB,EAST,SAST,CRNN和Attention等算法相关的损失函数模块，对于没有内置的模块可通过如下步骤添加: + +1. 在 [ppocr/losses](../../ppocr/losses) 文件夹下新建文件，如 my_loss.py。 +2. 在 my_loss.py 文件内添加相关代码，示例代码如下: + +```python +import paddle +from paddle import nn + + +class MyLoss(nn.Layer): + def __init__(self, **kwargs): + super(MyLoss, self).__init__() + # you init code + pass + + def __call__(self, predicts, batch): + label = batch[1] + # your loss code + loss = self.loss(input=predicts, label=label) + return {'loss': loss} +``` + +3. 在 [ppocr/losses/\__init\__.py](../../ppocr/losses/__init__.py)文件内导入添加的模块。 + +在损失函数添加之后，只需要配置文件中进行配置即可使用，如: + +```yaml +Loss: + name: MyLoss + args1: args1 + args2: args2 +``` + +## 指标评估 + +指标评估用于计算网络在当前batch上的性能。这一部分在[ppocr/metrics](../../ppocr/metrics)下。 PaddleOCR内置了检测，分类和识别等算法相关的指标评估模块，对于没有内置的模块可通过如下步骤添加: + +1. 在 [ppocr/metrics](../../ppocr/metrics) 文件夹下新建文件，如my_metric.py。 +2. 在 my_metric.py 文件内添加相关代码，示例代码如下: + +```python + +class MyMetric(object): + def __init__(self, main_indicator='acc', **kwargs): + # main_indicator is used for select best model + self.main_indicator = main_indicator + self.reset() + + def __call__(self, preds, batch, *args, **kwargs): + # preds is out of postprocess + # batch is out of dataloader + labels = batch[1] + cur_correct_num = 0 + cur_all_num = 0 + # you metric code + self.correct_num += cur_correct_num + self.all_num += cur_all_num + return {'acc': cur_correct_num / cur_all_num, } + + def get_metric(self): + """ + return metircs { + 'acc': 0, + 'norm_edit_dis': 0, + } + """ + acc = self.correct_num / self.all_num + self.reset() + return {'acc': acc} + + def reset(self): + # reset metric + self.correct_num = 0 + self.all_num = 0 + +``` + +3. 在 [ppocr/metrics/\__init\__.py](../../ppocr/metrics/__init__.py)文件内导入添加的模块。 + +在指标评估模块添加之后，只需要配置文件中进行配置即可使用，如: + +```yaml +Metric: + name: MyMetric + main_indicator: acc +``` + +## 优化器 + +优化器用于训练网络。优化器内部还包含了网络正则化和学习率衰减模块。 这一部分在[ppocr/optimizer](../../ppocr/optimizer)下。 PaddleOCR内置了`Momentum`,`Adam` +和`RMSProp`等常用的优化器模块，`Linear`,`Cosine`,`Step`和`Piecewise`等常用的正则化模块与`L1Decay`和`L2Decay`等常用的学习率衰减模块。 +对于没有内置的模块可通过如下步骤添加，以`optimizer`为例: + +1. 在 [ppocr/optimizer/optimizer.py](../../ppocr/optimizer/optimizer.py) 文件内创建自己的优化器，示例代码如下: + +```python +from paddle import optimizer as optim + + +class MyOptim(object): + def __init__(self, learning_rate=0.001, *args, **kwargs): + self.learning_rate = learning_rate + + def __call__(self, parameters): + # It is recommended to wrap the built-in optimizer of paddle + opt = optim.XXX( + learning_rate=self.learning_rate, + parameters=parameters) + return opt + +``` + +在优化器模块添加之后，只需要配置文件中进行配置即可使用，如: + +```yaml +Optimizer: + name: MyOptim + args1: args1 + args2: args2 + lr: + name: Cosine + learning_rate: 0.001 + regularizer: + name: 'L2' + factor: 0 +``` \ No newline at end of file diff --git a/doc/doc_ch/config.md b/doc/doc_ch/config.md index f9c664d4ea38e2e52dc76bfb5b63d9a515b106a7..2cc502cadf5101c4321ca7543647dd90ea7e0466 100644 --- a/doc/doc_ch/config.md +++ b/doc/doc_ch/config.md @@ -1,4 +1,4 @@ -# 可选参数列表 +## 可选参数列表 以下列表可以通过`--help`查看 @@ -8,65 +8,115 @@ | -o | ALL | 设置配置文件里的参数内容 | None | 使用-o配置相较于-c选择的配置文件具有更高的优先级。例如：`-o Global.use_gpu=false` | -## 配置文件 Global 参数介绍 +## 配置文件参数介绍 以 `rec_chinese_lite_train_v1.1.yml ` 为例 - +### Global | 字段 | 用途 | 默认值 | 备注 | | :----------------------: | :---------------------: | :--------------: | :--------------------: | -| algorithm | 设置算法 | 与配置文件同步 | 选择模型，支持模型请参考[简介](https://github.com/PaddlePaddle/PaddleOCR/blob/develop/README.md) | -| use_gpu | 设置代码运行场所 | true | \ | -| epoch_num | 最大训练epoch数 | 3000 | \ | +| use_gpu | 设置代码是否在gpu运行 | true | \ | +| epoch_num | 最大训练epoch数 | 500 | \ | | log_smooth_window | 滑动窗口大小 | 20 | \ | | print_batch_step | 设置打印log间隔 | 10 | \ | | save_model_dir | 设置模型保存路径 | output/{算法名称} | \ | | save_epoch_step | 设置模型保存间隔 | 3 | \ | | eval_batch_step | 设置模型评估间隔 | 2000 或 [1000, 2000] | 2000 表示每2000次迭代评估一次，[1000， 2000]表示从1000次迭代开始，每2000次评估一次 | -|train_batch_size_per_card | 设置训练时单卡batch size | 256 | \ | -| test_batch_size_per_card | 设置评估时单卡batch size | 256 | \ | -| image_shape | 设置输入图片尺寸 | [3, 32, 100] | \ | +| cal_metric_during_train | 设置是否在训练过程中评估指标，此时评估的是模型在当前batch下的指标 | true | \ | +| load_static_weights | 设置预训练模型是否是静态图模式保存(目前仅检测算法需要) | true | \ | +| pretrained_model | 设置加载预训练模型路径 | ./pretrain_models/CRNN/best_accuracy | \ | +| checkpoints | 加载模型参数路径 | None | 用于中断后加载参数继续训练 | +| use_visualdl | 设置是否启用visualdl进行可视化log展示 | False | [教程地址](https://www.paddlepaddle.org.cn/paddle/visualdl) | +| infer_img | 设置预测图像路径或文件夹路径 | ./infer_img | \| +| character_dict_path | 设置字典路径 | ./ppocr/utils/ppocr_keys_v1.txt | \ | | max_text_length | 设置文本最大长度 | 25 | \ | | character_type | 设置字符类型 | ch | en/ch, en时将使用默认dict，ch时使用自定义dict| -| character_dict_path | 设置字典路径 | ./ppocr/utils/ic15_dict.txt | \ | -| loss_type | 设置 loss 类型 | ctc | 支持两种loss： ctc / attention | -| distort | 设置是否使用数据增强 | false | 设置为true时，将在训练时随机进行扰动，支持的扰动操作可阅读[img_tools.py](https://github.com/PaddlePaddle/PaddleOCR/blob/develop/ppocr/data/rec/img_tools.py) | -| use_space_char | 设置是否识别空格 | false | 仅在 character_type=ch 时支持空格 | +| use_space_char | 设置是否识别空格 | True | 仅在 character_type=ch 时支持空格 | | label_list | 设置方向分类器支持的角度 | ['0','180'] | 仅在方向分类器中生效 | -| average_window | ModelAverage优化器中的窗口长度计算比例 | 0.15 | 目前仅应用与SRN | -| max_average_window | 平均值计算窗口长度的最大值 | 15625 | 推荐设置为一轮训练中mini-batchs的数目| -| min_average_window | 平均值计算窗口长度的最小值 | 10000 | \ | -| reader_yml | 设置reader配置文件 | ./configs/rec/rec_icdar15_reader.yml | \ | -| pretrain_weights | 加载预训练模型路径 | ./pretrain_models/CRNN/best_accuracy | \ | -| checkpoints | 加载模型参数路径 | None | 用于中断后加载参数继续训练 | -| save_inference_dir | inference model 保存路径 | None | 用于保存inference model | +| save_res_path | 设置检测模型的结果保存地址 | ./output/det_db/predicts_db.txt | 仅在检测模型中生效 | -## 配置文件 Reader 系列参数介绍 +### Optimizer ([ppocr/optimizer](../../ppocr/optimizer)) -以 `rec_chinese_reader.yml` 为例 +| 字段 | 用途 | 默认值 | 备注 | +| :---------------------: | :---------------------: | :--------------: | :--------------------: | +| name | 优化器类名 | Adam | 目前支持`Momentum`,`Adam`,`RMSProp`, 见[ppocr/optimizer/optimizer.py](../../ppocr/optimizer/optimizer.py) | +| beta1 | 设置一阶矩估计的指数衰减率 | 0.9 | \ | +| beta2 | 设置二阶矩估计的指数衰减率 | 0.999 | \ | +| **lr** | 设置学习率decay方式 | - | \ | +| name | 学习率decay类名 | Cosine | 目前支持`Linear`,`Cosine`,`Step`,`Piecewise`, 见[ppocr/optimizer/learning_rate.py](../../ppocr/optimizer/learning_rate.py) | +| learning_rate | 基础学习率 | 0.001 | \ | +| **regularizer** | 设置网络正则化方式 | - | \ | +| name | 正则化类名 | L2 | 目前支持`L1`,`L2`, 见[ppocr/optimizer/regularizer.py](../../ppocr/optimizer/regularizer.py) | +| factor | 学习率衰减系数 | 0.00004 | \ | -| 字段 | 用途 | 默认值 | 备注 | -| :----------------------: | :---------------------: | :--------------: | :--------------------: | -| reader_function | 选择数据读取方式 | ppocr.data.rec.dataset_traversal,SimpleReader | 支持SimpleReader / LMDBReader 两种数据读取方式 | -| num_workers | 设置数据读取线程数 | 8 | \ | -| img_set_dir | 数据集路径 | ./train_data | \ | -| label_file_path | 数据标签路径 | ./train_data/rec_gt_train.txt| \ | -| infer_img | 预测图像文件夹路径 | ./infer_img | \| -## 配置文件 Optimizer 系列参数介绍 +### Architecture ([ppocr/modeling](../../ppocr/modeling)) +在ppocr中，网络被划分为Transform,Backbone,Neck和Head四个阶段 + +| 字段 | 用途 | 默认值 | 备注 | +| :---------------------: | :---------------------: | :--------------: | :--------------------: | +| model_type | 网络类型 | rec | 目前支持`rec`,`det`,`cls` | +| algorithm | 模型名称 | CRNN | 支持列表见[algorithm_overview](./algorithm_overview.md) | +| **Transform** | 设置变换方式 | - | 目前仅rec类型的算法支持, 具体见[ppocr/modeling/transform](../../ppocr/modeling/transform) | +| name | 变换方式类名 | TPS | 目前支持`TPS` | +| num_fiducial | TPS控制点数 | 20 | 上下边各十个 | +| loc_lr | 定位网络学习率 | 0.1 | \ | +| model_name | 定位网络大小 | small | 目前支持`small`,`large` | +| **Backbone** | 设置网络backbone类名 | - | 具体见[ppocr/modeling/backbones](../../ppocr/modeling/backbones) | +| name | backbone类名 | ResNet | 目前支持`MobileNetV3`,`ResNet` | +| layers | resnet层数 | 34 | 支持18,34,50,101,152,200 | +| model_name | MobileNetV3 网络大小 | small | 支持`small`,`large` | +| **Neck** | 设置网络neck | - | 具体见[ppocr/modeling/necks](../../ppocr/modeling/necks) | +| name | neck类名 | SequenceEncoder | 目前支持`SequenceEncoder`,`DBFPN` | +| encoder_type | SequenceEncoder编码器类型 | rnn | 支持`reshape`,`fc`,`rnn` | +| hidden_size | rnn内部单元数 | 48 | \ | +| out_channels | DBFPN输出通道数 | 256 | \ | +| **Head** | 设置网络Head | - | 具体见[ppocr/modeling/heads](../../ppocr/modeling/heads) | +| name | head类名 | CTCHead | 目前支持`CTCHead`,`DBHead`,`ClsHead` | +| fc_decay | CTCHead正则化系数 | 0.0004 | \ | +| k | DBHead二值化系数 | 50 | \ | +| class_dim | ClsHead输出分类数 | 2 | \ | + -以 `rec_icdar15_train.yml` 为例 +### Loss ([ppocr/losses](../../ppocr/losses)) + +| 字段 | 用途 | 默认值 | 备注 | +| :---------------------: | :---------------------: | :--------------: | :--------------------: | +| name | 网络loss类名 | CTCLoss | 目前支持`CTCLoss`,`DBLoss`,`ClsLoss` | +| balance_loss | DBLossloss中是否对正负样本数量进行均衡(使用OHEM) | True | \ | +| ohem_ratio | DBLossloss中的OHEM的负正样本比例 | 3 | \ | +| main_loss_type | DBLossloss中shrink_map所采用的的loss | DiceLoss | 支持`DiceLoss`,`BCELoss` | +| alpha | DBLossloss中shrink_map_loss的系数 | 5 | \ | +| beta | DBLossloss中threshold_map_loss的系数 | 10 | \ | + +### PostProcess ([ppocr/postprocess](../../ppocr/postprocess)) + +| 字段 | 用途 | 默认值 | 备注 | +| :---------------------: | :---------------------: | :--------------: | :--------------------: | +| name | 后处理类名 | CTCLabelDecode | 目前支持`CTCLoss`,`AttnLabelDecode`,`DBPostProcess`,`ClsPostProcess` | +| thresh | DBPostProcess中分割图进行二值化的阈值 | 0.3 | \ | +| box_thresh | DBPostProcess中对输出框进行过滤的阈值，低于此阈值的框不会输出 | 0.7 | \ | +| max_candidates | DBPostProcess中输出的最大文本框数量 | 1000 | | +| unclip_ratio | DBPostProcess中对文本框进行放大的比例 | 2.0 | \ | + +### Metric ([ppocr/metrics](../../ppocr/metrics)) + +| 字段 | 用途 | 默认值 | 备注 | +| :---------------------: | :---------------------: | :--------------: | :--------------------: | +| name | 指标评估方法名称 | CTCLabelDecode | 目前支持`DetMetric`,`RecMetric`,`ClsMetric` | +| main_indicator | 主要指标,用于选取最优模型 | acc | 对于检测方法为hmean，识别和分类方法为acc | +### Dataset ([ppocr/data](../../ppocr/data)) | 字段 | 用途 | 默认值 | 备注 | | :---------------------: | :---------------------: | :--------------: | :--------------------: | -| function | 选择优化器 | pocr.optimizer,AdamDecay | 目前只支持Adam方式 | -| base_lr | 设置初始学习率 | 0.0005 | \ | -| beta1 | 设置一阶矩估计的指数衰减率 | 0.9 | \ | -| beta2 | 设置二阶矩估计的指数衰减率 | 0.999 | \ | -| decay | 是否使用decay | \ | \ | -| function(decay) | 设置decay方式 | - | 目前支持cosine_decay, cosine_decay_warmup与piecewise_decay | -| step_each_epoch | 每个epoch包含多少次迭代, cosine_decay/cosine_decay_warmup时有效 | 20 | 计算方式：total_image_num / (batch_size_per_card * card_size) | -| total_epoch | 总共迭代多少个epoch, cosine_decay/cosine_decay_warmup时有效 | 1000 | 与Global.epoch_num 一致 | -| warmup_minibatch | 线性warmup的迭代次数, cosine_decay_warmup时有效 | 1000 | \ | -| boundaries | 学习率下降时的迭代次数间隔, piecewise_decay时有效 | - | 参数为列表形式 | -| decay_rate | 学习率衰减系数, piecewise_decay时有效 | - | \ | +| **dataset** | 每次迭代返回一个样本 | - | - | +| name | dataset类名 | SimpleDataSet | 目前支持`SimpleDataSet`和`LMDBDateSet` | +| data_dir | 数据集图片存放路径 | ./train_data | \ | +| label_file_list | 数据标签路径 | ["./train_data/train_list.txt"] | dataset为LMDBDateSet时不需要此参数 | +| ratio_list | 数据集的比例 | [1.0] | 若label_file_list中有两个train_list，且ratio_list为[0.4,0.6]，则从train_list1中采样40%，从train_list2中采样60%组合整个dataset | +| transforms | 对图片和标签进行变换的方法列表 | [DecodeImage,CTCLabelEncode,RecResizeImg,KeepKeys] | 见[ppocr/data/imaug](../../ppocr/data/imaug) | +| **loader** | dataloader相关 | - | | +| shuffle | 每个epoch是否将数据集顺序打乱 | True | \ | +| batch_size_per_card | 训练时单卡batch size | 256 | \ | +| drop_last | 是否丢弃因数据集样本数不能被 batch_size 整除而产生的最后一个不完整的mini-batch | True | \ | +| num_workers | 用于加载数据的子进程个数，若为0即为不开启子进程，在主进程中进行数据加载 | 8 | \ | \ No newline at end of file diff --git a/doc/doc_ch/tree.md b/doc/doc_ch/tree.md index ba8503d119a27b2cac34b593243caa68404b95ec..511860869188020ccfe7531db55ecc5008ccb56b 100644 --- a/doc/doc_ch/tree.md +++ b/doc/doc_ch/tree.md @@ -4,38 +4,34 @@ PaddleOCR 的整体目录结构介绍如下： ``` PaddleOCR -├── configs // 配置文件，可通过yml文件选择模型结构并修改超参 -│ ├── cls // 方向分类器相关配置文件 -│ │ ├── cls_mv3.yml // 训练配置相关，包括骨干网络、head、loss、优化器 -│ │ └── cls_reader.yml // 数据读取相关，数据读取方式、数据存储路径 -│ ├── det // 检测相关配置文件 -│ │ ├── det_db_icdar15_reader.yml // 数据读取 -│ │ ├── det_mv3_db.yml // 训练配置 +├── configs // 配置文件，可通过 yml 文件选择模型结构并修改超参 +│ ├── cls // 方向分类器相关配置文件 +│ │ ├── cls_mv3.yml // 训练配置相关，包括骨干网络、head、loss、优化器和数据 +│ ├── det // 检测相关配置文件 +│ │ ├── det_mv3_db.yml // 训练配置 │ │ ... -│ └── rec // 识别相关配置文件 -│ ├── rec_benchmark_reader.yml // LMDB 格式数据读取相关 -│ ├── rec_chinese_common_train.yml // 通用中文训练配置 -│ ├── rec_icdar15_reader.yml // simple 数据读取相关，包括数据读取函数、数据路径、标签文件 +│ └── rec // 识别相关配置文件 +│ ├── rec_mv3_none_bilstm_ctc.yml // crnn 训练配置 │ ... -├── deploy // 部署相关 -│ ├── android_demo // android_demo +├── deploy // 部署相关 +│ ├── android_demo // android_demo │ │ ... -│ ├── cpp_infer // C++ infer -│ │ ├── CMakeLists.txt // Cmake 文件 -│ │ ├── docs // 说明文档 +│ ├── cpp_infer // C++ infer +│ │ ├── CMakeLists.txt // Cmake 文件 +│ │ ├── docs // 说明文档 │ │ │ └── windows_vs2019_build.md -│ │ ├── include // 头文件 -│ │ │ ├── clipper.h // clipper 库 -│ │ │ ├── config.h // 预测配置 -│ │ │ ├── ocr_cls.h // 方向分类器 -│ │ │ ├── ocr_det.h // 文字检测 -│ │ │ ├── ocr_rec.h // 文字识别 -│ │ │ ├── postprocess_op.h // 检测后处理 -│ │ │ ├── preprocess_op.h // 检测预处理 -│ │ │ └── utility.h // 工具 -│ │ ├── readme.md // 说明文档 +│ │ ├── include // 头文件 +│ │ │ ├── clipper.h // clipper 库 +│ │ │ ├── config.h // 预测配置 +│ │ │ ├── ocr_cls.h // 方向分类器 +│ │ │ ├── ocr_det.h // 文字检测 +│ │ │ ├── ocr_rec.h // 文字识别 +│ │ │ ├── postprocess_op.h // 检测后处理 +│ │ │ ├── preprocess_op.h // 检测预处理 +│ │ │ └── utility.h // 工具 +│ │ ├── readme.md // 说明文档 │ │ ├── ... -│ │ ├── src // 源文件 +│ │ ├── src // 源文件 │ │ │ ├── clipper.cpp │ │ │ ├── config.cpp │ │ │ ├── main.cpp @@ -45,10 +41,10 @@ PaddleOCR │ │ │ ├── postprocess_op.cpp │ │ │ ├── preprocess_op.cpp │ │ │ └── utility.cpp -│ │ └── tools // 编译、执行脚本 -│ │ ├── build.sh // 编译脚本 -│ │ ├── config.txt // 配置文件 -│ │ └── run.sh // 测试启动脚本 +│ │ └── tools // 编译、执行脚本 +│ │ ├── build.sh // 编译脚本 +│ │ ├── config.txt // 配置文件 +│ │ └── run.sh // 测试启动脚本 │ ├── docker │ │ └── hubserving │ │ ├── cpu @@ -58,151 +54,163 @@ PaddleOCR │ │ ├── README_cn.md │ │ ├── README.md │ │ └── sample_request.txt -│ ├── hubserving // hubserving -│ │ ├── ocr_det // 文字检测 -│ │ │ ├── config.json // serving 配置 +│ ├── hubserving // hubserving +│ │ ├── ocr_cls // 方向分类器 +│ │ │ ├── config.json // serving 配置 │ │ │ ├── __init__.py -│ │ │ ├── module.py // 预测模型 -│ │ │ └── params.py // 预测参数 -│ │ ├── ocr_rec // 文字识别 +│ │ │ ├── module.py // 预测模型 +│ │ │ └── params.py // 预测参数 +│ │ ├── ocr_det // 文字检测 +│ │ │ ├── config.json // serving 配置 +│ │ │ ├── __init__.py +│ │ │ ├── module.py // 预测模型 +│ │ │ └── params.py // 预测参数 +│ │ ├── ocr_rec // 文字识别 │ │ │ ├── config.json │ │ │ ├── __init__.py │ │ │ ├── module.py │ │ │ └── params.py -│ │ └── ocr_system // 系统预测 +│ │ └── ocr_system // 系统预测 │ │ ├── config.json │ │ ├── __init__.py │ │ ├── module.py │ │ └── params.py -│ ├── imgs // 预测图片 +│ ├── imgs // 预测图片 │ │ ├── cpp_infer_pred_12.png │ │ └── demo.png -│ ├── ios_demo // ios demo +│ ├── ios_demo // ios demo │ │ ... -│ ├── lite // lite 部署 -│ │ ├── cls_process.cc // 方向分类器数据处理 +│ ├── lite // lite 部署 +│ │ ├── cls_process.cc // 方向分类器数据处理 │ │ ├── cls_process.h -│ │ ├── config.txt // 检测配置参数 -│ │ ├── crnn_process.cc // crnn数据处理 +│ │ ├── config.txt // 检测配置参数 +│ │ ├── crnn_process.cc // crnn 数据处理 │ │ ├── crnn_process.h -│ │ ├── db_post_process.cc // db数据处理 +│ │ ├── db_post_process.cc // db 数据处理 │ │ ├── db_post_process.h -│ │ ├── Makefile // 编译文件 -│ │ ├── ocr_db_crnn.cc // 串联预测 -│ │ ├── prepare.sh // 数据准备 -│ │ ├── readme.md // 说明文档 +│ │ ├── Makefile // 编译文件 +│ │ ├── ocr_db_crnn.cc // 串联预测 +│ │ ├── prepare.sh // 数据准备 +│ │ ├── readme.md // 说明文档 │ │ ... -│ ├── pdserving // pdserving 部署 -│ │ ├── det_local_server.py // 检测 快速版，部署方便预测速度快 -│ │ ├── det_web_server.py // 检测 完整版，稳定性高分布式部署 -│ │ ├── ocr_local_server.py // 检测+识别 快速版 -│ │ ├── ocr_web_client.py // 客户端 -│ │ ├── ocr_web_server.py // 检测+识别 完整版 -│ │ ├── readme.md // 说明文档 -│ │ ├── rec_local_server.py // 识别 快速版 -│ │ └── rec_web_server.py // 识别 完整版 +│ ├── pdserving // pdserving 部署 +│ │ ├── det_local_server.py // 检测 快速版，部署方便预测速度快 +│ │ ├── det_web_server.py // 检测 完整版，稳定性高分布式部署 +│ │ ├── ocr_local_server.py // 检测+识别 快速版 +│ │ ├── ocr_web_client.py // 客户端 +│ │ ├── ocr_web_server.py // 检测+识别 完整版 +│ │ ├── readme.md // 说明文档 +│ │ ├── rec_local_server.py // 识别 快速版 +│ │ └── rec_web_server.py // 识别 完整版 │ └── slim -│ └── quantization // 量化相关 -│ ├── export_model.py // 导出模型 -│ ├── quant.py // 量化 -│ └── README.md // 说明文档 -├── doc // 文档教程 +│ └── quantization // 量化相关 +│ ├── export_model.py // 导出模型 +│ ├── quant.py // 量化 +│ └── README.md // 说明文档 +├── doc // 文档教程 │ ... -├── paddleocr.py -├── ppocr // 网络核心代码 -│ ├── data // 数据处理 -│ │ ├── cls // 方向分类器 -│ │ │ ├── dataset_traversal.py // 数据传输，定义数据读取器，读取数据并组成batch -│ │ │ └── randaugment.py // 随机数据增广操作 -│ │ ├── det // 检测 -│ │ │ ├── data_augment.py // 数据增广操作 -│ │ │ ├── dataset_traversal.py // 数据传输，定义数据读取器，读取数据并组成batch -│ │ │ ├── db_process.py // db 数据处理 -│ │ │ ├── east_process.py // east 数据处理 -│ │ │ ├── make_border_map.py // 生成边界图 -│ │ │ ├── make_shrink_map.py // 生成收缩图 -│ │ │ ├── random_crop_data.py // 随机切割 -│ │ │ └── sast_process.py // sast 数据处理 -│ │ ├── reader_main.py // 数据读取器主函数 -│ │ └── rec // 识别 -│ │ ├── dataset_traversal.py // 数据传输，定义数据读取器，包含 LMDB_Reader 和 Simple_Reader -│ │ └── img_tools.py // 数据处理相关，包括数据归一化、扰动 -│ ├── __init__.py -│ ├── modeling // 组网相关 -│ │ ├── architectures // 模型架构，定义模型所需的各个模块 -│ │ │ ├── cls_model.py // 方向分类器 -│ │ │ ├── det_model.py // 检测 -│ │ │ └── rec_model.py // 识别 -│ │ ├── backbones // 骨干网络 -│ │ │ ├── det_mobilenet_v3.py // 检测 mobilenet_v3 -│ │ │ ├── det_resnet_vd.py -│ │ │ ├── det_resnet_vd_sast.py -│ │ │ ├── rec_mobilenet_v3.py // 识别 mobilenet_v3 -│ │ │ ├── rec_resnet_fpn.py -│ │ │ └── rec_resnet_vd.py -│ │ ├── common_functions.py // 公共函数 -│ │ ├── heads // 头函数 -│ │ │ ├── cls_head.py // 分类头 -│ │ │ ├── det_db_head.py // db 检测头 -│ │ │ ├── det_east_head.py // east 检测头 -│ │ │ ├── det_sast_head.py // sast 检测头 -│ │ │ ├── rec_attention_head.py // 识别 attention -│ │ │ ├── rec_ctc_head.py // 识别 ctc -│ │ │ ├── rec_seq_encoder.py // 识别 序列编码 -│ │ │ ├── rec_srn_all_head.py // 识别 srn 相关 -│ │ │ └── self_attention // srn attention -│ │ │ └── model.py -│ │ ├── losses // 损失函数 -│ │ │ ├── cls_loss.py // 方向分类器损失函数 -│ │ │ ├── det_basic_loss.py // 检测基础loss -│ │ │ ├── det_db_loss.py // DB loss -│ │ │ ├── det_east_loss.py // EAST loss -│ │ │ ├── det_sast_loss.py // SAST loss -│ │ │ ├── rec_attention_loss.py // attention loss -│ │ │ ├── rec_ctc_loss.py // ctc loss -│ │ │ └── rec_srn_loss.py // srn loss -│ │ └── stns // 空间变换网络 -│ │ └── tps.py // TPS 变换 -│ ├── optimizer.py // 优化器 -│ ├── postprocess // 后处理 -│ │ ├── db_postprocess.py // DB 后处理 -│ │ ├── east_postprocess.py // East 后处理 -│ │ ├── lanms // lanms 相关 -│ │ │ ... -│ │ ├── locality_aware_nms.py // nms -│ │ └── sast_postprocess.py // sast 后处理 -│ └── utils // 工具 -│ ├── character.py // 字符处理，包括对文本的编码和解码，计算预测准确率 -│ ├── check.py // 参数加载检查 -│ ├── ic15_dict.txt // 英文数字字典，区分大小写 -│ ├── ppocr_keys_v1.txt // 中文字典，用于训练中文模型 -│ ├── save_load.py // 模型保存和加载函数 -│ ├── stats.py // 统计 -│ └── utility.py // 工具函数，包含输入参数是否合法等相关检查工具 -├── README_en.md // 说明文档 -├── README.md -├── requirments.txt // 安装依赖 -├── setup.py // whl包打包脚本 -└── tools // 启动工具 - ├── eval.py // 评估函数 - ├── eval_utils // 评估工具 - │ ├── eval_cls_utils.py // 分类相关 - │ ├── eval_det_iou.py // 检测 iou 相关 - │ ├── eval_det_utils.py // 检测相关 - │ ├── eval_rec_utils.py // 识别相关 - │ └── __init__.py - ├── export_model.py // 导出 infer 模型 - ├── infer // 基于预测引擎预测 - │ ├── predict_cls.py - │ ├── predict_det.py - │ ├── predict_rec.py - │ ├── predict_system.py - │ └── utility.py - ├── infer_cls.py // 基于训练引擎 预测分类 - ├── infer_det.py // 基于训练引擎 预测检测 - ├── infer_rec.py // 基于训练引擎 预测识别 - ├── program.py // 整体流程 - ├── test_hubserving.py - └── train.py // 启动训练 - -``` +├── ppocr // 网络核心代码 +│ ├── data // 数据处理 +│ │ ├── imaug // 图片和 label 处理代码 +│ │ │ ├── text_image_aug // 文本识别的 tia 数据扩充 +│ │ │ │ ├── __init__.py +│ │ │ │ ├── augment.py // tia_distort,tia_stretch 和 tia_perspective 的代码 +│ │ │ │ ├── warp_mls.py +│ │ │ ├── __init__.py +│ │ │ ├── east_process.py // EAST 算法的数据处理步骤 +│ │ │ ├── make_border_map.py // 生成边界图 +│ │ │ ├── make_shrink_map.py // 生成收缩图 +│ │ │ ├── operators.py // 图像基本操作，如读取和归一化 +│ │ │ ├── randaugment.py // 随机数据增广操作 +│ │ │ ├── random_crop_data.py // 随机裁剪 +│ │ │ ├── rec_img_aug.py // 文本识别的数据扩充 +│ │ │ └── sast_process.py // SAST 算法的数据处理步骤 +│ │ ├── __init__.py // 构造 dataloader 相关代码 +│ │ ├── lmdb_dataset.py // 读取lmdb数据集的 dataset +│ │ ├── simple_dataset.py // 读取文本格式存储数据集的 dataset +│ ├── losses // 损失函数 +│ │ ├── __init__.py // 构造 loss 相关代码 +│ │ ├── cls_loss.py // 方向分类器 loss +│ │ ├── det_basic_loss.py // 检测基础 loss +│ │ ├── det_db_loss.py // DB loss +│ │ ├── det_east_loss.py // EAST loss +│ │ ├── det_sast_loss.py // SAST loss +│ │ ├── rec_ctc_loss.py // CTC loss +│ │ ├── rec_att_loss.py // Attention loss +│ ├── metrics // 评估指标 +│ │ ├── __init__.py // 构造 metric 相关代码 +│ │ ├── cls_metric.py // 方向分类器 metric +│ │ ├── det_metric.py // 检测 metric + │ ├── eval_det_iou.py // 检测 iou 相关 +│ │ ├── rec_metric.py // 识别 metric +│ ├── modeling // 组网相关 +│ │ ├── architectures // 网络 +│ │ │ ├── __init__.py // 构造 model 相关代码 +│ │ │ ├── base_model.py // 组网代码 +│ │ ├── backbones // 骨干网络 +│ │ │ ├── __init__.py // 构造 backbone 相关代码 +│ │ │ ├── det_mobilenet_v3.py // 检测 mobilenet_v3 +│ │ │ ├── det_resnet_vd.py // 检测 resnet +│ │ │ ├── det_resnet_vd_sast.py // 检测 SAST算法的resnet backbone +│ │ │ ├── rec_mobilenet_v3.py // 识别 mobilenet_v3 +│ │ │ └── rec_resnet_vd.py // 识别 resnet +│ │ ├── necks // 颈函数 +│ │ │ ├── __init__.py // 构造 neck 相关代码 +│ │ │ ├── db_fpn.py // 标准 fpn 网络 +│ │ │ ├── east_fpn.py // EAST 算法的 fpn 网络 +│ │ │ ├── sast_fpn.py // SAST 算法的 fpn 网络 +│ │ │ ├── rnn.py // 识别 序列编码 +│ │ ├── heads // 头函数 +│ │ │ ├── __init__.py // 构造 head 相关代码 +│ │ │ ├── cls_head.py // 方向分类器 分类头 +│ │ │ ├── det_db_head.py // DB 检测头 +│ │ │ ├── det_east_head.py // EAST 检测头 +│ │ │ ├── det_sast_head.py // SAST 检测头 +│ │ │ ├── rec_ctc_head.py // 识别 ctc +│ │ │ ├── rec_att_head.py // 识别 attention +│ │ ├── transforms // 图像变换 +│ │ │ ├── __init__.py // 构造 transform 相关代码 +│ │ │ └── tps.py // TPS 变换 +│ ├── optimizer // 优化器 +│ │ ├── __init__.py // 构造 optimizer 相关代码 +│ │ └── learning_rate.py // 学习率衰减 +│ │ └── optimizer.py // 优化器 +│ │ └── regularizer.py // 网络正则化 +│ ├── postprocess // 后处理 +│ │ ├── cls_postprocess.py // 方向分类器 后处理 +│ │ ├── db_postprocess.py // DB 后处理 +│ │ ├── east_postprocess.py // EAST 后处理 +│ │ ├── locality_aware_nms.py // NMS +│ │ ├── rec_postprocess.py // 识别网络 后处理 +│ │ └── sast_postprocess.py // SAST 后处理 +│ └── utils // 工具 +│ ├── dict // 小语种字典 +│ .... +│ ├── ic15_dict.txt // 英文数字字典，区分大小写 +│ ├── ppocr_keys_v1.txt // 中文字典，用于训练中文模型 +│ ├── logging.py // logger +│ ├── save_load.py // 模型保存和加载函数 +│ ├── stats.py // 统计 +│ └── utility.py // 工具函数 +├── tools +│ ├── eval.py // 评估函数 +│ ├── export_model.py // 导出 inference 模型 +│ ├── infer // 基于预测引擎预测 +│ │ ├── predict_cls.py +│ │ ├── predict_det.py +│ │ ├── predict_rec.py +│ │ ├── predict_system.py +│ │ └── utility.py +│ ├── infer_cls.py // 基于训练引擎 预测分类 +│ ├── infer_det.py // 基于训练引擎 预测检测 +│ ├── infer_rec.py // 基于训练引擎 预测识别 +│ ├── program.py // 整体流程 +│ ├── test_hubserving.py +│ └── train.py // 启动训练 +├── paddleocr.py +├── README_ch.md // 中文说明文档 +├── README_en.md // 英文说明文档 +├── README.md // 主页说明文档 +├── requirments.txt // 安装依赖 +├── setup.py // whl包打包脚本 +├── train.sh // 启动训练脚本 \ No newline at end of file diff --git a/doc/doc_en/add_new_algorithm_en.md b/doc/doc_en/add_new_algorithm_en.md new file mode 100644 index 0000000000000000000000000000000000000000..48b505b6994a99e66ce38ad2a59d47975534663b --- /dev/null +++ b/doc/doc_en/add_new_algorithm_en.md @@ -0,0 +1,304 @@ +# Add new algorithm + +PaddleOCR decomposes an algorithm into the following parts, and modularizes each part to make it more convenient to develop new algorithms. + +* Data loading and processing +* Network +* Post-processing +* Loss +* Metric +* Optimizer + +The following will introduce each part separately, and introduce how to add the modules required for the new algorithm. + + +## Data loading and processing + +Data loading and processing are composed of different modules, which complete the image reading, data augment and label production. This part is under [ppocr/data](../../ppocr/data). The explanation of each file and folder are as follows: + +```bash +ppocr/data/ +├── imaug # Scripts for image reading, data augment and label production +│ ├── label_ops.py # Modules that transform the label +│ ├── operators.py # Modules that transform the image +│ ├──..... +├── __init__.py +├── lmdb_dataset.py # The dataset that reads the lmdb +└── simple_dataset.py # Read the dataset saved in the form of `image_path\tgt` +``` + +PaddleOCR has a large number of built-in image operation related modules. For modules that are not built-in, you can add them through the following steps: + +1. Create a new file under the [ppocr/data/imaug](../../ppocr/data/imaug) folder, such as my_module.py. +2. Add code in the my_module.py file, the sample code is as follows: + +```python +class MyModule: + def __init__(self, *args, **kwargs): + # your init code + pass + + def __call__(self, data): + img = data['image'] + label = data['label'] + # your process code + + data['image'] = img + data['label'] = label + return data +``` + +3. Import the added module in the [ppocr/data/imaug/\__init\__.py](../../ppocr/data/imaug/__init__.py) file. + +All different modules of data processing are executed by sequence, combined and executed in the form of a list in the config file. Such as: + +```yaml +# angle class data process +transforms: + - DecodeImage: # load image + img_mode: BGR + channel_first: False + - MyModule: + args1: args1 + args2: args2 + - KeepKeys: + keep_keys: [ 'image', 'label' ] # dataloader will return list in this order +``` + +## Network + +The network part completes the construction of the network, and PaddleOCR divides the network into four parts, which are under [ppocr/modeling](../../ppocr/modeling). The data entering the network will pass through these four parts in sequence(transforms->backbones-> +necks->heads). + +```bash +├── architectures # Code for building network +├── transforms # Image Transformation Module +├── backbones # Feature extraction module +├── necks # Feature enhancement module +└── heads # Output module +``` + +PaddleOCR has built-in commonly used modules related to algorithms such as DB, EAST, SAST, CRNN and Attention. For modules that do not have built-in, you can add them through the following steps, the four parts are added in the same steps, take backbones as an example: + +1. Create a new file under the [ppocr/modeling/backbones](../../ppocr/modeling/backbones) folder, such as my_backbone.py. +2. Add code in the my_backbone.py file, the sample code is as follows: + +```python +import paddle +import paddle.nn as nn +import paddle.nn.functional as F + + +class MyBackbone(nn.Layer): + def __init__(self, *args, **kwargs): + super(MyBackbone, self).__init__() + # your init code + self.conv = nn.xxxx + + def forward(self, inputs): + # your necwork forward + y = self.conv(inputs) + return y +``` + +3. Import the added module in the [ppocr/modeling/backbones/\__init\__.py](../../ppocr/modeling/backbones/__init__.py) file. + +After adding the four-part modules of the network, you only need to configure them in the configuration file to use, such as: + +```yaml +Architecture: + model_type: rec + algorithm: CRNN + Transform: + name: MyTransform + args1: args1 + args2: args2 + Backbone: + name: MyBackbone + args1: args1 + Neck: + name: MyNeck + args1: args1 + Head: + name: MyHead + args1: args1 +``` + +## Post-processing + +Post-processing realizes decoding network output to obtain text box or recognized text. This part is under [ppocr/postprocess](../../ppocr/postprocess). +PaddleOCR has built-in post-processing modules related to algorithms such as DB, EAST, SAST, CRNN and Attention. For components that are not built-in, they can be added through the following steps: + +1. Create a new file under the [ppocr/postprocess](../../ppocr/postprocess) folder, such as my_postprocess.py. +2. Add code in the my_postprocess.py file, the sample code is as follows: + +```python +import paddle + + +class MyPostProcess: + def __init__(self, *args, **kwargs): + # your init code + pass + + def __call__(self, preds, label=None, *args, **kwargs): + if isinstance(preds, paddle.Tensor): + preds = preds.numpy() + # you preds decode code + preds = self.decode_preds(preds) + if label is None: + return preds + # you label decode code + label = self.decode_label(label) + return preds, label + + def decode_preds(self, preds): + # you preds decode code + pass + + def decode_label(self, preds): + # you label decode code + pass +``` + +3. Import the added module in the [ppocr/postprocess/\__init\__.py](../../ppocr/postprocess/__init__.py) file. + +After the post-processing module is added, you only need to configure it in the configuration file to use, such as: + +```yaml +PostProcess: + name: MyPostProcess + args1: args1 + args2: args2 +``` + +## Loss + +The loss function is used to calculate the distance between the network output and the label. This part is under [ppocr/losses](../../ppocr/losses). +PaddleOCR has built-in loss function modules related to algorithms such as DB, EAST, SAST, CRNN and Attention. For modules that do not have built-in modules, you can add them through the following steps: + +1. Create a new file in the [ppocr/losses](../../ppocr/losses) folder, such as my_loss.py. +2. Add code in the my_loss.py file, the sample code is as follows: + +```python +import paddle +from paddle import nn + + +class MyLoss(nn.Layer): + def __init__(self, **kwargs): + super(MyLoss, self).__init__() + # you init code + pass + + def __call__(self, predicts, batch): + label = batch[1] + # your loss code + loss = self.loss(input=predicts, label=label) + return {'loss': loss} +``` + +3. Import the added module in the [ppocr/losses/\__init\__.py](../../ppocr/losses/__init__.py) file. + +After the loss function module is added, you only need to configure it in the configuration file to use it, such as: + +```yaml +Loss: + name: MyLoss + args1: args1 + args2: args2 +``` + +## Metric + +Metric is used to calculate the performance of the network on the current batch. This part is under [ppocr/metrics](../../ppocr/metrics). PaddleOCR has built-in evaluation modules related to algorithms such as detection, classification and recognition. For modules that do not have built-in modules, you can add them through the following steps: + +1. Create a new file under the [ppocr/metrics](../../ppocr/metrics) folder, such as my_metric.py. +2. Add code in the my_metric.py file, the sample code is as follows: + +```python + +class MyMetric(object): + def __init__(self, main_indicator='acc', **kwargs): + # main_indicator is used for select best model + self.main_indicator = main_indicator + self.reset() + + def __call__(self, preds, batch, *args, **kwargs): + # preds is out of postprocess + # batch is out of dataloader + labels = batch[1] + cur_correct_num = 0 + cur_all_num = 0 + # you metric code + self.correct_num += cur_correct_num + self.all_num += cur_all_num + return {'acc': cur_correct_num / cur_all_num, } + + def get_metric(self): + """ + return metircs { + 'acc': 0, + 'norm_edit_dis': 0, + } + """ + acc = self.correct_num / self.all_num + self.reset() + return {'acc': acc} + + def reset(self): + # reset metric + self.correct_num = 0 + self.all_num = 0 + +``` + +3. Import the added module in the [ppocr/metrics/\__init\__.py](../../ppocr/metrics/__init__.py) file. + +After the metric module is added, you only need to configure it in the configuration file to use it, such as: + +```yaml +Metric: + name: MyMetric + main_indicator: acc +``` + +## 优化器 + +The optimizer is used to train the network. The optimizer also contains network regularization and learning rate decay modules. This part is under [ppocr/optimizer](../../ppocr/optimizer). PaddleOCR has built-in +Commonly used optimizer modules such as `Momentum`, `Adam` and `RMSProp`, common regularization modules such as `Linear`, `Cosine`, `Step` and `Piecewise`, and common learning rate decay modules such as `L1Decay` and `L2Decay`. +Modules without built-in can be added through the following steps, take `optimizer` as an example: + +1. Create your own optimizer in the [ppocr/optimizer/optimizer.py](../../ppocr/optimizer/optimizer.py) file, the sample code is as follows: + +```python +from paddle import optimizer as optim + + +class MyOptim(object): + def __init__(self, learning_rate=0.001, *args, **kwargs): + self.learning_rate = learning_rate + + def __call__(self, parameters): + # It is recommended to wrap the built-in optimizer of paddle + opt = optim.XXX( + learning_rate=self.learning_rate, + parameters=parameters) + return opt + +``` + +After the optimizer module is added, you only need to configure it in the configuration file to use, such as: + +```yaml +Optimizer: + name: MyOptim + args1: args1 + args2: args2 + lr: + name: Cosine + learning_rate: 0.001 + regularizer: + name: 'L2' + factor: 0 +``` \ No newline at end of file diff --git a/doc/doc_en/config_en.md b/doc/doc_en/config_en.md index 722da6620fd03912c48a47679e7c13a23f15752e..574bb41b6b1735271f9c794b856c5efb32db424f 100644 --- a/doc/doc_en/config_en.md +++ b/doc/doc_en/config_en.md @@ -1,69 +1,121 @@ -# OPTIONAL PARAMETERS LIST +## Optional parameter list -The following list can be viewed via `--help` +The following list can be viewed through `--help` | FLAG | Supported script | Use | Defaults | Note | | :----------------------: | :------------: | :---------------: | :--------------: | :-----------------: | -| -c | ALL | Specify configuration file to use | None | **Please refer to the parameter introduction for configuration file usage** | -| -o | ALL | set configuration options | None | Configuration using -o has higher priority than the configuration file selected with -c. E.g: `-o Global.use_gpu=false` | - +| -c | ALL | Specify configuration file to use | None | **Please refer to the parameter introduction for configuration file usage** | +| -o | ALL | set configuration options | None | Configuration using -o has higher priority than the configuration file selected with -c. E.g: -o Global.use_gpu=false | ## INTRODUCTION TO GLOBAL PARAMETERS OF CONFIGURATION FILE -Take `rec_chinese_lite_train_v1.1.yml` as an example - +Take rec_chinese_lite_train_v1.1.yml as an example +### Global -| Parameter | Use | Default | Note | +| Parameter | Use | Defaults | Note | | :----------------------: | :---------------------: | :--------------: | :--------------------: | -| algorithm | Select algorithm to use | Synchronize with configuration file | For selecting model, please refer to the supported model [list](https://github.com/PaddlePaddle/PaddleOCR/blob/develop/README_en.md) | -| use_gpu | Set using GPU or not | true | \ | -| epoch_num | Maximum training epoch number | 3000 | \ | +| use_gpu | Set using GPU or not | true | \ | +| epoch_num | Maximum training epoch number | 500 | \ | | log_smooth_window | Sliding window size | 20 | \ | | print_batch_step | Set print log interval | 10 | \ | -| save_model_dir | Set model save path | output/{model_name} | \ | +| save_model_dir | Set model save path | output/{算法名称} | \ | | save_epoch_step | Set model save interval | 3 | \ | -| eval_batch_step | Set the model evaluation interval |2000 or [1000, 2000] |runing evaluation every 2000 iters or evaluation is run every 2000 iterations after the 1000th iteration | -|train_batch_size_per_card | Set the batch size during training | 256 | \ | -| test_batch_size_per_card | Set the batch size during testing | 256 | \ | -| image_shape | Set input image size | [3, 32, 100] | \ | -| max_text_length | Set the maximum text length | 25 | \ | -| character_type | Set character type | ch | en/ch, the default dict will be used for en, and the custom dict will be used for ch| -| character_dict_path | Set dictionary path | ./ppocr/utils/ic15_dict.txt | \ | -| loss_type | Set loss type | ctc | Supports two types of loss: ctc / attention | -| distort | Set use distort | false | Support distort type ,read [img_tools.py](https://github.com/PaddlePaddle/PaddleOCR/blob/develop/ppocr/data/rec/img_tools.py) | -| use_space_char | Wether to recognize space | false | Only support in character_type=ch mode | - label_list | Set the angle supported by the direction classifier | ['0','180'] | Only valid in the direction classifier | -| reader_yml | Set the reader configuration file | ./configs/rec/rec_icdar15_reader.yml | \ | -| pretrain_weights | Load pre-trained model path | ./pretrain_models/CRNN/best_accuracy | \ | -| checkpoints | Load saved model path | None | Used to load saved parameters to continue training after interruption | -| save_inference_dir | path to save model for inference | None | Use to save inference model | - -## INTRODUCTION TO READER PARAMETERS OF CONFIGURATION FILE - -Take `rec_chinese_reader.yml` as an example: - -| Parameter | Use | Default | Note | -| :----------------------: | :---------------------: | :--------------: | :--------------------: | -| reader_function | Select data reading method | ppocr.data.rec.dataset_traversal,SimpleReader | Support two data reading methods: SimpleReader / LMDBReader | -| num_workers | Set the number of data reading threads | 8 | \ | -| img_set_dir | Image folder path | ./train_data | \ | -| label_file_path | Groundtruth file path | ./train_data/rec_gt_train.txt| \ | -| infer_img | Result folder path | ./infer_img | \| +| eval_batch_step | Set the model evaluation interval | 2000 or [1000, 2000] | runing evaluation every 2000 iters or evaluation is run every 2000 iterations after the 1000th iteration | +| cal_metric_during_train | Set whether to evaluate the metric during the training process. At this time, the metric of the model under the current batch is evaluated | true | \ | +| load_static_weights | Set whether the pre-training model is saved in static graph mode (currently only required by the detection algorithm) | true | \ | +| pretrained_model | Set the path of the pre-trained model | ./pretrain_models/CRNN/best_accuracy | \ | +| checkpoints | set model parameter path | None | Used to load parameters after interruption to continue training| +| use_visualdl | Set whether to enable visualdl for visual log display | False | [Tutorial](https://www.paddlepaddle.org.cn/paddle/visualdl) | +| infer_img | Set inference image path or folder path | ./infer_img | \| +| character_dict_path | Set dictionary path | ./ppocr/utils/ppocr_keys_v1.txt | \ | +| max_text_length | Set the maximum length of text | 25 | \ | +| character_type | Set character type | ch | en/ch, the default dict will be used for en, and the custom dict will be used for ch | +| use_space_char | Set whether to recognize spaces | True | Only support in character_type=ch mode | +| label_list | Set the angle supported by the direction classifier | ['0','180'] | Only valid in angle classifier model | +| save_res_path | Set the save address of the test model results | ./output/det_db/predicts_db.txt | Only valid in the text detection model | + +### Optimizer ([ppocr/optimizer](../../ppocr/optimizer)) + +| Parameter | Use | Defaults | Note | +| :---------------------: | :---------------------: | :--------------: | :--------------------: | +| name | Optimizer class name | Adam | Currently supports`Momentum`,`Adam`,`RMSProp`, see [ppocr/optimizer/optimizer.py](../../ppocr/optimizer/optimizer.py) | +| beta1 | Set the exponential decay rate for the 1st moment estimates | 0.9 | \ | +| beta2 | Set the exponential decay rate for the 2nd moment estimates | 0.999 | \ | +| **lr** | Set the learning rate decay method | - | \ | +| name | Learning rate decay class name | Cosine | Currently supports`Linear`,`Cosine`,`Step`,`Piecewise`, see[ppocr/optimizer/learning_rate.py](../../ppocr/optimizer/learning_rate.py) | +| learning_rate | Set the base learning rate | 0.001 | \ | +| **regularizer** | Set network regularization method | - | \ | +| name | Regularizer class name | L2 | Currently support`L1`,`L2`, see[ppocr/optimizer/regularizer.py](../../ppocr/optimizer/regularizer.py) | +| factor | Learning rate decay coefficient | 0.00004 | \ | + + +### Architecture ([ppocr/modeling](../../ppocr/modeling)) +In ppocr, the network is divided into four stages: Transform, Backbone, Neck and Head + +| Parameter | Use | Defaults | Note | +| :---------------------: | :---------------------: | :--------------: | :--------------------: | +| model_type | Network Type | rec | Currently support`rec`,`det`,`cls` | +| algorithm | Model name | CRNN | See [algorithm_overview](./algorithm_overview.md) for the support list | +| **Transform** | Set the transformation method | - | Currently only recognition algorithms are supported, see [ppocr/modeling/transform](../../ppocr/modeling/transform) for details | +| name | Transformation class name | TPS | Currently supports `TPS` | +| num_fiducial | Number of TPS control points | 20 | Ten on the top and bottom | +| loc_lr | Localization network learning rate | 0.1 | \ | +| model_name | Localization network size | small | Currently support`small`,`large` | +| **Backbone** | Set the network backbone class name | - | see [ppocr/modeling/backbones](../../ppocr/modeling/backbones) | +| name | backbone class name | ResNet | Currently support`MobileNetV3`,`ResNet` | +| layers | resnet layers | 34 | Currently support18,34,50,101,152,200 | +| model_name | MobileNetV3 network size | small | Currently support`small`,`large` | +| **Neck** | Set network neck | - | see[ppocr/modeling/necks](../../ppocr/modeling/necks) | +| name | neck class name | SequenceEncoder | Currently support`SequenceEncoder`,`DBFPN` | +| encoder_type | SequenceEncoder encoder type | rnn | Currently support`reshape`,`fc`,`rnn` | +| hidden_size | rnn number of internal units | 48 | \ | +| out_channels | Number of DBFPN output channels | 256 | \ | +| **Head** | Set the network head | - | see[ppocr/modeling/heads](../../ppocr/modeling/heads) | +| name | head class name | CTCHead | Currently support`CTCHead`,`DBHead`,`ClsHead` | +| fc_decay | CTCHead regularization coefficient | 0.0004 | \ | +| k | DBHead binarization coefficient | 50 | \ | +| class_dim | ClsHead output category number | 2 | \ | + + +### Loss ([ppocr/losses](../../ppocr/losses)) + +| Parameter | Use | Defaults | Note | +| :---------------------: | :---------------------: | :--------------: | :--------------------: | +| name | loss class name | CTCLoss | Currently support`CTCLoss`,`DBLoss`,`ClsLoss` | +| balance_loss | Whether to balance the number of positive and negative samples in DBLossloss (using OHEM) | True | \ | +| ohem_ratio | The negative and positive sample ratio of OHEM in DBLossloss | 3 | \ | +| main_loss_type | The loss used by shrink_map in DBLossloss | DiceLoss | Currently support`DiceLoss`,`BCELoss` | +| alpha | The coefficient of shrink_map_loss in DBLossloss | 5 | \ | +| beta | The coefficient of threshold_map_loss in DBLossloss | 10 | \ | -## INTRODUCTION TO OPTIMIZER PARAMETERS OF CONFIGURATION FILE +### PostProcess ([ppocr/postprocess](../../ppocr/postprocess)) -Take `rec_icdar15_train.yml` as an example: +| Parameter | Use | Defaults | Note | +| :---------------------: | :---------------------: | :--------------: | :--------------------: | +| name | Post-processing class name | CTCLabelDecode | Currently support`CTCLoss`,`AttnLabelDecode`,`DBPostProcess`,`ClsPostProcess` | +| thresh | The threshold for binarization of the segmentation map in DBPostProcess | 0.3 | \ | +| box_thresh | The threshold for filtering output boxes in DBPostProcess. Boxes below this threshold will not be output | 0.7 | \ | +| max_candidates | The maximum number of text boxes output in DBPostProcess | 1000 | | +| unclip_ratio | The unclip ratio of the text box in DBPostProcess | 2.0 | \ | + +### Metric ([ppocr/metrics](../../ppocr/metrics)) + +| Parameter | Use | Defaults | Note | +| :---------------------: | :---------------------: | :--------------: | :--------------------: | +| name | Metric method name | CTCLabelDecode | Currently support`DetMetric`,`RecMetric`,`ClsMetric` | +| main_indicator | Main indicators, used to select the best model | acc | For the detection method is hmean, the recognition and classification method is acc | -| Parameter | Use | Default | None | +### Dataset ([ppocr/data](../../ppocr/data)) +| Parameter | Use | Defaults | Note | | :---------------------: | :---------------------: | :--------------: | :--------------------: | -| function | Select Optimizer function | pocr.optimizer,AdamDecay | Only support Adam | -| base_lr | Set the base lr | 0.0005 | \ | -| beta1 | Set the exponential decay rate for the 1st moment estimates | 0.9 | \ | -| beta2 | Set the exponential decay rate for the 2nd moment estimates | 0.999 | \ | -| decay | Whether to use decay | \ | \ | -| function(decay) | Set the decay function | cosine_decay | Support cosine_decay, cosine_decay_warmup and piecewise_decay | -| step_each_epoch | The number of steps in an epoch. Used in cosine_decay/cosine_decay_warmup | 20 | Calculation: total_image_num / (batch_size_per_card * card_size) | -| total_epoch | The number of epochs. Used in cosine_decay/cosine_decay_warmup | 1000 | Consistent with Global.epoch_num | -| warmup_minibatch | Number of steps for linear warmup. Used in cosine_decay_warmup | 1000 | \ | -| boundaries | The step intervals to reduce learning rate. Used in piecewise_decay | - | The format is list | -| decay_rate | Learning rate decay rate. Used in piecewise_decay | - | \ | +| **dataset** | Return one sample per iteration | - | - | +| name | dataset class name | SimpleDataSet | Currently support`SimpleDataSet`,`LMDBDateSet` | +| data_dir | Image folder path | ./train_data | \ | +| label_file_list | Groundtruth file path | ["./train_data/train_list.txt"] | This parameter is not required when dataset is LMDBDateSet | +| ratio_list | Ratio of data set | [1.0] | If there are two train_lists in label_file_list and ratio_list is [0.4,0.6], 40% will be sampled from train_list1, and 60% will be sampled from train_list2 to combine the entire dataset | +| transforms | List of methods to transform images and labels | [DecodeImage,CTCLabelEncode,RecResizeImg,KeepKeys] | see[ppocr/data/imaug](../../ppocr/data/imaug) | +| **loader** | dataloader related | - | | +| shuffle | Does each epoch disrupt the order of the data set | True | \ | +| batch_size_per_card | Single card batch size during training | 256 | \ | +| drop_last | Whether to discard the last incomplete mini-batch because the number of samples in the data set cannot be divisible by batch_size | True | \ | +| num_workers | The number of sub-processes used to load data, if it is 0, the sub-process is not started, and the data is loaded in the main process | 8 | \ | \ No newline at end of file diff --git a/doc/doc_en/tree_en.md b/doc/doc_en/tree_en.md index 8f05de7e76b243bdcc6358bfea4e153e769dd3d5..be2f5733d3c30af8c83f1072cd0aea07865bacc3 100644 --- a/doc/doc_en/tree_en.md +++ b/doc/doc_en/tree_en.md @@ -2,40 +2,37 @@ The overall directory structure of PaddleOCR is introduced as follows: + ``` PaddleOCR -├── configs // configuration file, you can select model structure and modify hyperparameters through yml file -│ ├── cls // Related configuration files of direction classifier -│ │ ├── cls_mv3.yml // training configuration related, including backbone network, head, loss, optimizer -│ │ └── cls_reader.yml // Data reading related, data reading method, data storage path -│ ├── det // Detection related configuration files -│ │ ├── det_db_icdar15_reader.yml // data read -│ │ ├── det_mv3_db.yml // training configuration +├── configs // Configuration file, you can config the model structure and modify the hyperparameters through the yml file +│ ├── cls // Angle classifier config files +│ │ ├── cls_mv3.yml // Training config, including backbone network, head, loss, optimizer and data +│ ├── det // Text detection config files +│ │ ├── det_mv3_db.yml // Training config │ │ ... -│ └── rec // Identify related configuration files -│ ├── rec_benchmark_reader.yml // LMDB format data reading related -│ ├── rec_chinese_common_train.yml // General Chinese training configuration -│ ├── rec_icdar15_reader.yml // simple data reading related, including data reading function, data path, label file +│ └── rec // Text recognition config files +│ ├── rec_mv3_none_bilstm_ctc.yml // CRNN config │ ... -├── deploy // deployment related -│ ├── android_demo // android_demo +├── deploy // Depoly +│ ├── android_demo // Android demo │ │ ... -│ ├── cpp_infer // C++ infer -│ │ ├── CMakeLists.txt // Cmake file -│ │ ├── docs // documentation +│ ├── cpp_infer // C++ infer +│ │ ├── CMakeLists.txt // Cmake file +│ │ ├── docs // Docs │ │ │ └── windows_vs2019_build.md -│ │ ├── include -│ │ │ ├── clipper.h // clipper library -│ │ │ ├── config.h // infer configuration -│ │ │ ├── ocr_cls.h // direction classifier -│ │ │ ├── ocr_det.h // text detection -│ │ │ ├── ocr_rec.h // text recognition -│ │ │ ├── postprocess_op.h // postprocess after detection -│ │ │ ├── preprocess_op.h // preprocess detection -│ │ │ └── utility.h // tools -│ │ ├── readme.md // documentation +│ │ ├── include // Head Files +│ │ │ ├── clipper.h // clipper +│ │ │ ├── config.h // Inference config +│ │ │ ├── ocr_cls.h // Angle class +│ │ │ ├── ocr_det.h // Text detection +│ │ │ ├── ocr_rec.h // Text recognition +│ │ │ ├── postprocess_op.h // Post-processing +│ │ │ ├── preprocess_op.h // Pre-processing +│ │ │ └── utility.h // tools +│ │ ├── readme.md // Documentation │ │ ├── ... -│ │ ├── src // source file +│ │ ├── src // Source code files │ │ │ ├── clipper.cpp │ │ │ ├── config.cpp │ │ │ ├── main.cpp @@ -45,10 +42,10 @@ PaddleOCR │ │ │ ├── postprocess_op.cpp │ │ │ ├── preprocess_op.cpp │ │ │ └── utility.cpp -│ │ └── tools // compile and execute script -│ │ ├── build.sh // compile script -│ │ ├── config.txt // configuration file -│ │ └── run.sh // Test startup script +│ │ └── tools // Compile and execute script +│ │ ├── build.sh // Compile script +│ │ ├── config.txt // Config file +│ │ └── run.sh // Execute script │ ├── docker │ │ └── hubserving │ │ ├── cpu @@ -58,151 +55,165 @@ PaddleOCR │ │ ├── README_cn.md │ │ ├── README.md │ │ └── sample_request.txt -│ ├── hubserving // hubserving -│ │ ├── ocr_det // text detection -│ │ │ ├── config.json // serving configuration +│ ├── hubserving // hubserving +│ │ ├── ocr_cls // Angle class +│ │ │ ├── config.json // Serving config +│ │ │ ├── __init__.py +│ │ │ ├── module.py // Model +│ │ │ └── params.py // Parameters +│ │ ├── ocr_det // Text detection +│ │ │ ├── config.json // serving config │ │ │ ├── __init__.py -│ │ │ ├── module.py // prediction model -│ │ │ └── params.py // prediction parameters -│ │ ├── ocr_rec // text recognition +│ │ │ ├── module.py // Model +│ │ │ └── params.py // Parameters +│ │ ├── ocr_rec // Text recognition │ │ │ ├── config.json │ │ │ ├── __init__.py │ │ │ ├── module.py │ │ │ └── params.py -│ │ └── ocr_system // system forecast +│ │ └── ocr_system // Inference System │ │ ├── config.json │ │ ├── __init__.py │ │ ├── module.py │ │ └── params.py -│ ├── imgs // prediction picture +│ ├── imgs // Inference images │ │ ├── cpp_infer_pred_12.png │ │ └── demo.png -│ ├── ios_demo // ios demo +│ ├── ios_demo // IOS demo │ │ ... -│ ├── lite // lite deployment -│ │ ├── cls_process.cc // direction classifier data processing +│ ├── lite // Lite depoly +│ │ ├── cls_process.cc // Pre-process for angle class │ │ ├── cls_process.h -│ │ ├── config.txt // check configuration parameters -│ │ ├── crnn_process.cc // crnn data processing +│ │ ├── config.txt // Config file +│ │ ├── crnn_process.cc // Pre-process for CRNN │ │ ├── crnn_process.h -│ │ ├── db_post_process.cc // db data processing +│ │ ├── db_post_process.cc // Pre-process for DB │ │ ├── db_post_process.h -│ │ ├── Makefile // compile file -│ │ ├── ocr_db_crnn.cc // series prediction -│ │ ├── prepare.sh // data preparation -│ │ ├── readme.md // documentation +│ │ ├── Makefile // Compile file +│ │ ├── ocr_db_crnn.cc // Inference system +│ │ ├── prepare.sh // Prepare bash script +│ │ ├── readme.md // Documentation │ │ ... -│ ├── pdserving // pdserving deployment -│ │ ├── det_local_server.py // fast detection version, easy deployment and fast prediction -│ │ ├── det_web_server.py // Full version of detection, high stability and distributed deployment -│ │ ├── ocr_local_server.py // detection + identification quick version -│ │ ├── ocr_web_client.py // client -│ │ ├── ocr_web_server.py // detection + identification full version -│ │ ├── readme.md // documentation -│ │ ├── rec_local_server.py // recognize quick version -│ │ └── rec_web_server.py // Identify the full version +│ ├── pdserving // Pdserving depoly +│ │ ├── det_local_server.py // Text detection fast version, easy to deploy and fast to predict +│ │ ├── det_web_server.py // Text detection full version, high stability distributed deployment +│ │ ├── ocr_local_server.py // Text detection + recognition fast version +│ │ ├── ocr_web_client.py // client +│ │ ├── ocr_web_server.py // Text detection + recognition full version +│ │ ├── readme.md // Documentation +│ │ ├── rec_local_server.py // Text recognition fast version +│ │ └── rec_web_server.py // Text recognition full version │ └── slim -│ └── quantization // quantization related -│ ├── export_model.py // export model -│ ├── quant.py // quantization -│ └── README.md // Documentation -├── doc // Documentation tutorial +│ └── quantization // Quantization +│ ├── export_model.py // Export model +│ ├── quant.py // Quantization script +│ └── README.md // Documentation +├── doc // Documentation and Tutorials │ ... -├── paddleocr.py -├── ppocr // network core code -│ ├── data // data processing -│ │ ├── cls // direction classifier -│ │ │ ├── dataset_traversal.py // Data transmission, define data reader, read data and form batch -│ │ │ └── randaugment.py // Random data augmentation operation -│ │ ├── det // detection -│ │ │ ├── data_augment.py // data augmentation operation -│ │ │ ├── dataset_traversal.py // Data transmission, define data reader, read data and form batch -│ │ │ ├── db_process.py // db data processing -│ │ │ ├── east_process.py // east data processing -│ │ │ ├── make_border_map.py // Generate boundary map -│ │ │ ├── make_shrink_map.py // Generate shrink map -│ │ │ ├── random_crop_data.py // random crop -│ │ │ └── sast_process.py // sast data processing -│ │ ├── reader_main.py // main function of data reader -│ │ └── rec // recognation -│ │ ├── dataset_traversal.py // Data transmission, define data reader, including LMDB_Reader and Simple_Reader -│ │ └── img_tools.py // Data processing related, including data normalization and disturbance -│ ├── __init__.py -│ ├── modeling // networking related -│ │ ├── architectures // Model architecture, which defines the various modules required by the model -│ │ │ ├── cls_model.py // direction classifier -│ │ │ ├── det_model.py // detection -│ │ │ └── rec_model.py // recognition -│ │ ├── backbones // backbone network -│ │ │ ├── det_mobilenet_v3.py // detect mobilenet_v3 -│ │ │ ├── det_resnet_vd.py -│ │ │ ├── det_resnet_vd_sast.py -│ │ │ ├── rec_mobilenet_v3.py // recognize mobilenet_v3 -│ │ │ ├── rec_resnet_fpn.py -│ │ │ └── rec_resnet_vd.py -│ │ ├── common_functions.py // common functions -│ │ ├── heads -│ │ │ ├── cls_head.py // class header -│ │ │ ├── det_db_head.py // db detection head -│ │ │ ├── det_east_head.py // east detection head -│ │ │ ├── det_sast_head.py // sast detection head -│ │ │ ├── rec_attention_head.py // recognition attention -│ │ │ ├── rec_ctc_head.py // recognition ctc -│ │ │ ├── rec_seq_encoder.py // recognition sequence code -│ │ │ ├── rec_srn_all_head.py // srn related -│ │ │ └── self_attention // srn attention -│ │ │ └── model.py -│ │ ├── losses // loss function -│ │ │ ├── cls_loss.py // Directional classifier loss function -│ │ │ ├── det_basic_loss.py // detect basic loss -│ │ │ ├── det_db_loss.py // DB loss -│ │ │ ├── det_east_loss.py // EAST loss -│ │ │ ├── det_sast_loss.py // SAST loss -│ │ │ ├── rec_attention_loss.py // attention loss -│ │ │ ├── rec_ctc_loss.py // ctc loss -│ │ │ └── rec_srn_loss.py // srn loss -│ │ └── stns // Spatial transformation network -│ │ └── tps.py // TPS conversion -│ ├── optimizer.py // optimizer -│ ├── postprocess // post-processing -│ │ ├── db_postprocess.py // DB postprocess -│ │ ├── east_postprocess.py // East postprocess -│ │ ├── lanms // lanms related -│ │ │ ... -│ │ ├── locality_aware_nms.py // nms -│ │ └── sast_postprocess.py // sast post-processing -│ └── utils // tools -│ ├── character.py // Character processing, including text encoding and decoding, and calculation of prediction accuracy -│ ├── check.py // parameter loading check -│ ├── ic15_dict.txt // English number dictionary, case sensitive -│ ├── ppocr_keys_v1.txt // Chinese dictionary, used to train Chinese models -│ ├── save_load.py // model save and load function -│ ├── stats.py // Statistics -│ └── utility.py // Tool functions, including related check tools such as whether the input parameters are legal -├── README_en.md // documentation -├── README.md -├── requirments.txt // installation dependencies -├── setup.py // whl package packaging script -└── tools // start tool - ├── eval.py // evaluation function - ├── eval_utils // evaluation tools - │ ├── eval_cls_utils.py // category related - │ ├── eval_det_iou.py // detect iou related - │ ├── eval_det_utils.py // detection related - │ ├── eval_rec_utils.py // recognition related - │ └── __init__.py - ├── export_model.py // export infer model - ├── infer // Forecast based on prediction engine - │ ├── predict_cls.py - │ ├── predict_det.py - │ ├── predict_rec.py - │ ├── predict_system.py - │ └── utility.py - ├── infer_cls.py // Predict classification based on training engine - ├── infer_det.py // Predictive detection based on training engine - ├── infer_rec.py // Predictive recognition based on training engine - ├── program.py // overall process - ├── test_hubserving.py - └── train.py // start training - -``` +├── ppocr // Core code +│ ├── data // Data processing +│ │ ├── imaug // Image and label processing code +│ │ │ ├── text_image_aug // Tia data augment for text recognition +│ │ │ │ ├── __init__.py +│ │ │ │ ├── augment.py // Tia_distort,tia_stretch and tia_perspective +│ │ │ │ ├── warp_mls.py +│ │ │ ├── __init__.py +│ │ │ ├── east_process.py // Data processing steps of EAST algorithm +│ │ │ ├── iaa_augment.py // Data augmentation operations +│ │ │ ├── label_ops.py // label encode operations +│ │ │ ├── make_border_map.py // Generate boundary map +│ │ │ ├── make_shrink_map.py // Generate shrink graph +│ │ │ ├── operators.py // Basic image operations, such as reading and normalization +│ │ │ ├── randaugment.py // Random data augmentation operation +│ │ │ ├── random_crop_data.py // Random crop +│ │ │ ├── rec_img_aug.py // Data augmentation for text recognition +│ │ │ └── sast_process.py // Data processing steps of SAST algorithm +│ │ ├── __init__.py // Construct dataloader code +│ │ ├── lmdb_dataset.py // Read lmdb dataset +│ │ ├── simple_dataset.py // Read the dataset stored in text format +│ ├── losses // Loss function +│ │ ├── __init__.py // Construct loss code +│ │ ├── cls_loss.py // Angle class loss +│ │ ├── det_basic_loss.py // Text detection basic loss +│ │ ├── det_db_loss.py // DB loss +│ │ ├── det_east_loss.py // EAST loss +│ │ ├── det_sast_loss.py // SAST loss +│ │ ├── rec_ctc_loss.py // CTC loss +│ │ ├── rec_att_loss.py // Attention loss +│ ├── metrics // Metrics +│ │ ├── __init__.py // Construct metric code +│ │ ├── cls_metric.py // Angle class metric +│ │ ├── det_metric.py // Text detection metric + │ ├── eval_det_iou.py // Text detection iou code +│ │ ├── rec_metric.py // Text recognition metric +│ ├── modeling // Network +│ │ ├── architectures // Architecture +│ │ │ ├── __init__.py // Construct model code +│ │ │ ├── base_model.py // Base model +│ │ ├── backbones // backbones +│ │ │ ├── __init__.py // Construct backbone code +│ │ │ ├── det_mobilenet_v3.py // Text detection mobilenet_v3 +│ │ │ ├── det_resnet_vd.py // Text detection resnet +│ │ │ ├── det_resnet_vd_sast.py // Text detection resnet backbone of the SAST algorithm +│ │ │ ├── rec_mobilenet_v3.py // Text recognition mobilenet_v3 +│ │ │ └── rec_resnet_vd.py // Text recognition resnet +│ │ ├── necks // Necks +│ │ │ ├── __init__.py // Construct neck code +│ │ │ ├── db_fpn.py // Standard fpn +│ │ │ ├── east_fpn.py // EAST algorithm fpn network +│ │ │ ├── sast_fpn.py // SAST algorithm fpn network +│ │ │ ├── rnn.py // Character recognition sequence encoding +│ │ ├── heads // Heads +│ │ │ ├── __init__.py // Construct head code +│ │ │ ├── cls_head.py // Angle class head +│ │ │ ├── det_db_head.py // DB head +│ │ │ ├── det_east_head.py // EAST head +│ │ │ ├── det_sast_head.py // SAST head +│ │ │ ├── rec_ctc_head.py // CTC head +│ │ │ ├── rec_att_head.py // Attention head +│ │ ├── transforms // Transforms +│ │ │ ├── __init__.py // Construct transform code +│ │ │ └── tps.py // TPS transform +│ ├── optimizer // Optimizer +│ │ ├── __init__.py // Construct optimizer code +│ │ └── learning_rate.py // Learning rate decay +│ │ └── optimizer.py // Optimizer +│ │ └── regularizer.py // Network regularization +│ ├── postprocess // Post-processing +│ │ ├── cls_postprocess.py // Angle class post-processing +│ │ ├── db_postprocess.py // DB post-processing +│ │ ├── east_postprocess.py // EAST post-processing +│ │ ├── locality_aware_nms.py // NMS +│ │ ├── rec_postprocess.py // Text recognition post-processing +│ │ └── sast_postprocess.py // SAST post-processing +│ └── utils // utils +│ ├── dict // Minor language dictionary +│ .... +│ ├── ic15_dict.txt // English number dictionary, case sensitive +│ ├── ppocr_keys_v1.txt // Chinese dictionary for training Chinese models +│ ├── logging.py // logger +│ ├── save_load.py // Model saving and loading functions +│ ├── stats.py // Training status statistics +│ └── utility.py // Utility function +├── tools +│ ├── eval.py // Evaluation function +│ ├── export_model.py // Export inference model +│ ├── infer // Inference based on Inference engine +│ │ ├── predict_cls.py +│ │ ├── predict_det.py +│ │ ├── predict_rec.py +│ │ ├── predict_system.py +│ │ └── utility.py +│ ├── infer_cls.py // Angle classification inference based on training engine +│ ├── infer_det.py // Text detection inference based on training engine +│ ├── infer_rec.py // Text recognition inference based on training engine +│ ├── program.py // Inference system +│ ├── test_hubserving.py +│ └── train.py // Start training script +├── paddleocr.py +├── README_ch.md // Chinese documentation +├── README_en.md // English documentation +├── README.md // Home page documentation +├── requirments.txt // Requirments +├── setup.py // Whl package packaging script +├── train.sh // Start training bash script \ No newline at end of file diff --git a/ppocr/data/imaug/__init__.py b/ppocr/data/imaug/__init__.py index fd143b53202a6116df90a1df2b75b254a73bd5f5..6ea4dd8ed6d0f58fbee3362e8eb82a0eda65e812 100644 --- a/ppocr/data/imaug/__init__.py +++ b/ppocr/data/imaug/__init__.py @@ -26,6 +26,9 @@ from .randaugment import RandAugment from .operators import * from .label_ops import * +from .east_process import * +from .sast_process import * + def transform(data, ops=None): """ transform """ diff --git a/ppocr/data/imaug/east_process.py b/ppocr/data/imaug/east_process.py new file mode 100644 index 0000000000000000000000000000000000000000..b1d7a5e51939af981dd62c269c930f4bf9ba4179 --- /dev/null +++ b/ppocr/data/imaug/east_process.py @@ -0,0 +1,439 @@ +#copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve. +# +#Licensed under the Apache License, Version 2.0 (the "License"); +#you may not use this file except in compliance with the License. +#You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +#Unless required by applicable law or agreed to in writing, software +#distributed under the License is distributed on an "AS IS" BASIS, +#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +#See the License for the specific language governing permissions and +#limitations under the License. + +import math +import cv2 +import numpy as np +import json +import sys +import os + +__all__ = ['EASTProcessTrain'] + + +class EASTProcessTrain(object): + def __init__(self, + image_shape = [512, 512], + background_ratio = 0.125, + min_crop_side_ratio = 0.1, + min_text_size = 10, + **kwargs): + self.input_size = image_shape[1] + self.random_scale = np.array([0.5, 1, 2.0, 3.0]) + self.background_ratio = background_ratio + self.min_crop_side_ratio = min_crop_side_ratio + self.min_text_size = min_text_size + + def preprocess(self, im): + input_size = self.input_size + im_shape = im.shape + im_size_min = np.min(im_shape[0:2]) + im_size_max = np.max(im_shape[0:2]) + im_scale = float(input_size) / float(im_size_max) + im = cv2.resize(im, None, None, fx=im_scale, fy=im_scale) + img_mean = [0.485, 0.456, 0.406] + img_std = [0.229, 0.224, 0.225] + # im = im[:, :, ::-1].astype(np.float32) + im = im / 255 + im -= img_mean + im /= img_std + new_h, new_w, _ = im.shape + im_padded = np.zeros((input_size, input_size, 3), dtype=np.float32) + im_padded[:new_h, :new_w, :] = im + im_padded = im_padded.transpose((2, 0, 1)) + im_padded = im_padded[np.newaxis, :] + return im_padded, im_scale + + def rotate_im_poly(self, im, text_polys): + """ + rotate image with 90 / 180 / 270 degre + """ + im_w, im_h = im.shape[1], im.shape[0] + dst_im = im.copy() + dst_polys = [] + rand_degree_ratio = np.random.rand() + rand_degree_cnt = 1 + if 0.333 < rand_degree_ratio < 0.666: + rand_degree_cnt = 2 + elif rand_degree_ratio > 0.666: + rand_degree_cnt = 3 + for i in range(rand_degree_cnt): + dst_im = np.rot90(dst_im) + rot_degree = -90 * rand_degree_cnt + rot_angle = rot_degree * math.pi / 180.0 + n_poly = text_polys.shape[0] + cx, cy = 0.5 * im_w, 0.5 * im_h + ncx, ncy = 0.5 * dst_im.shape[1], 0.5 * dst_im.shape[0] + for i in range(n_poly): + wordBB = text_polys[i] + poly = [] + for j in range(4): + sx, sy = wordBB[j][0], wordBB[j][1] + dx = math.cos(rot_angle) * (sx - cx)\ + - math.sin(rot_angle) * (sy - cy) + ncx + dy = math.sin(rot_angle) * (sx - cx)\ + + math.cos(rot_angle) * (sy - cy) + ncy + poly.append([dx, dy]) + dst_polys.append(poly) + dst_polys = np.array(dst_polys, dtype=np.float32) + return dst_im, dst_polys + + def polygon_area(self, poly): + """ + compute area of a polygon + :param poly: + :return: + """ + edge = [(poly[1][0] - poly[0][0]) * (poly[1][1] + poly[0][1]), + (poly[2][0] - poly[1][0]) * (poly[2][1] + poly[1][1]), + (poly[3][0] - poly[2][0]) * (poly[3][1] + poly[2][1]), + (poly[0][0] - poly[3][0]) * (poly[0][1] + poly[3][1])] + return np.sum(edge) / 2. + + def check_and_validate_polys(self, polys, tags, img_height, img_width): + """ + check so that the text poly is in the same direction, + and also filter some invalid polygons + :param polys: + :param tags: + :return: + """ + h, w = img_height, img_width + if polys.shape[0] == 0: + return polys + polys[:, :, 0] = np.clip(polys[:, :, 0], 0, w - 1) + polys[:, :, 1] = np.clip(polys[:, :, 1], 0, h - 1) + + validated_polys = [] + validated_tags = [] + for poly, tag in zip(polys, tags): + p_area = self.polygon_area(poly) + #invalid poly + if abs(p_area) < 1: + continue + if p_area > 0: + #'poly in wrong direction' + if not tag: + tag = True #reversed cases should be ignore + poly = poly[(0, 3, 2, 1), :] + validated_polys.append(poly) + validated_tags.append(tag) + return np.array(validated_polys), np.array(validated_tags) + + def draw_img_polys(self, img, polys): + if len(img.shape) == 4: + img = np.squeeze(img, axis=0) + if img.shape[0] == 3: + img = img.transpose((1, 2, 0)) + img[:, :, 2] += 123.68 + img[:, :, 1] += 116.78 + img[:, :, 0] += 103.94 + cv2.imwrite("tmp.jpg", img) + img = cv2.imread("tmp.jpg") + for box in polys: + box = box.astype(np.int32).reshape((-1, 1, 2)) + cv2.polylines(img, [box], True, color=(255, 255, 0), thickness=2) + import random + ino = random.randint(0, 100) + cv2.imwrite("tmp_%d.jpg" % ino, img) + return + + def shrink_poly(self, poly, r): + """ + fit a poly inside the origin poly, maybe bugs here... + used for generate the score map + :param poly: the text poly + :param r: r in the paper + :return: the shrinked poly + """ + # shrink ratio + R = 0.3 + # find the longer pair + dist0 = np.linalg.norm(poly[0] - poly[1]) + dist1 = np.linalg.norm(poly[2] - poly[3]) + dist2 = np.linalg.norm(poly[0] - poly[3]) + dist3 = np.linalg.norm(poly[1] - poly[2]) + if dist0 + dist1 > dist2 + dist3: + # first move (p0, p1), (p2, p3), then (p0, p3), (p1, p2) + ## p0, p1 + theta = np.arctan2((poly[1][1] - poly[0][1]), + (poly[1][0] - poly[0][0])) + poly[0][0] += R * r[0] * np.cos(theta) + poly[0][1] += R * r[0] * np.sin(theta) + poly[1][0] -= R * r[1] * np.cos(theta) + poly[1][1] -= R * r[1] * np.sin(theta) + ## p2, p3 + theta = np.arctan2((poly[2][1] - poly[3][1]), + (poly[2][0] - poly[3][0])) + poly[3][0] += R * r[3] * np.cos(theta) + poly[3][1] += R * r[3] * np.sin(theta) + poly[2][0] -= R * r[2] * np.cos(theta) + poly[2][1] -= R * r[2] * np.sin(theta) + ## p0, p3 + theta = np.arctan2((poly[3][0] - poly[0][0]), + (poly[3][1] - poly[0][1])) + poly[0][0] += R * r[0] * np.sin(theta) + poly[0][1] += R * r[0] * np.cos(theta) + poly[3][0] -= R * r[3] * np.sin(theta) + poly[3][1] -= R * r[3] * np.cos(theta) + ## p1, p2 + theta = np.arctan2((poly[2][0] - poly[1][0]), + (poly[2][1] - poly[1][1])) + poly[1][0] += R * r[1] * np.sin(theta) + poly[1][1] += R * r[1] * np.cos(theta) + poly[2][0] -= R * r[2] * np.sin(theta) + poly[2][1] -= R * r[2] * np.cos(theta) + else: + ## p0, p3 + # print poly + theta = np.arctan2((poly[3][0] - poly[0][0]), + (poly[3][1] - poly[0][1])) + poly[0][0] += R * r[0] * np.sin(theta) + poly[0][1] += R * r[0] * np.cos(theta) + poly[3][0] -= R * r[3] * np.sin(theta) + poly[3][1] -= R * r[3] * np.cos(theta) + ## p1, p2 + theta = np.arctan2((poly[2][0] - poly[1][0]), + (poly[2][1] - poly[1][1])) + poly[1][0] += R * r[1] * np.sin(theta) + poly[1][1] += R * r[1] * np.cos(theta) + poly[2][0] -= R * r[2] * np.sin(theta) + poly[2][1] -= R * r[2] * np.cos(theta) + ## p0, p1 + theta = np.arctan2((poly[1][1] - poly[0][1]), + (poly[1][0] - poly[0][0])) + poly[0][0] += R * r[0] * np.cos(theta) + poly[0][1] += R * r[0] * np.sin(theta) + poly[1][0] -= R * r[1] * np.cos(theta) + poly[1][1] -= R * r[1] * np.sin(theta) + ## p2, p3 + theta = np.arctan2((poly[2][1] - poly[3][1]), + (poly[2][0] - poly[3][0])) + poly[3][0] += R * r[3] * np.cos(theta) + poly[3][1] += R * r[3] * np.sin(theta) + poly[2][0] -= R * r[2] * np.cos(theta) + poly[2][1] -= R * r[2] * np.sin(theta) + return poly + + def generate_quad(self, im_size, polys, tags): + """ + Generate quadrangle. + """ + h, w = im_size + poly_mask = np.zeros((h, w), dtype=np.uint8) + score_map = np.zeros((h, w), dtype=np.uint8) + # (x1, y1, ..., x4, y4, short_edge_norm) + geo_map = np.zeros((h, w, 9), dtype=np.float32) + # mask used during traning, to ignore some hard areas + training_mask = np.ones((h, w), dtype=np.uint8) + for poly_idx, poly_tag in enumerate(zip(polys, tags)): + poly = poly_tag[0] + tag = poly_tag[1] + + r = [None, None, None, None] + for i in range(4): + dist1 = np.linalg.norm(poly[i] - poly[(i + 1) % 4]) + dist2 = np.linalg.norm(poly[i] - poly[(i - 1) % 4]) + r[i] = min(dist1, dist2) + # score map + shrinked_poly = self.shrink_poly( + poly.copy(), r).astype(np.int32)[np.newaxis, :, :] + cv2.fillPoly(score_map, shrinked_poly, 1) + cv2.fillPoly(poly_mask, shrinked_poly, poly_idx + 1) + # if the poly is too small, then ignore it during training + poly_h = min( + np.linalg.norm(poly[0] - poly[3]), + np.linalg.norm(poly[1] - poly[2])) + poly_w = min( + np.linalg.norm(poly[0] - poly[1]), + np.linalg.norm(poly[2] - poly[3])) + if min(poly_h, poly_w) < self.min_text_size: + cv2.fillPoly(training_mask, + poly.astype(np.int32)[np.newaxis, :, :], 0) + + if tag: + cv2.fillPoly(training_mask, + poly.astype(np.int32)[np.newaxis, :, :], 0) + + xy_in_poly = np.argwhere(poly_mask == (poly_idx + 1)) + # geo map. + y_in_poly = xy_in_poly[:, 0] + x_in_poly = xy_in_poly[:, 1] + poly[:, 0] = np.minimum(np.maximum(poly[:, 0], 0), w) + poly[:, 1] = np.minimum(np.maximum(poly[:, 1], 0), h) + for pno in range(4): + geo_channel_beg = pno * 2 + geo_map[y_in_poly, x_in_poly, geo_channel_beg] =\ + x_in_poly - poly[pno, 0] + geo_map[y_in_poly, x_in_poly, geo_channel_beg+1] =\ + y_in_poly - poly[pno, 1] + geo_map[y_in_poly, x_in_poly, 8] = \ + 1.0 / max(min(poly_h, poly_w), 1.0) + return score_map, geo_map, training_mask + + def crop_area(self, + im, + polys, + tags, + crop_background=False, + max_tries=50): + """ + make random crop from the input image + :param im: + :param polys: + :param tags: + :param crop_background: + :param max_tries: + :return: + """ + h, w, _ = im.shape + pad_h = h // 10 + pad_w = w // 10 + h_array = np.zeros((h + pad_h * 2), dtype=np.int32) + w_array = np.zeros((w + pad_w * 2), dtype=np.int32) + for poly in polys: + poly = np.round(poly, decimals=0).astype(np.int32) + minx = np.min(poly[:, 0]) + maxx = np.max(poly[:, 0]) + w_array[minx + pad_w:maxx + pad_w] = 1 + miny = np.min(poly[:, 1]) + maxy = np.max(poly[:, 1]) + h_array[miny + pad_h:maxy + pad_h] = 1 + # ensure the cropped area not across a text + h_axis = np.where(h_array == 0)[0] + w_axis = np.where(w_array == 0)[0] + if len(h_axis) == 0 or len(w_axis) == 0: + return im, polys, tags + + for i in range(max_tries): + xx = np.random.choice(w_axis, size=2) + xmin = np.min(xx) - pad_w + xmax = np.max(xx) - pad_w + xmin = np.clip(xmin, 0, w - 1) + xmax = np.clip(xmax, 0, w - 1) + yy = np.random.choice(h_axis, size=2) + ymin = np.min(yy) - pad_h + ymax = np.max(yy) - pad_h + ymin = np.clip(ymin, 0, h - 1) + ymax = np.clip(ymax, 0, h - 1) + if xmax - xmin < self.min_crop_side_ratio * w or \ + ymax - ymin < self.min_crop_side_ratio * h: + # area too small + continue + if polys.shape[0] != 0: + poly_axis_in_area = (polys[:, :, 0] >= xmin)\ + & (polys[:, :, 0] <= xmax)\ + & (polys[:, :, 1] >= ymin)\ + & (polys[:, :, 1] <= ymax) + selected_polys = np.where( + np.sum(poly_axis_in_area, axis=1) == 4)[0] + else: + selected_polys = [] + + if len(selected_polys) == 0: + # no text in this area + if crop_background: + im = im[ymin:ymax + 1, xmin:xmax + 1, :] + polys = [] + tags = [] + return im, polys, tags + else: + continue + + im = im[ymin:ymax + 1, xmin:xmax + 1, :] + polys = polys[selected_polys] + tags = tags[selected_polys] + polys[:, :, 0] -= xmin + polys[:, :, 1] -= ymin + return im, polys, tags + return im, polys, tags + + def crop_background_infor(self, im, text_polys, text_tags): + im, text_polys, text_tags = self.crop_area( + im, text_polys, text_tags, crop_background=True) + + if len(text_polys) > 0: + return None + # pad and resize image + input_size = self.input_size + im, ratio = self.preprocess(im) + score_map = np.zeros((input_size, input_size), dtype=np.float32) + geo_map = np.zeros((input_size, input_size, 9), dtype=np.float32) + training_mask = np.ones((input_size, input_size), dtype=np.float32) + return im, score_map, geo_map, training_mask + + def crop_foreground_infor(self, im, text_polys, text_tags): + im, text_polys, text_tags = self.crop_area( + im, text_polys, text_tags, crop_background=False) + + if text_polys.shape[0] == 0: + return None + #continue for all ignore case + if np.sum((text_tags * 1.0)) >= text_tags.size: + return None + # pad and resize image + input_size = self.input_size + im, ratio = self.preprocess(im) + text_polys[:, :, 0] *= ratio + text_polys[:, :, 1] *= ratio + _, _, new_h, new_w = im.shape + # print(im.shape) + # self.draw_img_polys(im, text_polys) + score_map, geo_map, training_mask = self.generate_quad( + (new_h, new_w), text_polys, text_tags) + return im, score_map, geo_map, training_mask + + def __call__(self, data): + im = data['image'] + text_polys = data['polys'] + text_tags = data['ignore_tags'] + if im is None: + return None + if text_polys.shape[0] == 0: + return None + + #add rotate cases + if np.random.rand() < 0.5: + im, text_polys = self.rotate_im_poly(im, text_polys) + h, w, _ = im.shape + text_polys, text_tags = self.check_and_validate_polys(text_polys, + text_tags, h, w) + if text_polys.shape[0] == 0: + return None + + # random scale this image + rd_scale = np.random.choice(self.random_scale) + im = cv2.resize(im, dsize=None, fx=rd_scale, fy=rd_scale) + text_polys *= rd_scale + if np.random.rand() < self.background_ratio: + outs = self.crop_background_infor(im, text_polys, text_tags) + else: + outs = self.crop_foreground_infor(im, text_polys, text_tags) + + if outs is None: + return None + im, score_map, geo_map, training_mask = outs + score_map = score_map[np.newaxis, ::4, ::4].astype(np.float32) + geo_map = np.swapaxes(geo_map, 1, 2) + geo_map = np.swapaxes(geo_map, 1, 0) + geo_map = geo_map[:, ::4, ::4].astype(np.float32) + training_mask = training_mask[np.newaxis, ::4, ::4] + training_mask = training_mask.astype(np.float32) + + data['image'] = im[0] + data['score_map'] = score_map + data['geo_map'] = geo_map + data['training_mask'] = training_mask + # print(im.shape, score_map.shape, geo_map.shape, training_mask.shape) + return data \ No newline at end of file diff --git a/ppocr/data/imaug/label_ops.py b/ppocr/data/imaug/label_ops.py index 0b1d46c62404978563d24698018ff24d2d552f2f..af3308a553b40b61fb206a8970038e91b718c56f 100644 --- a/ppocr/data/imaug/label_ops.py +++ b/ppocr/data/imaug/label_ops.py @@ -52,6 +52,7 @@ class DetLabelEncode(object): txt_tags.append(True) else: txt_tags.append(False) + boxes = self.expand_points_num(boxes) boxes = np.array(boxes, dtype=np.float32) txt_tags = np.array(txt_tags, dtype=np.bool) @@ -70,6 +71,17 @@ class DetLabelEncode(object): rect[3] = pts[np.argmax(diff)] return rect + def expand_points_num(self, boxes): + max_points_num = 0 + for box in boxes: + if len(box) > max_points_num: + max_points_num = len(box) + ex_boxes = [] + for box in boxes: + ex_box = box + [box[-1]] * (max_points_num - len(box)) + ex_boxes.append(ex_box) + return ex_boxes + class BaseRecLabelEncode(object): """ Convert between text-label and text-index """ @@ -83,7 +95,7 @@ class BaseRecLabelEncode(object): 'ch', 'en', 'en_sensitive', 'french', 'german', 'japan', 'korean' ] assert character_type in support_character_type, "Only {} are supported now but get {}".format( - support_character_type, self.character_str) + support_character_type, character_type) self.max_text_len = max_text_length if character_type == "en": diff --git a/ppocr/data/imaug/operators.py b/ppocr/data/imaug/operators.py index 927aa6407efceffc2dfb60410f33cc1addf004ac..57cd3b4b172cc52d073999a007fe5a9a343c9186 100644 --- a/ppocr/data/imaug/operators.py +++ b/ppocr/data/imaug/operators.py @@ -122,26 +122,37 @@ class DetResizeForTest(object): if 'limit_side_len' in kwargs: self.limit_side_len = kwargs['limit_side_len'] self.limit_type = kwargs.get('limit_type', 'min') + if 'resize_long' in kwargs: + self.resize_type = 2 + self.resize_long = kwargs.get('resize_long', 960) else: self.limit_side_len = 736 self.limit_type = 'min' def __call__(self, data): img = data['image'] + src_h, src_w, _ = img.shape if self.resize_type == 0: - img, shape = self.resize_image_type0(img) + # img, shape = self.resize_image_type0(img) + img, [ratio_h, ratio_w] = self.resize_image_type0(img) + elif self.resize_type == 2: + img, [ratio_h, ratio_w] = self.resize_image_type2(img) else: - img, shape = self.resize_image_type1(img) + # img, shape = self.resize_image_type1(img) + img, [ratio_h, ratio_w] = self.resize_image_type1(img) data['image'] = img - data['shape'] = shape + data['shape'] = np.array([src_h, src_w, ratio_h, ratio_w]) return data def resize_image_type1(self, img): resize_h, resize_w = self.image_shape ori_h, ori_w = img.shape[:2] # (h, w, c) + ratio_h = float(resize_h) / ori_h + ratio_w = float(resize_w) / ori_w img = cv2.resize(img, (int(resize_w), int(resize_h))) - return img, np.array([ori_h, ori_w]) + # return img, np.array([ori_h, ori_w]) + return img, [ratio_h, ratio_w] def resize_image_type0(self, img): """ @@ -184,4 +195,31 @@ class DetResizeForTest(object): except: print(img.shape, resize_w, resize_h) sys.exit(0) - return img, np.array([h, w]) + ratio_h = resize_h / float(h) + ratio_w = resize_w / float(w) + # return img, np.array([h, w]) + return img, [ratio_h, ratio_w] + + def resize_image_type2(self, img): + h, w, _ = img.shape + + resize_w = w + resize_h = h + + # Fix the longer side + if resize_h > resize_w: + ratio = float(self.resize_long) / resize_h + else: + ratio = float(self.resize_long) / resize_w + + resize_h = int(resize_h * ratio) + resize_w = int(resize_w * ratio) + + max_stride = 128 + resize_h = (resize_h + max_stride - 1) // max_stride * max_stride + resize_w = (resize_w + max_stride - 1) // max_stride * max_stride + img = cv2.resize(img, (int(resize_w), int(resize_h))) + ratio_h = resize_h / float(h) + ratio_w = resize_w / float(w) + + return img, [ratio_h, ratio_w] diff --git a/ppocr/data/imaug/sast_process.py b/ppocr/data/imaug/sast_process.py new file mode 100644 index 0000000000000000000000000000000000000000..b8d6ff89eed5e5061f6f07def25d4fef53125810 --- /dev/null +++ b/ppocr/data/imaug/sast_process.py @@ -0,0 +1,689 @@ +#copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve. +# +#Licensed under the Apache License, Version 2.0 (the "License"); +#you may not use this file except in compliance with the License. +#You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +#Unless required by applicable law or agreed to in writing, software +#distributed under the License is distributed on an "AS IS" BASIS, +#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +#See the License for the specific language governing permissions and +#limitations under the License. + +import math +import cv2 +import numpy as np +import json +import sys +import os + +__all__ = ['SASTProcessTrain'] + + +class SASTProcessTrain(object): + def __init__(self, + image_shape = [512, 512], + min_crop_size = 24, + min_crop_side_ratio = 0.3, + min_text_size = 10, + max_text_size = 512, + **kwargs): + self.input_size = image_shape[1] + self.min_crop_size = min_crop_size + self.min_crop_side_ratio = min_crop_side_ratio + self.min_text_size = min_text_size + self.max_text_size = max_text_size + + def quad_area(self, poly): + """ + compute area of a polygon + :param poly: + :return: + """ + edge = [ + (poly[1][0] - poly[0][0]) * (poly[1][1] + poly[0][1]), + (poly[2][0] - poly[1][0]) * (poly[2][1] + poly[1][1]), + (poly[3][0] - poly[2][0]) * (poly[3][1] + poly[2][1]), + (poly[0][0] - poly[3][0]) * (poly[0][1] + poly[3][1]) + ] + return np.sum(edge) / 2. + + def gen_quad_from_poly(self, poly): + """ + Generate min area quad from poly. + """ + point_num = poly.shape[0] + min_area_quad = np.zeros((4, 2), dtype=np.float32) + if True: + rect = cv2.minAreaRect(poly.astype(np.int32)) # (center (x,y), (width, height), angle of rotation) + center_point = rect[0] + box = np.array(cv2.boxPoints(rect)) + + first_point_idx = 0 + min_dist = 1e4 + for i in range(4): + dist = np.linalg.norm(box[(i + 0) % 4] - poly[0]) + \ + np.linalg.norm(box[(i + 1) % 4] - poly[point_num // 2 - 1]) + \ + np.linalg.norm(box[(i + 2) % 4] - poly[point_num // 2]) + \ + np.linalg.norm(box[(i + 3) % 4] - poly[-1]) + if dist < min_dist: + min_dist = dist + first_point_idx = i + for i in range(4): + min_area_quad[i] = box[(first_point_idx + i) % 4] + + return min_area_quad + + def check_and_validate_polys(self, polys, tags, xxx_todo_changeme): + """ + check so that the text poly is in the same direction, + and also filter some invalid polygons + :param polys: + :param tags: + :return: + """ + (h, w) = xxx_todo_changeme + if polys.shape[0] == 0: + return polys, np.array([]), np.array([]) + polys[:, :, 0] = np.clip(polys[:, :, 0], 0, w - 1) + polys[:, :, 1] = np.clip(polys[:, :, 1], 0, h - 1) + + validated_polys = [] + validated_tags = [] + hv_tags = [] + for poly, tag in zip(polys, tags): + quad = self.gen_quad_from_poly(poly) + p_area = self.quad_area(quad) + if abs(p_area) < 1: + print('invalid poly') + continue + if p_area > 0: + if tag == False: + print('poly in wrong direction') + tag = True # reversed cases should be ignore + poly = poly[(0, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1), :] + quad = quad[(0, 3, 2, 1), :] + + len_w = np.linalg.norm(quad[0] - quad[1]) + np.linalg.norm(quad[3] - quad[2]) + len_h = np.linalg.norm(quad[0] - quad[3]) + np.linalg.norm(quad[1] - quad[2]) + hv_tag = 1 + + if len_w * 2.0 < len_h: + hv_tag = 0 + + validated_polys.append(poly) + validated_tags.append(tag) + hv_tags.append(hv_tag) + return np.array(validated_polys), np.array(validated_tags), np.array(hv_tags) + + def crop_area(self, im, polys, tags, hv_tags, crop_background=False, max_tries=25): + """ + make random crop from the input image + :param im: + :param polys: + :param tags: + :param crop_background: + :param max_tries: 50 -> 25 + :return: + """ + h, w, _ = im.shape + pad_h = h // 10 + pad_w = w // 10 + h_array = np.zeros((h + pad_h * 2), dtype=np.int32) + w_array = np.zeros((w + pad_w * 2), dtype=np.int32) + for poly in polys: + poly = np.round(poly, decimals=0).astype(np.int32) + minx = np.min(poly[:, 0]) + maxx = np.max(poly[:, 0]) + w_array[minx + pad_w: maxx + pad_w] = 1 + miny = np.min(poly[:, 1]) + maxy = np.max(poly[:, 1]) + h_array[miny + pad_h: maxy + pad_h] = 1 + # ensure the cropped area not across a text + h_axis = np.where(h_array == 0)[0] + w_axis = np.where(w_array == 0)[0] + if len(h_axis) == 0 or len(w_axis) == 0: + return im, polys, tags, hv_tags + for i in range(max_tries): + xx = np.random.choice(w_axis, size=2) + xmin = np.min(xx) - pad_w + xmax = np.max(xx) - pad_w + xmin = np.clip(xmin, 0, w - 1) + xmax = np.clip(xmax, 0, w - 1) + yy = np.random.choice(h_axis, size=2) + ymin = np.min(yy) - pad_h + ymax = np.max(yy) - pad_h + ymin = np.clip(ymin, 0, h - 1) + ymax = np.clip(ymax, 0, h - 1) + # if xmax - xmin < ARGS.min_crop_side_ratio * w or \ + # ymax - ymin < ARGS.min_crop_side_ratio * h: + if xmax - xmin < self.min_crop_size or \ + ymax - ymin < self.min_crop_size: + # area too small + continue + if polys.shape[0] != 0: + poly_axis_in_area = (polys[:, :, 0] >= xmin) & (polys[:, :, 0] <= xmax) \ + & (polys[:, :, 1] >= ymin) & (polys[:, :, 1] <= ymax) + selected_polys = np.where(np.sum(poly_axis_in_area, axis=1) == 4)[0] + else: + selected_polys = [] + if len(selected_polys) == 0: + # no text in this area + if crop_background: + return im[ymin : ymax + 1, xmin : xmax + 1, :], \ + polys[selected_polys], tags[selected_polys], hv_tags[selected_polys], txts + else: + continue + im = im[ymin: ymax + 1, xmin: xmax + 1, :] + polys = polys[selected_polys] + tags = tags[selected_polys] + hv_tags = hv_tags[selected_polys] + polys[:, :, 0] -= xmin + polys[:, :, 1] -= ymin + return im, polys, tags, hv_tags + + return im, polys, tags, hv_tags + + def generate_direction_map(self, poly_quads, direction_map): + """ + """ + width_list = [] + height_list = [] + for quad in poly_quads: + quad_w = (np.linalg.norm(quad[0] - quad[1]) + np.linalg.norm(quad[2] - quad[3])) / 2.0 + quad_h = (np.linalg.norm(quad[0] - quad[3]) + np.linalg.norm(quad[2] - quad[1])) / 2.0 + width_list.append(quad_w) + height_list.append(quad_h) + norm_width = max(sum(width_list) / (len(width_list) + 1e-6), 1.0) + average_height = max(sum(height_list) / (len(height_list) + 1e-6), 1.0) + + for quad in poly_quads: + direct_vector_full = ((quad[1] + quad[2]) - (quad[0] + quad[3])) / 2.0 + direct_vector = direct_vector_full / (np.linalg.norm(direct_vector_full) + 1e-6) * norm_width + direction_label = tuple(map(float, [direct_vector[0], direct_vector[1], 1.0 / (average_height + 1e-6)])) + cv2.fillPoly(direction_map, quad.round().astype(np.int32)[np.newaxis, :, :], direction_label) + return direction_map + + def calculate_average_height(self, poly_quads): + """ + """ + height_list = [] + for quad in poly_quads: + quad_h = (np.linalg.norm(quad[0] - quad[3]) + np.linalg.norm(quad[2] - quad[1])) / 2.0 + height_list.append(quad_h) + average_height = max(sum(height_list) / len(height_list), 1.0) + return average_height + + def generate_tcl_label(self, hw, polys, tags, ds_ratio, + tcl_ratio=0.3, shrink_ratio_of_width=0.15): + """ + Generate polygon. + """ + h, w = hw + h, w = int(h * ds_ratio), int(w * ds_ratio) + polys = polys * ds_ratio + + score_map = np.zeros((h, w,), dtype=np.float32) + tbo_map = np.zeros((h, w, 5), dtype=np.float32) + training_mask = np.ones((h, w,), dtype=np.float32) + direction_map = np.ones((h, w, 3)) * np.array([0, 0, 1]).reshape([1, 1, 3]).astype(np.float32) + + for poly_idx, poly_tag in enumerate(zip(polys, tags)): + poly = poly_tag[0] + tag = poly_tag[1] + + # generate min_area_quad + min_area_quad, center_point = self.gen_min_area_quad_from_poly(poly) + min_area_quad_h = 0.5 * (np.linalg.norm(min_area_quad[0] - min_area_quad[3]) + + np.linalg.norm(min_area_quad[1] - min_area_quad[2])) + min_area_quad_w = 0.5 * (np.linalg.norm(min_area_quad[0] - min_area_quad[1]) + + np.linalg.norm(min_area_quad[2] - min_area_quad[3])) + + if min(min_area_quad_h, min_area_quad_w) < self.min_text_size * ds_ratio \ + or min(min_area_quad_h, min_area_quad_w) > self.max_text_size * ds_ratio: + continue + + if tag: + # continue + cv2.fillPoly(training_mask, poly.astype(np.int32)[np.newaxis, :, :], 0.15) + else: + tcl_poly = self.poly2tcl(poly, tcl_ratio) + tcl_quads = self.poly2quads(tcl_poly) + poly_quads = self.poly2quads(poly) + # stcl map + stcl_quads, quad_index = self.shrink_poly_along_width(tcl_quads, shrink_ratio_of_width=shrink_ratio_of_width, + expand_height_ratio=1.0 / tcl_ratio) + # generate tcl map + cv2.fillPoly(score_map, np.round(stcl_quads).astype(np.int32), 1.0) + + # generate tbo map + for idx, quad in enumerate(stcl_quads): + quad_mask = np.zeros((h, w), dtype=np.float32) + quad_mask = cv2.fillPoly(quad_mask, np.round(quad[np.newaxis, :, :]).astype(np.int32), 1.0) + tbo_map = self.gen_quad_tbo(poly_quads[quad_index[idx]], quad_mask, tbo_map) + return score_map, tbo_map, training_mask + + def generate_tvo_and_tco(self, hw, polys, tags, tcl_ratio=0.3, ds_ratio=0.25): + """ + Generate tcl map, tvo map and tbo map. + """ + h, w = hw + h, w = int(h * ds_ratio), int(w * ds_ratio) + polys = polys * ds_ratio + poly_mask = np.zeros((h, w), dtype=np.float32) + + tvo_map = np.ones((9, h, w), dtype=np.float32) + tvo_map[0:-1:2] = np.tile(np.arange(0, w), (h, 1)) + tvo_map[1:-1:2] = np.tile(np.arange(0, w), (h, 1)).T + poly_tv_xy_map = np.zeros((8, h, w), dtype=np.float32) + + # tco map + tco_map = np.ones((3, h, w), dtype=np.float32) + tco_map[0] = np.tile(np.arange(0, w), (h, 1)) + tco_map[1] = np.tile(np.arange(0, w), (h, 1)).T + poly_tc_xy_map = np.zeros((2, h, w), dtype=np.float32) + + poly_short_edge_map = np.ones((h, w), dtype=np.float32) + + for poly, poly_tag in zip(polys, tags): + + if poly_tag == True: + continue + + # adjust point order for vertical poly + poly = self.adjust_point(poly) + + # generate min_area_quad + min_area_quad, center_point = self.gen_min_area_quad_from_poly(poly) + min_area_quad_h = 0.5 * (np.linalg.norm(min_area_quad[0] - min_area_quad[3]) + + np.linalg.norm(min_area_quad[1] - min_area_quad[2])) + min_area_quad_w = 0.5 * (np.linalg.norm(min_area_quad[0] - min_area_quad[1]) + + np.linalg.norm(min_area_quad[2] - min_area_quad[3])) + + # generate tcl map and text, 128 * 128 + tcl_poly = self.poly2tcl(poly, tcl_ratio) + + # generate poly_tv_xy_map + for idx in range(4): + cv2.fillPoly(poly_tv_xy_map[2 * idx], + np.round(tcl_poly[np.newaxis, :, :]).astype(np.int32), + float(min(max(min_area_quad[idx, 0], 0), w))) + cv2.fillPoly(poly_tv_xy_map[2 * idx + 1], + np.round(tcl_poly[np.newaxis, :, :]).astype(np.int32), + float(min(max(min_area_quad[idx, 1], 0), h))) + + # generate poly_tc_xy_map + for idx in range(2): + cv2.fillPoly(poly_tc_xy_map[idx], + np.round(tcl_poly[np.newaxis, :, :]).astype(np.int32), float(center_point[idx])) + + # generate poly_short_edge_map + cv2.fillPoly(poly_short_edge_map, + np.round(tcl_poly[np.newaxis, :, :]).astype(np.int32), + float(max(min(min_area_quad_h, min_area_quad_w), 1.0))) + + # generate poly_mask and training_mask + cv2.fillPoly(poly_mask, np.round(tcl_poly[np.newaxis, :, :]).astype(np.int32), 1) + + tvo_map *= poly_mask + tvo_map[:8] -= poly_tv_xy_map + tvo_map[-1] /= poly_short_edge_map + tvo_map = tvo_map.transpose((1, 2, 0)) + + tco_map *= poly_mask + tco_map[:2] -= poly_tc_xy_map + tco_map[-1] /= poly_short_edge_map + tco_map = tco_map.transpose((1, 2, 0)) + + return tvo_map, tco_map + + def adjust_point(self, poly): + """ + adjust point order. + """ + point_num = poly.shape[0] + if point_num == 4: + len_1 = np.linalg.norm(poly[0] - poly[1]) + len_2 = np.linalg.norm(poly[1] - poly[2]) + len_3 = np.linalg.norm(poly[2] - poly[3]) + len_4 = np.linalg.norm(poly[3] - poly[0]) + + if (len_1 + len_3) * 1.5 < (len_2 + len_4): + poly = poly[[1, 2, 3, 0], :] + + elif point_num > 4: + vector_1 = poly[0] - poly[1] + vector_2 = poly[1] - poly[2] + cos_theta = np.dot(vector_1, vector_2) / (np.linalg.norm(vector_1) * np.linalg.norm(vector_2) + 1e-6) + theta = np.arccos(np.round(cos_theta, decimals=4)) + + if abs(theta) > (70 / 180 * math.pi): + index = list(range(1, point_num)) + [0] + poly = poly[np.array(index), :] + return poly + + def gen_min_area_quad_from_poly(self, poly): + """ + Generate min area quad from poly. + """ + point_num = poly.shape[0] + min_area_quad = np.zeros((4, 2), dtype=np.float32) + if point_num == 4: + min_area_quad = poly + center_point = np.sum(poly, axis=0) / 4 + else: + rect = cv2.minAreaRect(poly.astype(np.int32)) # (center (x,y), (width, height), angle of rotation) + center_point = rect[0] + box = np.array(cv2.boxPoints(rect)) + + first_point_idx = 0 + min_dist = 1e4 + for i in range(4): + dist = np.linalg.norm(box[(i + 0) % 4] - poly[0]) + \ + np.linalg.norm(box[(i + 1) % 4] - poly[point_num // 2 - 1]) + \ + np.linalg.norm(box[(i + 2) % 4] - poly[point_num // 2]) + \ + np.linalg.norm(box[(i + 3) % 4] - poly[-1]) + if dist < min_dist: + min_dist = dist + first_point_idx = i + + for i in range(4): + min_area_quad[i] = box[(first_point_idx + i) % 4] + + return min_area_quad, center_point + + def shrink_quad_along_width(self, quad, begin_width_ratio=0., end_width_ratio=1.): + """ + Generate shrink_quad_along_width. + """ + ratio_pair = np.array([[begin_width_ratio], [end_width_ratio]], dtype=np.float32) + p0_1 = quad[0] + (quad[1] - quad[0]) * ratio_pair + p3_2 = quad[3] + (quad[2] - quad[3]) * ratio_pair + return np.array([p0_1[0], p0_1[1], p3_2[1], p3_2[0]]) + + def shrink_poly_along_width(self, quads, shrink_ratio_of_width, expand_height_ratio=1.0): + """ + shrink poly with given length. + """ + upper_edge_list = [] + + def get_cut_info(edge_len_list, cut_len): + for idx, edge_len in enumerate(edge_len_list): + cut_len -= edge_len + if cut_len <= 0.000001: + ratio = (cut_len + edge_len_list[idx]) / edge_len_list[idx] + return idx, ratio + + for quad in quads: + upper_edge_len = np.linalg.norm(quad[0] - quad[1]) + upper_edge_list.append(upper_edge_len) + + # length of left edge and right edge. + left_length = np.linalg.norm(quads[0][0] - quads[0][3]) * expand_height_ratio + right_length = np.linalg.norm(quads[-1][1] - quads[-1][2]) * expand_height_ratio + + shrink_length = min(left_length, right_length, sum(upper_edge_list)) * shrink_ratio_of_width + # shrinking length + upper_len_left = shrink_length + upper_len_right = sum(upper_edge_list) - shrink_length + + left_idx, left_ratio = get_cut_info(upper_edge_list, upper_len_left) + left_quad = self.shrink_quad_along_width(quads[left_idx], begin_width_ratio=left_ratio, end_width_ratio=1) + right_idx, right_ratio = get_cut_info(upper_edge_list, upper_len_right) + right_quad = self.shrink_quad_along_width(quads[right_idx], begin_width_ratio=0, end_width_ratio=right_ratio) + + out_quad_list = [] + if left_idx == right_idx: + out_quad_list.append([left_quad[0], right_quad[1], right_quad[2], left_quad[3]]) + else: + out_quad_list.append(left_quad) + for idx in range(left_idx + 1, right_idx): + out_quad_list.append(quads[idx]) + out_quad_list.append(right_quad) + + return np.array(out_quad_list), list(range(left_idx, right_idx + 1)) + + def vector_angle(self, A, B): + """ + Calculate the angle between vector AB and x-axis positive direction. + """ + AB = np.array([B[1] - A[1], B[0] - A[0]]) + return np.arctan2(*AB) + + def theta_line_cross_point(self, theta, point): + """ + Calculate the line through given point and angle in ax + by + c =0 form. + """ + x, y = point + cos = np.cos(theta) + sin = np.sin(theta) + return [sin, -cos, cos * y - sin * x] + + def line_cross_two_point(self, A, B): + """ + Calculate the line through given point A and B in ax + by + c =0 form. + """ + angle = self.vector_angle(A, B) + return self.theta_line_cross_point(angle, A) + + def average_angle(self, poly): + """ + Calculate the average angle between left and right edge in given poly. + """ + p0, p1, p2, p3 = poly + angle30 = self.vector_angle(p3, p0) + angle21 = self.vector_angle(p2, p1) + return (angle30 + angle21) / 2 + + def line_cross_point(self, line1, line2): + """ + line1 and line2 in 0=ax+by+c form, compute the cross point of line1 and line2 + """ + a1, b1, c1 = line1 + a2, b2, c2 = line2 + d = a1 * b2 - a2 * b1 + + if d == 0: + #print("line1", line1) + #print("line2", line2) + print('Cross point does not exist') + return np.array([0, 0], dtype=np.float32) + else: + x = (b1 * c2 - b2 * c1) / d + y = (a2 * c1 - a1 * c2) / d + + return np.array([x, y], dtype=np.float32) + + def quad2tcl(self, poly, ratio): + """ + Generate center line by poly clock-wise point. (4, 2) + """ + ratio_pair = np.array([[0.5 - ratio / 2], [0.5 + ratio / 2]], dtype=np.float32) + p0_3 = poly[0] + (poly[3] - poly[0]) * ratio_pair + p1_2 = poly[1] + (poly[2] - poly[1]) * ratio_pair + return np.array([p0_3[0], p1_2[0], p1_2[1], p0_3[1]]) + + def poly2tcl(self, poly, ratio): + """ + Generate center line by poly clock-wise point. + """ + ratio_pair = np.array([[0.5 - ratio / 2], [0.5 + ratio / 2]], dtype=np.float32) + tcl_poly = np.zeros_like(poly) + point_num = poly.shape[0] + + for idx in range(point_num // 2): + point_pair = poly[idx] + (poly[point_num - 1 - idx] - poly[idx]) * ratio_pair + tcl_poly[idx] = point_pair[0] + tcl_poly[point_num - 1 - idx] = point_pair[1] + return tcl_poly + + def gen_quad_tbo(self, quad, tcl_mask, tbo_map): + """ + Generate tbo_map for give quad. + """ + # upper and lower line function: ax + by + c = 0; + up_line = self.line_cross_two_point(quad[0], quad[1]) + lower_line = self.line_cross_two_point(quad[3], quad[2]) + + quad_h = 0.5 * (np.linalg.norm(quad[0] - quad[3]) + np.linalg.norm(quad[1] - quad[2])) + quad_w = 0.5 * (np.linalg.norm(quad[0] - quad[1]) + np.linalg.norm(quad[2] - quad[3])) + + # average angle of left and right line. + angle = self.average_angle(quad) + + xy_in_poly = np.argwhere(tcl_mask == 1) + for y, x in xy_in_poly: + point = (x, y) + line = self.theta_line_cross_point(angle, point) + cross_point_upper = self.line_cross_point(up_line, line) + cross_point_lower = self.line_cross_point(lower_line, line) + ##FIX, offset reverse + upper_offset_x, upper_offset_y = cross_point_upper - point + lower_offset_x, lower_offset_y = cross_point_lower - point + tbo_map[y, x, 0] = upper_offset_y + tbo_map[y, x, 1] = upper_offset_x + tbo_map[y, x, 2] = lower_offset_y + tbo_map[y, x, 3] = lower_offset_x + tbo_map[y, x, 4] = 1.0 / max(min(quad_h, quad_w), 1.0) * 2 + return tbo_map + + def poly2quads(self, poly): + """ + Split poly into quads. + """ + quad_list = [] + point_num = poly.shape[0] + + # point pair + point_pair_list = [] + for idx in range(point_num // 2): + point_pair = [poly[idx], poly[point_num - 1 - idx]] + point_pair_list.append(point_pair) + + quad_num = point_num // 2 - 1 + for idx in range(quad_num): + # reshape and adjust to clock-wise + quad_list.append((np.array(point_pair_list)[[idx, idx + 1]]).reshape(4, 2)[[0, 2, 3, 1]]) + + return np.array(quad_list) + + def __call__(self, data): + im = data['image'] + text_polys = data['polys'] + text_tags = data['ignore_tags'] + if im is None: + return None + if text_polys.shape[0] == 0: + return None + + h, w, _ = im.shape + text_polys, text_tags, hv_tags = self.check_and_validate_polys(text_polys, text_tags, (h, w)) + + if text_polys.shape[0] == 0: + return None + + #set aspect ratio and keep area fix + asp_scales = np.arange(1.0, 1.55, 0.1) + asp_scale = np.random.choice(asp_scales) + + if np.random.rand() < 0.5: + asp_scale = 1.0 / asp_scale + asp_scale = math.sqrt(asp_scale) + + asp_wx = asp_scale + asp_hy = 1.0 / asp_scale + im = cv2.resize(im, dsize=None, fx=asp_wx, fy=asp_hy) + text_polys[:, :, 0] *= asp_wx + text_polys[:, :, 1] *= asp_hy + + h, w, _ = im.shape + if max(h, w) > 2048: + rd_scale = 2048.0 / max(h, w) + im = cv2.resize(im, dsize=None, fx=rd_scale, fy=rd_scale) + text_polys *= rd_scale + h, w, _ = im.shape + if min(h, w) < 16: + return None + + #no background + im, text_polys, text_tags, hv_tags = self.crop_area(im, \ + text_polys, text_tags, hv_tags, crop_background=False) + + if text_polys.shape[0] == 0: + return None + #continue for all ignore case + if np.sum((text_tags * 1.0)) >= text_tags.size: + return None + new_h, new_w, _ = im.shape + if (new_h is None) or (new_w is None): + return None + #resize image + std_ratio = float(self.input_size) / max(new_w, new_h) + rand_scales = np.array([0.25, 0.375, 0.5, 0.625, 0.75, 0.875, 1.0, 1.0, 1.0, 1.0, 1.0]) + rz_scale = std_ratio * np.random.choice(rand_scales) + im = cv2.resize(im, dsize=None, fx=rz_scale, fy=rz_scale) + text_polys[:, :, 0] *= rz_scale + text_polys[:, :, 1] *= rz_scale + + #add gaussian blur + if np.random.rand() < 0.1 * 0.5: + ks = np.random.permutation(5)[0] + 1 + ks = int(ks/2)*2 + 1 + im = cv2.GaussianBlur(im, ksize=(ks, ks), sigmaX=0, sigmaY=0) + #add brighter + if np.random.rand() < 0.1 * 0.5: + im = im * (1.0 + np.random.rand() * 0.5) + im = np.clip(im, 0.0, 255.0) + #add darker + if np.random.rand() < 0.1 * 0.5: + im = im * (1.0 - np.random.rand() * 0.5) + im = np.clip(im, 0.0, 255.0) + + # Padding the im to [input_size, input_size] + new_h, new_w, _ = im.shape + if min(new_w, new_h) < self.input_size * 0.5: + return None + + im_padded = np.ones((self.input_size, self.input_size, 3), dtype=np.float32) + im_padded[:, :, 2] = 0.485 * 255 + im_padded[:, :, 1] = 0.456 * 255 + im_padded[:, :, 0] = 0.406 * 255 + + # Random the start position + del_h = self.input_size - new_h + del_w = self.input_size - new_w + sh, sw = 0, 0 + if del_h > 1: + sh = int(np.random.rand() * del_h) + if del_w > 1: + sw = int(np.random.rand() * del_w) + + # Padding + im_padded[sh: sh + new_h, sw: sw + new_w, :] = im.copy() + text_polys[:, :, 0] += sw + text_polys[:, :, 1] += sh + + score_map, border_map, training_mask = self.generate_tcl_label((self.input_size, self.input_size), + text_polys, text_tags, 0.25) + + # SAST head + tvo_map, tco_map = self.generate_tvo_and_tco((self.input_size, self.input_size), text_polys, text_tags, tcl_ratio=0.3, ds_ratio=0.25) + # print("test--------tvo_map shape:", tvo_map.shape) + + im_padded[:, :, 2] -= 0.485 * 255 + im_padded[:, :, 1] -= 0.456 * 255 + im_padded[:, :, 0] -= 0.406 * 255 + im_padded[:, :, 2] /= (255.0 * 0.229) + im_padded[:, :, 1] /= (255.0 * 0.224) + im_padded[:, :, 0] /= (255.0 * 0.225) + im_padded = im_padded.transpose((2, 0, 1)) + + data['image'] = im_padded[::-1, :, :] + data['score_map'] = score_map[np.newaxis, :, :] + data['border_map'] = border_map.transpose((2, 0, 1)) + data['training_mask'] = training_mask[np.newaxis, :, :] + data['tvo_map'] = tvo_map.transpose((2, 0, 1)) + data['tco_map'] = tco_map.transpose((2, 0, 1)) + return data \ No newline at end of file diff --git a/ppocr/losses/__init__.py b/ppocr/losses/__init__.py index 564956e000a1a6eb1567a884f806801059687994..4673d35cece0af221d41132a34c2932246567383 100755 --- a/ppocr/losses/__init__.py +++ b/ppocr/losses/__init__.py @@ -18,6 +18,8 @@ import copy def build_loss(config): # det loss from .det_db_loss import DBLoss + from .det_east_loss import EASTLoss + from .det_sast_loss import SASTLoss # rec loss from .rec_ctc_loss import CTCLoss @@ -25,7 +27,7 @@ def build_loss(config): # cls loss from .cls_loss import ClsLoss - support_dict = ['DBLoss', 'CTCLoss', 'ClsLoss'] + support_dict = ['DBLoss', 'EASTLoss', 'SASTLoss', 'CTCLoss', 'ClsLoss'] config = copy.deepcopy(config) module_name = config.pop('name') diff --git a/ppocr/losses/det_east_loss.py b/ppocr/losses/det_east_loss.py new file mode 100644 index 0000000000000000000000000000000000000000..bcf5372b72d49ebeb34c0ce127e7221d563a6c35 --- /dev/null +++ b/ppocr/losses/det_east_loss.py @@ -0,0 +1,63 @@ +# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import paddle +from paddle import nn +from .det_basic_loss import DiceLoss + + +class EASTLoss(nn.Layer): + """ + """ + + def __init__(self, + eps=1e-6, + **kwargs): + super(EASTLoss, self).__init__() + self.dice_loss = DiceLoss(eps=eps) + + def forward(self, predicts, labels): + l_score, l_geo, l_mask = labels[1:] + f_score = predicts['f_score'] + f_geo = predicts['f_geo'] + + dice_loss = self.dice_loss(f_score, l_score, l_mask) + + #smoooth_l1_loss + channels = 8 + l_geo_split = paddle.split( + l_geo, num_or_sections=channels + 1, axis=1) + f_geo_split = paddle.split(f_geo, num_or_sections=channels, axis=1) + smooth_l1 = 0 + for i in range(0, channels): + geo_diff = l_geo_split[i] - f_geo_split[i] + abs_geo_diff = paddle.abs(geo_diff) + smooth_l1_sign = paddle.less_than(abs_geo_diff, l_score) + smooth_l1_sign = paddle.cast(smooth_l1_sign, dtype='float32') + in_loss = abs_geo_diff * abs_geo_diff * smooth_l1_sign + \ + (abs_geo_diff - 0.5) * (1.0 - smooth_l1_sign) + out_loss = l_geo_split[-1] / channels * in_loss * l_score + smooth_l1 += out_loss + smooth_l1_loss = paddle.mean(smooth_l1 * l_score) + + dice_loss = dice_loss * 0.01 + total_loss = dice_loss + smooth_l1_loss + losses = {"loss":total_loss, \ + "dice_loss":dice_loss,\ + "smooth_l1_loss":smooth_l1_loss} + return losses diff --git a/ppocr/losses/det_sast_loss.py b/ppocr/losses/det_sast_loss.py new file mode 100644 index 0000000000000000000000000000000000000000..a07af6a4598ffb97167094b40a7851ec1089e798 --- /dev/null +++ b/ppocr/losses/det_sast_loss.py @@ -0,0 +1,121 @@ +# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import paddle +from paddle import nn +from .det_basic_loss import DiceLoss +import paddle.fluid as fluid +import numpy as np + + +class SASTLoss(nn.Layer): + """ + """ + + def __init__(self, + eps=1e-6, + **kwargs): + super(SASTLoss, self).__init__() + self.dice_loss = DiceLoss(eps=eps) + + def forward(self, predicts, labels): + """ + tcl_pos: N x 128 x 3 + tcl_mask: N x 128 x 1 + tcl_label: N x X list or LoDTensor + """ + + f_score = predicts['f_score'] + f_border = predicts['f_border'] + f_tvo = predicts['f_tvo'] + f_tco = predicts['f_tco'] + + l_score, l_border, l_mask, l_tvo, l_tco = labels[1:] + + #score_loss + intersection = paddle.sum(f_score * l_score * l_mask) + union = paddle.sum(f_score * l_mask) + paddle.sum(l_score * l_mask) + score_loss = 1.0 - 2 * intersection / (union + 1e-5) + + #border loss + l_border_split, l_border_norm = paddle.split(l_border, num_or_sections=[4, 1], axis=1) + f_border_split = f_border + border_ex_shape = l_border_norm.shape * np.array([1, 4, 1, 1]) + l_border_norm_split = paddle.expand(x=l_border_norm, shape=border_ex_shape) + l_border_score = paddle.expand(x=l_score, shape=border_ex_shape) + l_border_mask = paddle.expand(x=l_mask, shape=border_ex_shape) + + border_diff = l_border_split - f_border_split + abs_border_diff = paddle.abs(border_diff) + border_sign = abs_border_diff < 1.0 + border_sign = paddle.cast(border_sign, dtype='float32') + border_sign.stop_gradient = True + border_in_loss = 0.5 * abs_border_diff * abs_border_diff * border_sign + \ + (abs_border_diff - 0.5) * (1.0 - border_sign) + border_out_loss = l_border_norm_split * border_in_loss + border_loss = paddle.sum(border_out_loss * l_border_score * l_border_mask) / \ + (paddle.sum(l_border_score * l_border_mask) + 1e-5) + + #tvo_loss + l_tvo_split, l_tvo_norm = paddle.split(l_tvo, num_or_sections=[8, 1], axis=1) + f_tvo_split = f_tvo + tvo_ex_shape = l_tvo_norm.shape * np.array([1, 8, 1, 1]) + l_tvo_norm_split = paddle.expand(x=l_tvo_norm, shape=tvo_ex_shape) + l_tvo_score = paddle.expand(x=l_score, shape=tvo_ex_shape) + l_tvo_mask = paddle.expand(x=l_mask, shape=tvo_ex_shape) + # + tvo_geo_diff = l_tvo_split - f_tvo_split + abs_tvo_geo_diff = paddle.abs(tvo_geo_diff) + tvo_sign = abs_tvo_geo_diff < 1.0 + tvo_sign = paddle.cast(tvo_sign, dtype='float32') + tvo_sign.stop_gradient = True + tvo_in_loss = 0.5 * abs_tvo_geo_diff * abs_tvo_geo_diff * tvo_sign + \ + (abs_tvo_geo_diff - 0.5) * (1.0 - tvo_sign) + tvo_out_loss = l_tvo_norm_split * tvo_in_loss + tvo_loss = paddle.sum(tvo_out_loss * l_tvo_score * l_tvo_mask) / \ + (paddle.sum(l_tvo_score * l_tvo_mask) + 1e-5) + + #tco_loss + l_tco_split, l_tco_norm = paddle.split(l_tco, num_or_sections=[2, 1], axis=1) + f_tco_split = f_tco + tco_ex_shape = l_tco_norm.shape * np.array([1, 2, 1, 1]) + l_tco_norm_split = paddle.expand(x=l_tco_norm, shape=tco_ex_shape) + l_tco_score = paddle.expand(x=l_score, shape=tco_ex_shape) + l_tco_mask = paddle.expand(x=l_mask, shape=tco_ex_shape) + + tco_geo_diff = l_tco_split - f_tco_split + abs_tco_geo_diff = paddle.abs(tco_geo_diff) + tco_sign = abs_tco_geo_diff < 1.0 + tco_sign = paddle.cast(tco_sign, dtype='float32') + tco_sign.stop_gradient = True + tco_in_loss = 0.5 * abs_tco_geo_diff * abs_tco_geo_diff * tco_sign + \ + (abs_tco_geo_diff - 0.5) * (1.0 - tco_sign) + tco_out_loss = l_tco_norm_split * tco_in_loss + tco_loss = paddle.sum(tco_out_loss * l_tco_score * l_tco_mask) / \ + (paddle.sum(l_tco_score * l_tco_mask) + 1e-5) + + + # total loss + tvo_lw, tco_lw = 1.5, 1.5 + score_lw, border_lw = 1.0, 1.0 + total_loss = score_loss * score_lw + border_loss * border_lw + \ + tvo_loss * tvo_lw + tco_loss * tco_lw + + losses = {'loss':total_loss, "score_loss":score_loss,\ + "border_loss":border_loss, 'tvo_loss':tvo_loss, 'tco_loss':tco_loss} + return losses \ No newline at end of file diff --git a/ppocr/modeling/backbones/__init__.py b/ppocr/modeling/backbones/__init__.py index 7085d3af317be5664d8d8a10a5fd9bb8deb9267b..43103e53d2413eb63c2831cfd54e91f357b3b496 100755 --- a/ppocr/modeling/backbones/__init__.py +++ b/ppocr/modeling/backbones/__init__.py @@ -19,6 +19,7 @@ def build_backbone(config, model_type): if model_type == 'det': from .det_mobilenet_v3 import MobileNetV3 from .det_resnet_vd import ResNet + from .det_resnet_vd_sast import ResNet_SAST support_dict = ['MobileNetV3', 'ResNet', 'ResNet_SAST'] elif model_type == 'rec' or model_type == 'cls': from .rec_mobilenet_v3 import MobileNetV3 diff --git a/ppocr/modeling/backbones/det_resnet_vd_sast.py b/ppocr/modeling/backbones/det_resnet_vd_sast.py new file mode 100644 index 0000000000000000000000000000000000000000..c9376a8d56e9fa8eff5d2980f63b0b8b955efe88 --- /dev/null +++ b/ppocr/modeling/backbones/det_resnet_vd_sast.py @@ -0,0 +1,285 @@ +# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import paddle +from paddle import ParamAttr +import paddle.nn as nn +import paddle.nn.functional as F + +__all__ = ["ResNet_SAST"] + + +class ConvBNLayer(nn.Layer): + def __init__( + self, + in_channels, + out_channels, + kernel_size, + stride=1, + groups=1, + is_vd_mode=False, + act=None, + name=None, ): + super(ConvBNLayer, self).__init__() + + self.is_vd_mode = is_vd_mode + self._pool2d_avg = nn.AvgPool2D( + kernel_size=2, stride=2, padding=0, ceil_mode=True) + self._conv = nn.Conv2D( + in_channels=in_channels, + out_channels=out_channels, + kernel_size=kernel_size, + stride=stride, + padding=(kernel_size - 1) // 2, + groups=groups, + weight_attr=ParamAttr(name=name + "_weights"), + bias_attr=False) + if name == "conv1": + bn_name = "bn_" + name + else: + bn_name = "bn" + name[3:] + self._batch_norm = nn.BatchNorm( + out_channels, + act=act, + param_attr=ParamAttr(name=bn_name + '_scale'), + bias_attr=ParamAttr(bn_name + '_offset'), + moving_mean_name=bn_name + '_mean', + moving_variance_name=bn_name + '_variance') + + def forward(self, inputs): + if self.is_vd_mode: + inputs = self._pool2d_avg(inputs) + y = self._conv(inputs) + y = self._batch_norm(y) + return y + + +class BottleneckBlock(nn.Layer): + def __init__(self, + in_channels, + out_channels, + stride, + shortcut=True, + if_first=False, + name=None): + super(BottleneckBlock, self).__init__() + + self.conv0 = ConvBNLayer( + in_channels=in_channels, + out_channels=out_channels, + kernel_size=1, + act='relu', + name=name + "_branch2a") + self.conv1 = ConvBNLayer( + in_channels=out_channels, + out_channels=out_channels, + kernel_size=3, + stride=stride, + act='relu', + name=name + "_branch2b") + self.conv2 = ConvBNLayer( + in_channels=out_channels, + out_channels=out_channels * 4, + kernel_size=1, + act=None, + name=name + "_branch2c") + + if not shortcut: + self.short = ConvBNLayer( + in_channels=in_channels, + out_channels=out_channels * 4, + kernel_size=1, + stride=1, + is_vd_mode=False if if_first else True, + name=name + "_branch1") + + self.shortcut = shortcut + + def forward(self, inputs): + y = self.conv0(inputs) + conv1 = self.conv1(y) + conv2 = self.conv2(conv1) + + if self.shortcut: + short = inputs + else: + short = self.short(inputs) + y = paddle.add(x=short, y=conv2) + y = F.relu(y) + return y + + +class BasicBlock(nn.Layer): + def __init__(self, + in_channels, + out_channels, + stride, + shortcut=True, + if_first=False, + name=None): + super(BasicBlock, self).__init__() + self.stride = stride + self.conv0 = ConvBNLayer( + in_channels=in_channels, + out_channels=out_channels, + kernel_size=3, + stride=stride, + act='relu', + name=name + "_branch2a") + self.conv1 = ConvBNLayer( + in_channels=out_channels, + out_channels=out_channels, + kernel_size=3, + act=None, + name=name + "_branch2b") + + if not shortcut: + self.short = ConvBNLayer( + in_channels=in_channels, + out_channels=out_channels, + kernel_size=1, + stride=1, + is_vd_mode=False if if_first else True, + name=name + "_branch1") + + self.shortcut = shortcut + + def forward(self, inputs): + y = self.conv0(inputs) + conv1 = self.conv1(y) + + if self.shortcut: + short = inputs + else: + short = self.short(inputs) + y = paddle.add(x=short, y=conv1) + y = F.relu(y) + return y + + +class ResNet_SAST(nn.Layer): + def __init__(self, in_channels=3, layers=50, **kwargs): + super(ResNet_SAST, self).__init__() + + self.layers = layers + supported_layers = [18, 34, 50, 101, 152, 200] + assert layers in supported_layers, \ + "supported layers are {} but input layer is {}".format( + supported_layers, layers) + + if layers == 18: + depth = [2, 2, 2, 2] + elif layers == 34 or layers == 50: + # depth = [3, 4, 6, 3] + depth = [3, 4, 6, 3, 3] + elif layers == 101: + depth = [3, 4, 23, 3] + elif layers == 152: + depth = [3, 8, 36, 3] + elif layers == 200: + depth = [3, 12, 48, 3] + # num_channels = [64, 256, 512, + # 1024] if layers >= 50 else [64, 64, 128, 256] + # num_filters = [64, 128, 256, 512] + num_channels = [64, 256, 512, + 1024, 2048] if layers >= 50 else [64, 64, 128, 256] + num_filters = [64, 128, 256, 512, 512] + + self.conv1_1 = ConvBNLayer( + in_channels=in_channels, + out_channels=32, + kernel_size=3, + stride=2, + act='relu', + name="conv1_1") + self.conv1_2 = ConvBNLayer( + in_channels=32, + out_channels=32, + kernel_size=3, + stride=1, + act='relu', + name="conv1_2") + self.conv1_3 = ConvBNLayer( + in_channels=32, + out_channels=64, + kernel_size=3, + stride=1, + act='relu', + name="conv1_3") + self.pool2d_max = nn.MaxPool2D(kernel_size=3, stride=2, padding=1) + + self.stages = [] + self.out_channels = [3, 64] + if layers >= 50: + for block in range(len(depth)): + block_list = [] + shortcut = False + for i in range(depth[block]): + if layers in [101, 152] and block == 2: + if i == 0: + conv_name = "res" + str(block + 2) + "a" + else: + conv_name = "res" + str(block + 2) + "b" + str(i) + else: + conv_name = "res" + str(block + 2) + chr(97 + i) + bottleneck_block = self.add_sublayer( + 'bb_%d_%d' % (block, i), + BottleneckBlock( + in_channels=num_channels[block] + if i == 0 else num_filters[block] * 4, + out_channels=num_filters[block], + stride=2 if i == 0 and block != 0 else 1, + shortcut=shortcut, + if_first=block == i == 0, + name=conv_name)) + shortcut = True + block_list.append(bottleneck_block) + self.out_channels.append(num_filters[block] * 4) + self.stages.append(nn.Sequential(*block_list)) + else: + for block in range(len(depth)): + block_list = [] + shortcut = False + for i in range(depth[block]): + conv_name = "res" + str(block + 2) + chr(97 + i) + basic_block = self.add_sublayer( + 'bb_%d_%d' % (block, i), + BasicBlock( + in_channels=num_channels[block] + if i == 0 else num_filters[block], + out_channels=num_filters[block], + stride=2 if i == 0 and block != 0 else 1, + shortcut=shortcut, + if_first=block == i == 0, + name=conv_name)) + shortcut = True + block_list.append(basic_block) + self.out_channels.append(num_filters[block]) + self.stages.append(nn.Sequential(*block_list)) + + def forward(self, inputs): + out = [inputs] + y = self.conv1_1(inputs) + y = self.conv1_2(y) + y = self.conv1_3(y) + out.append(y) + y = self.pool2d_max(y) + for block in self.stages: + y = block(y) + out.append(y) + return out \ No newline at end of file diff --git a/ppocr/modeling/heads/__init__.py b/ppocr/modeling/heads/__init__.py index 673f6fb4b6ffa72abc9fe1571869169f51604b5c..78074709056f4865098600a28e63b54e68eecf80 100755 --- a/ppocr/modeling/heads/__init__.py +++ b/ppocr/modeling/heads/__init__.py @@ -18,13 +18,15 @@ __all__ = ['build_head'] def build_head(config): # det head from .det_db_head import DBHead + from .det_east_head import EASTHead + from .det_sast_head import SASTHead # rec head from .rec_ctc_head import CTCHead # cls head from .cls_head import ClsHead - support_dict = ['DBHead', 'CTCHead', 'ClsHead'] + support_dict = ['DBHead', 'EASTHead', 'SASTHead', 'CTCHead', 'ClsHead'] module_name = config.pop('name') assert module_name in support_dict, Exception('head only support {}'.format( diff --git a/ppocr/modeling/heads/det_east_head.py b/ppocr/modeling/heads/det_east_head.py new file mode 100644 index 0000000000000000000000000000000000000000..9d0c3c4cf83adb018fcc368374cbe305658e07a9 --- /dev/null +++ b/ppocr/modeling/heads/det_east_head.py @@ -0,0 +1,121 @@ +# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import math +import paddle +from paddle import nn +import paddle.nn.functional as F +from paddle import ParamAttr + + +class ConvBNLayer(nn.Layer): + def __init__(self, + in_channels, + out_channels, + kernel_size, + stride, + padding, + groups=1, + if_act=True, + act=None, + name=None): + super(ConvBNLayer, self).__init__() + self.if_act = if_act + self.act = act + self.conv = nn.Conv2D( + in_channels=in_channels, + out_channels=out_channels, + kernel_size=kernel_size, + stride=stride, + padding=padding, + groups=groups, + weight_attr=ParamAttr(name=name + '_weights'), + bias_attr=False) + + self.bn = nn.BatchNorm( + num_channels=out_channels, + act=act, + param_attr=ParamAttr(name="bn_" + name + "_scale"), + bias_attr=ParamAttr(name="bn_" + name + "_offset"), + moving_mean_name="bn_" + name + "_mean", + moving_variance_name="bn_" + name + "_variance") + + def forward(self, x): + x = self.conv(x) + x = self.bn(x) + return x + + +class EASTHead(nn.Layer): + """ + """ + def __init__(self, in_channels, model_name, **kwargs): + super(EASTHead, self).__init__() + self.model_name = model_name + if self.model_name == "large": + num_outputs = [128, 64, 1, 8] + else: + num_outputs = [64, 32, 1, 8] + + self.det_conv1 = ConvBNLayer( + in_channels=in_channels, + out_channels=num_outputs[0], + kernel_size=3, + stride=1, + padding=1, + if_act=True, + act='relu', + name="det_head1") + self.det_conv2 = ConvBNLayer( + in_channels=num_outputs[0], + out_channels=num_outputs[1], + kernel_size=3, + stride=1, + padding=1, + if_act=True, + act='relu', + name="det_head2") + self.score_conv = ConvBNLayer( + in_channels=num_outputs[1], + out_channels=num_outputs[2], + kernel_size=1, + stride=1, + padding=0, + if_act=False, + act=None, + name="f_score") + self.geo_conv = ConvBNLayer( + in_channels=num_outputs[1], + out_channels=num_outputs[3], + kernel_size=1, + stride=1, + padding=0, + if_act=False, + act=None, + name="f_geo") + + def forward(self, x): + f_det = self.det_conv1(x) + f_det = self.det_conv2(f_det) + f_score = self.score_conv(f_det) + f_score = F.sigmoid(f_score) + f_geo = self.geo_conv(f_det) + f_geo = (F.sigmoid(f_geo) - 0.5) * 2 * 800 + + pred = {'f_score': f_score, 'f_geo': f_geo} + return pred diff --git a/ppocr/modeling/heads/det_sast_head.py b/ppocr/modeling/heads/det_sast_head.py new file mode 100644 index 0000000000000000000000000000000000000000..263b28672299e733369938fa03952dca7685fabe --- /dev/null +++ b/ppocr/modeling/heads/det_sast_head.py @@ -0,0 +1,128 @@ +# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import math +import paddle +from paddle import nn +import paddle.nn.functional as F +from paddle import ParamAttr + + +class ConvBNLayer(nn.Layer): + def __init__(self, + in_channels, + out_channels, + kernel_size, + stride, + groups=1, + if_act=True, + act=None, + name=None): + super(ConvBNLayer, self).__init__() + self.if_act = if_act + self.act = act + self.conv = nn.Conv2D( + in_channels=in_channels, + out_channels=out_channels, + kernel_size=kernel_size, + stride=stride, + padding=(kernel_size - 1) // 2, + groups=groups, + weight_attr=ParamAttr(name=name + '_weights'), + bias_attr=False) + + self.bn = nn.BatchNorm( + num_channels=out_channels, + act=act, + param_attr=ParamAttr(name="bn_" + name + "_scale"), + bias_attr=ParamAttr(name="bn_" + name + "_offset"), + moving_mean_name="bn_" + name + "_mean", + moving_variance_name="bn_" + name + "_variance") + + def forward(self, x): + x = self.conv(x) + x = self.bn(x) + return x + + +class SAST_Header1(nn.Layer): + def __init__(self, in_channels, **kwargs): + super(SAST_Header1, self).__init__() + out_channels = [64, 64, 128] + self.score_conv = nn.Sequential( + ConvBNLayer(in_channels, out_channels[0], 1, 1, act='relu', name='f_score1'), + ConvBNLayer(out_channels[0], out_channels[1], 3, 1, act='relu', name='f_score2'), + ConvBNLayer(out_channels[1], out_channels[2], 1, 1, act='relu', name='f_score3'), + ConvBNLayer(out_channels[2], 1, 3, 1, act=None, name='f_score4') + ) + self.border_conv = nn.Sequential( + ConvBNLayer(in_channels, out_channels[0], 1, 1, act='relu', name='f_border1'), + ConvBNLayer(out_channels[0], out_channels[1], 3, 1, act='relu', name='f_border2'), + ConvBNLayer(out_channels[1], out_channels[2], 1, 1, act='relu', name='f_border3'), + ConvBNLayer(out_channels[2], 4, 3, 1, act=None, name='f_border4') + ) + + def forward(self, x): + f_score = self.score_conv(x) + f_score = F.sigmoid(f_score) + f_border = self.border_conv(x) + return f_score, f_border + + +class SAST_Header2(nn.Layer): + def __init__(self, in_channels, **kwargs): + super(SAST_Header2, self).__init__() + out_channels = [64, 64, 128] + self.tvo_conv = nn.Sequential( + ConvBNLayer(in_channels, out_channels[0], 1, 1, act='relu', name='f_tvo1'), + ConvBNLayer(out_channels[0], out_channels[1], 3, 1, act='relu', name='f_tvo2'), + ConvBNLayer(out_channels[1], out_channels[2], 1, 1, act='relu', name='f_tvo3'), + ConvBNLayer(out_channels[2], 8, 3, 1, act=None, name='f_tvo4') + ) + self.tco_conv = nn.Sequential( + ConvBNLayer(in_channels, out_channels[0], 1, 1, act='relu', name='f_tco1'), + ConvBNLayer(out_channels[0], out_channels[1], 3, 1, act='relu', name='f_tco2'), + ConvBNLayer(out_channels[1], out_channels[2], 1, 1, act='relu', name='f_tco3'), + ConvBNLayer(out_channels[2], 2, 3, 1, act=None, name='f_tco4') + ) + + def forward(self, x): + f_tvo = self.tvo_conv(x) + f_tco = self.tco_conv(x) + return f_tvo, f_tco + + +class SASTHead(nn.Layer): + """ + """ + def __init__(self, in_channels, **kwargs): + super(SASTHead, self).__init__() + + self.head1 = SAST_Header1(in_channels) + self.head2 = SAST_Header2(in_channels) + + def forward(self, x): + f_score, f_border = self.head1(x) + f_tvo, f_tco = self.head2(x) + + predicts = {} + predicts['f_score'] = f_score + predicts['f_border'] = f_border + predicts['f_tvo'] = f_tvo + predicts['f_tco'] = f_tco + return predicts \ No newline at end of file diff --git a/ppocr/modeling/necks/__init__.py b/ppocr/modeling/necks/__init__.py index a9bf414b71489985f3d63d5aabea0805dc312c27..405e062b352da759b743d5997fd6e4c8b89c038b 100644 --- a/ppocr/modeling/necks/__init__.py +++ b/ppocr/modeling/necks/__init__.py @@ -16,8 +16,10 @@ __all__ = ['build_neck'] def build_neck(config): from .db_fpn import DBFPN + from .east_fpn import EASTFPN + from .sast_fpn import SASTFPN from .rnn import SequenceEncoder - support_dict = ['DBFPN', 'SequenceEncoder'] + support_dict = ['DBFPN', 'EASTFPN', 'SASTFPN', 'SequenceEncoder'] module_name = config.pop('name') assert module_name in support_dict, Exception('neck only support {}'.format( diff --git a/ppocr/modeling/necks/east_fpn.py b/ppocr/modeling/necks/east_fpn.py new file mode 100644 index 0000000000000000000000000000000000000000..120ff156cba2f7f8f8a562e76c82f1fac8584d4a --- /dev/null +++ b/ppocr/modeling/necks/east_fpn.py @@ -0,0 +1,188 @@ +# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import paddle +from paddle import nn +import paddle.nn.functional as F +from paddle import ParamAttr + + +class ConvBNLayer(nn.Layer): + def __init__(self, + in_channels, + out_channels, + kernel_size, + stride, + padding, + groups=1, + if_act=True, + act=None, + name=None): + super(ConvBNLayer, self).__init__() + self.if_act = if_act + self.act = act + self.conv = nn.Conv2D( + in_channels=in_channels, + out_channels=out_channels, + kernel_size=kernel_size, + stride=stride, + padding=padding, + groups=groups, + weight_attr=ParamAttr(name=name + '_weights'), + bias_attr=False) + + self.bn = nn.BatchNorm( + num_channels=out_channels, + act=act, + param_attr=ParamAttr(name="bn_" + name + "_scale"), + bias_attr=ParamAttr(name="bn_" + name + "_offset"), + moving_mean_name="bn_" + name + "_mean", + moving_variance_name="bn_" + name + "_variance") + + def forward(self, x): + x = self.conv(x) + x = self.bn(x) + return x + + +class DeConvBNLayer(nn.Layer): + def __init__(self, + in_channels, + out_channels, + kernel_size, + stride, + padding, + groups=1, + if_act=True, + act=None, + name=None): + super(DeConvBNLayer, self).__init__() + self.if_act = if_act + self.act = act + self.deconv = nn.Conv2DTranspose( + in_channels=in_channels, + out_channels=out_channels, + kernel_size=kernel_size, + stride=stride, + padding=padding, + groups=groups, + weight_attr=ParamAttr(name=name + '_weights'), + bias_attr=False) + self.bn = nn.BatchNorm( + num_channels=out_channels, + act=act, + param_attr=ParamAttr(name="bn_" + name + "_scale"), + bias_attr=ParamAttr(name="bn_" + name + "_offset"), + moving_mean_name="bn_" + name + "_mean", + moving_variance_name="bn_" + name + "_variance") + + def forward(self, x): + x = self.deconv(x) + x = self.bn(x) + return x + + +class EASTFPN(nn.Layer): + def __init__(self, in_channels, model_name, **kwargs): + super(EASTFPN, self).__init__() + self.model_name = model_name + if self.model_name == "large": + self.out_channels = 128 + else: + self.out_channels = 64 + self.in_channels = in_channels[::-1] + self.h1_conv = ConvBNLayer( + in_channels=self.out_channels+self.in_channels[1], + out_channels=self.out_channels, + kernel_size=3, + stride=1, + padding=1, + if_act=True, + act='relu', + name="unet_h_1") + self.h2_conv = ConvBNLayer( + in_channels=self.out_channels+self.in_channels[2], + out_channels=self.out_channels, + kernel_size=3, + stride=1, + padding=1, + if_act=True, + act='relu', + name="unet_h_2") + self.h3_conv = ConvBNLayer( + in_channels=self.out_channels+self.in_channels[3], + out_channels=self.out_channels, + kernel_size=3, + stride=1, + padding=1, + if_act=True, + act='relu', + name="unet_h_3") + self.g0_deconv = DeConvBNLayer( + in_channels=self.in_channels[0], + out_channels=self.out_channels, + kernel_size=4, + stride=2, + padding=1, + if_act=True, + act='relu', + name="unet_g_0") + self.g1_deconv = DeConvBNLayer( + in_channels=self.out_channels, + out_channels=self.out_channels, + kernel_size=4, + stride=2, + padding=1, + if_act=True, + act='relu', + name="unet_g_1") + self.g2_deconv = DeConvBNLayer( + in_channels=self.out_channels, + out_channels=self.out_channels, + kernel_size=4, + stride=2, + padding=1, + if_act=True, + act='relu', + name="unet_g_2") + self.g3_conv = ConvBNLayer( + in_channels=self.out_channels, + out_channels=self.out_channels, + kernel_size=3, + stride=1, + padding=1, + if_act=True, + act='relu', + name="unet_g_3") + + def forward(self, x): + f = x[::-1] + + h = f[0] + g = self.g0_deconv(h) + h = paddle.concat([g, f[1]], axis=1) + h = self.h1_conv(h) + g = self.g1_deconv(h) + h = paddle.concat([g, f[2]], axis=1) + h = self.h2_conv(h) + g = self.g2_deconv(h) + h = paddle.concat([g, f[3]], axis=1) + h = self.h3_conv(h) + g = self.g3_conv(h) + + return g \ No newline at end of file diff --git a/ppocr/modeling/necks/sast_fpn.py b/ppocr/modeling/necks/sast_fpn.py new file mode 100644 index 0000000000000000000000000000000000000000..9b6024590ac574f725fc6c519dd96ee4462276d6 --- /dev/null +++ b/ppocr/modeling/necks/sast_fpn.py @@ -0,0 +1,284 @@ +# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import paddle +from paddle import nn +import paddle.nn.functional as F +from paddle import ParamAttr + + +class ConvBNLayer(nn.Layer): + def __init__(self, + in_channels, + out_channels, + kernel_size, + stride, + groups=1, + if_act=True, + act=None, + name=None): + super(ConvBNLayer, self).__init__() + self.if_act = if_act + self.act = act + self.conv = nn.Conv2D( + in_channels=in_channels, + out_channels=out_channels, + kernel_size=kernel_size, + stride=stride, + padding=(kernel_size - 1) // 2, + groups=groups, + weight_attr=ParamAttr(name=name + '_weights'), + bias_attr=False) + + self.bn = nn.BatchNorm( + num_channels=out_channels, + act=act, + param_attr=ParamAttr(name="bn_" + name + "_scale"), + bias_attr=ParamAttr(name="bn_" + name + "_offset"), + moving_mean_name="bn_" + name + "_mean", + moving_variance_name="bn_" + name + "_variance") + + def forward(self, x): + x = self.conv(x) + x = self.bn(x) + return x + + +class DeConvBNLayer(nn.Layer): + def __init__(self, + in_channels, + out_channels, + kernel_size, + stride, + groups=1, + if_act=True, + act=None, + name=None): + super(DeConvBNLayer, self).__init__() + self.if_act = if_act + self.act = act + self.deconv = nn.Conv2DTranspose( + in_channels=in_channels, + out_channels=out_channels, + kernel_size=kernel_size, + stride=stride, + padding=(kernel_size - 1) // 2, + groups=groups, + weight_attr=ParamAttr(name=name + '_weights'), + bias_attr=False) + self.bn = nn.BatchNorm( + num_channels=out_channels, + act=act, + param_attr=ParamAttr(name="bn_" + name + "_scale"), + bias_attr=ParamAttr(name="bn_" + name + "_offset"), + moving_mean_name="bn_" + name + "_mean", + moving_variance_name="bn_" + name + "_variance") + + def forward(self, x): + x = self.deconv(x) + x = self.bn(x) + return x + + +class FPN_Up_Fusion(nn.Layer): + def __init__(self, in_channels): + super(FPN_Up_Fusion, self).__init__() + in_channels = in_channels[::-1] + out_channels = [256, 256, 192, 192, 128] + + self.h0_conv = ConvBNLayer(in_channels[0], out_channels[0], 1, 1, act=None, name='fpn_up_h0') + self.h1_conv = ConvBNLayer(in_channels[1], out_channels[1], 1, 1, act=None, name='fpn_up_h1') + self.h2_conv = ConvBNLayer(in_channels[2], out_channels[2], 1, 1, act=None, name='fpn_up_h2') + self.h3_conv = ConvBNLayer(in_channels[3], out_channels[3], 1, 1, act=None, name='fpn_up_h3') + self.h4_conv = ConvBNLayer(in_channels[4], out_channels[4], 1, 1, act=None, name='fpn_up_h4') + + self.g0_conv = DeConvBNLayer(out_channels[0], out_channels[1], 4, 2, act=None, name='fpn_up_g0') + + self.g1_conv = nn.Sequential( + ConvBNLayer(out_channels[1], out_channels[1], 3, 1, act='relu', name='fpn_up_g1_1'), + DeConvBNLayer(out_channels[1], out_channels[2], 4, 2, act=None, name='fpn_up_g1_2') + ) + self.g2_conv = nn.Sequential( + ConvBNLayer(out_channels[2], out_channels[2], 3, 1, act='relu', name='fpn_up_g2_1'), + DeConvBNLayer(out_channels[2], out_channels[3], 4, 2, act=None, name='fpn_up_g2_2') + ) + self.g3_conv = nn.Sequential( + ConvBNLayer(out_channels[3], out_channels[3], 3, 1, act='relu', name='fpn_up_g3_1'), + DeConvBNLayer(out_channels[3], out_channels[4], 4, 2, act=None, name='fpn_up_g3_2') + ) + + self.g4_conv = nn.Sequential( + ConvBNLayer(out_channels[4], out_channels[4], 3, 1, act='relu', name='fpn_up_fusion_1'), + ConvBNLayer(out_channels[4], out_channels[4], 1, 1, act=None, name='fpn_up_fusion_2') + ) + + def _add_relu(self, x1, x2): + x = paddle.add(x=x1, y=x2) + x = F.relu(x) + return x + + def forward(self, x): + f = x[2:][::-1] + h0 = self.h0_conv(f[0]) + h1 = self.h1_conv(f[1]) + h2 = self.h2_conv(f[2]) + h3 = self.h3_conv(f[3]) + h4 = self.h4_conv(f[4]) + + g0 = self.g0_conv(h0) + g1 = self._add_relu(g0, h1) + g1 = self.g1_conv(g1) + g2 = self.g2_conv(self._add_relu(g1, h2)) + g3 = self.g3_conv(self._add_relu(g2, h3)) + g4 = self.g4_conv(self._add_relu(g3, h4)) + + return g4 + + +class FPN_Down_Fusion(nn.Layer): + def __init__(self, in_channels): + super(FPN_Down_Fusion, self).__init__() + out_channels = [32, 64, 128] + + self.h0_conv = ConvBNLayer(in_channels[0], out_channels[0], 3, 1, act=None, name='fpn_down_h0') + self.h1_conv = ConvBNLayer(in_channels[1], out_channels[1], 3, 1, act=None, name='fpn_down_h1') + self.h2_conv = ConvBNLayer(in_channels[2], out_channels[2], 3, 1, act=None, name='fpn_down_h2') + + self.g0_conv = ConvBNLayer(out_channels[0], out_channels[1], 3, 2, act=None, name='fpn_down_g0') + + self.g1_conv = nn.Sequential( + ConvBNLayer(out_channels[1], out_channels[1], 3, 1, act='relu', name='fpn_down_g1_1'), + ConvBNLayer(out_channels[1], out_channels[2], 3, 2, act=None, name='fpn_down_g1_2') + ) + + self.g2_conv = nn.Sequential( + ConvBNLayer(out_channels[2], out_channels[2], 3, 1, act='relu', name='fpn_down_fusion_1'), + ConvBNLayer(out_channels[2], out_channels[2], 1, 1, act=None, name='fpn_down_fusion_2') + ) + + def forward(self, x): + f = x[:3] + h0 = self.h0_conv(f[0]) + h1 = self.h1_conv(f[1]) + h2 = self.h2_conv(f[2]) + g0 = self.g0_conv(h0) + g1 = paddle.add(x=g0, y=h1) + g1 = F.relu(g1) + g1 = self.g1_conv(g1) + g2 = paddle.add(x=g1, y=h2) + g2 = F.relu(g2) + g2 = self.g2_conv(g2) + return g2 + + +class Cross_Attention(nn.Layer): + def __init__(self, in_channels): + super(Cross_Attention, self).__init__() + self.theta_conv = ConvBNLayer(in_channels, in_channels, 1, 1, act='relu', name='f_theta') + self.phi_conv = ConvBNLayer(in_channels, in_channels, 1, 1, act='relu', name='f_phi') + self.g_conv = ConvBNLayer(in_channels, in_channels, 1, 1, act='relu', name='f_g') + + self.fh_weight_conv = ConvBNLayer(in_channels, in_channels, 1, 1, act=None, name='fh_weight') + self.fh_sc_conv = ConvBNLayer(in_channels, in_channels, 1, 1, act=None, name='fh_sc') + + self.fv_weight_conv = ConvBNLayer(in_channels, in_channels, 1, 1, act=None, name='fv_weight') + self.fv_sc_conv = ConvBNLayer(in_channels, in_channels, 1, 1, act=None, name='fv_sc') + + self.f_attn_conv = ConvBNLayer(in_channels * 2, in_channels, 1, 1, act='relu', name='f_attn') + + def _cal_fweight(self, f, shape): + f_theta, f_phi, f_g = f + #flatten + f_theta = paddle.transpose(f_theta, [0, 2, 3, 1]) + f_theta = paddle.reshape(f_theta, [shape[0] * shape[1], shape[2], 128]) + f_phi = paddle.transpose(f_phi, [0, 2, 3, 1]) + f_phi = paddle.reshape(f_phi, [shape[0] * shape[1], shape[2], 128]) + f_g = paddle.transpose(f_g, [0, 2, 3, 1]) + f_g = paddle.reshape(f_g, [shape[0] * shape[1], shape[2], 128]) + #correlation + f_attn = paddle.matmul(f_theta, paddle.transpose(f_phi, [0, 2, 1])) + #scale + f_attn = f_attn / (128**0.5) + f_attn = F.softmax(f_attn) + #weighted sum + f_weight = paddle.matmul(f_attn, f_g) + f_weight = paddle.reshape( + f_weight, [shape[0], shape[1], shape[2], 128]) + return f_weight + + def forward(self, f_common): + f_shape = paddle.shape(f_common) + # print('f_shape: ', f_shape) + + f_theta = self.theta_conv(f_common) + f_phi = self.phi_conv(f_common) + f_g = self.g_conv(f_common) + + ######## horizon ######## + fh_weight = self._cal_fweight([f_theta, f_phi, f_g], + [f_shape[0], f_shape[2], f_shape[3]]) + fh_weight = paddle.transpose(fh_weight, [0, 3, 1, 2]) + fh_weight = self.fh_weight_conv(fh_weight) + #short cut + fh_sc = self.fh_sc_conv(f_common) + f_h = F.relu(fh_weight + fh_sc) + + ######## vertical ######## + fv_theta = paddle.transpose(f_theta, [0, 1, 3, 2]) + fv_phi = paddle.transpose(f_phi, [0, 1, 3, 2]) + fv_g = paddle.transpose(f_g, [0, 1, 3, 2]) + fv_weight = self._cal_fweight([fv_theta, fv_phi, fv_g], + [f_shape[0], f_shape[3], f_shape[2]]) + fv_weight = paddle.transpose(fv_weight, [0, 3, 2, 1]) + fv_weight = self.fv_weight_conv(fv_weight) + #short cut + fv_sc = self.fv_sc_conv(f_common) + f_v = F.relu(fv_weight + fv_sc) + + ######## merge ######## + f_attn = paddle.concat([f_h, f_v], axis=1) + f_attn = self.f_attn_conv(f_attn) + return f_attn + + +class SASTFPN(nn.Layer): + def __init__(self, in_channels, with_cab=False, **kwargs): + super(SASTFPN, self).__init__() + self.in_channels = in_channels + self.with_cab = with_cab + self.FPN_Down_Fusion = FPN_Down_Fusion(self.in_channels) + self.FPN_Up_Fusion = FPN_Up_Fusion(self.in_channels) + self.out_channels = 128 + self.cross_attention = Cross_Attention(self.out_channels) + + def forward(self, x): + #down fpn + f_down = self.FPN_Down_Fusion(x) + + #up fpn + f_up = self.FPN_Up_Fusion(x) + + #fusion + f_common = paddle.add(x=f_down, y=f_up) + f_common = F.relu(f_common) + + if self.with_cab: + # print('enhence f_common with CAB.') + f_common = self.cross_attention(f_common) + + return f_common diff --git a/ppocr/postprocess/__init__.py b/ppocr/postprocess/__init__.py index e08a217da867020e31183cc45893ed5cdd2aee1e..c9b42e08391d0405a58a071460eb6f57bbb62c43 100644 --- a/ppocr/postprocess/__init__.py +++ b/ppocr/postprocess/__init__.py @@ -24,11 +24,13 @@ __all__ = ['build_post_process'] def build_post_process(config, global_config=None): from .db_postprocess import DBPostProcess + from .east_postprocess import EASTPostProcess + from .sast_postprocess import SASTPostProcess from .rec_postprocess import CTCLabelDecode, AttnLabelDecode from .cls_postprocess import ClsPostProcess support_dict = [ - 'DBPostProcess', 'CTCLabelDecode', 'AttnLabelDecode', 'ClsPostProcess' + 'DBPostProcess', 'EASTPostProcess', 'SASTPostProcess', 'CTCLabelDecode', 'AttnLabelDecode', 'ClsPostProcess' ] config = copy.deepcopy(config) diff --git a/ppocr/postprocess/east_postprocess.py b/ppocr/postprocess/east_postprocess.py new file mode 100644 index 0000000000000000000000000000000000000000..0b669405562aef9812b9771977bf82f362beb75e --- /dev/null +++ b/ppocr/postprocess/east_postprocess.py @@ -0,0 +1,141 @@ +# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import numpy as np +from .locality_aware_nms import nms_locality +import cv2 + +import os +import sys +# __dir__ = os.path.dirname(os.path.abspath(__file__)) +# sys.path.append(__dir__) +# sys.path.append(os.path.abspath(os.path.join(__dir__, '..'))) + + +class EASTPostProcess(object): + """ + The post process for EAST. + """ + def __init__(self, + score_thresh=0.8, + cover_thresh=0.1, + nms_thresh=0.2, + **kwargs): + + self.score_thresh = score_thresh + self.cover_thresh = cover_thresh + self.nms_thresh = nms_thresh + + # c++ la-nms is faster, but only support python 3.5 + self.is_python35 = False + if sys.version_info.major == 3 and sys.version_info.minor == 5: + self.is_python35 = True + + def restore_rectangle_quad(self, origin, geometry): + """ + Restore rectangle from quadrangle. + """ + # quad + origin_concat = np.concatenate( + (origin, origin, origin, origin), axis=1) # (n, 8) + pred_quads = origin_concat - geometry + pred_quads = pred_quads.reshape((-1, 4, 2)) # (n, 4, 2) + return pred_quads + + def detect(self, + score_map, + geo_map, + score_thresh=0.8, + cover_thresh=0.1, + nms_thresh=0.2): + """ + restore text boxes from score map and geo map + """ + score_map = score_map[0] + geo_map = np.swapaxes(geo_map, 1, 0) + geo_map = np.swapaxes(geo_map, 1, 2) + # filter the score map + xy_text = np.argwhere(score_map > score_thresh) + if len(xy_text) == 0: + return [] + # sort the text boxes via the y axis + xy_text = xy_text[np.argsort(xy_text[:, 0])] + #restore quad proposals + text_box_restored = self.restore_rectangle_quad( + xy_text[:, ::-1] * 4, geo_map[xy_text[:, 0], xy_text[:, 1], :]) + boxes = np.zeros((text_box_restored.shape[0], 9), dtype=np.float32) + boxes[:, :8] = text_box_restored.reshape((-1, 8)) + boxes[:, 8] = score_map[xy_text[:, 0], xy_text[:, 1]] + if self.is_python35: + import lanms + boxes = lanms.merge_quadrangle_n9(boxes, nms_thresh) + else: + boxes = nms_locality(boxes.astype(np.float64), nms_thresh) + if boxes.shape[0] == 0: + return [] + # Here we filter some low score boxes by the average score map, + # this is different from the orginal paper. + for i, box in enumerate(boxes): + mask = np.zeros_like(score_map, dtype=np.uint8) + cv2.fillPoly(mask, box[:8].reshape( + (-1, 4, 2)).astype(np.int32) // 4, 1) + boxes[i, 8] = cv2.mean(score_map, mask)[0] + boxes = boxes[boxes[:, 8] > cover_thresh] + return boxes + + def sort_poly(self, p): + """ + Sort polygons. + """ + min_axis = np.argmin(np.sum(p, axis=1)) + p = p[[min_axis, (min_axis + 1) % 4,\ + (min_axis + 2) % 4, (min_axis + 3) % 4]] + if abs(p[0, 0] - p[1, 0]) > abs(p[0, 1] - p[1, 1]): + return p + else: + return p[[0, 3, 2, 1]] + + def __call__(self, outs_dict, shape_list): + score_list = outs_dict['f_score'] + geo_list = outs_dict['f_geo'] + img_num = len(shape_list) + dt_boxes_list = [] + for ino in range(img_num): + score = score_list[ino].numpy() + geo = geo_list[ino].numpy() + boxes = self.detect( + score_map=score, + geo_map=geo, + score_thresh=self.score_thresh, + cover_thresh=self.cover_thresh, + nms_thresh=self.nms_thresh) + boxes_norm = [] + if len(boxes) > 0: + h, w = score.shape[1:] + src_h, src_w, ratio_h, ratio_w = shape_list[ino] + boxes = boxes[:, :8].reshape((-1, 4, 2)) + boxes[:, :, 0] /= ratio_w + boxes[:, :, 1] /= ratio_h + for i_box, box in enumerate(boxes): + box = self.sort_poly(box.astype(np.int32)) + if np.linalg.norm(box[0] - box[1]) < 5 \ + or np.linalg.norm(box[3] - box[0]) < 5: + continue + boxes_norm.append(box) + dt_boxes_list.append({'points': np.array(boxes_norm)}) + return dt_boxes_list \ No newline at end of file diff --git a/ppocr/postprocess/locality_aware_nms.py b/ppocr/postprocess/locality_aware_nms.py new file mode 100644 index 0000000000000000000000000000000000000000..53280cc13ed7e41859e23e2517938d4f6eb07076 --- /dev/null +++ b/ppocr/postprocess/locality_aware_nms.py @@ -0,0 +1,199 @@ +""" +Locality aware nms. +""" + +import numpy as np +from shapely.geometry import Polygon + + +def intersection(g, p): + """ + Intersection. + """ + g = Polygon(g[:8].reshape((4, 2))) + p = Polygon(p[:8].reshape((4, 2))) + g = g.buffer(0) + p = p.buffer(0) + if not g.is_valid or not p.is_valid: + return 0 + inter = Polygon(g).intersection(Polygon(p)).area + union = g.area + p.area - inter + if union == 0: + return 0 + else: + return inter / union + + +def intersection_iog(g, p): + """ + Intersection_iog. + """ + g = Polygon(g[:8].reshape((4, 2))) + p = Polygon(p[:8].reshape((4, 2))) + if not g.is_valid or not p.is_valid: + return 0 + inter = Polygon(g).intersection(Polygon(p)).area + #union = g.area + p.area - inter + union = p.area + if union == 0: + print("p_area is very small") + return 0 + else: + return inter / union + + +def weighted_merge(g, p): + """ + Weighted merge. + """ + g[:8] = (g[8] * g[:8] + p[8] * p[:8]) / (g[8] + p[8]) + g[8] = (g[8] + p[8]) + return g + + +def standard_nms(S, thres): + """ + Standard nms. + """ + order = np.argsort(S[:, 8])[::-1] + keep = [] + while order.size > 0: + i = order[0] + keep.append(i) + ovr = np.array([intersection(S[i], S[t]) for t in order[1:]]) + + inds = np.where(ovr <= thres)[0] + order = order[inds + 1] + + return S[keep] + + +def standard_nms_inds(S, thres): + """ + Standard nms, retun inds. + """ + order = np.argsort(S[:, 8])[::-1] + keep = [] + while order.size > 0: + i = order[0] + keep.append(i) + ovr = np.array([intersection(S[i], S[t]) for t in order[1:]]) + + inds = np.where(ovr <= thres)[0] + order = order[inds + 1] + + return keep + + +def nms(S, thres): + """ + nms. + """ + order = np.argsort(S[:, 8])[::-1] + keep = [] + while order.size > 0: + i = order[0] + keep.append(i) + ovr = np.array([intersection(S[i], S[t]) for t in order[1:]]) + + inds = np.where(ovr <= thres)[0] + order = order[inds + 1] + + return keep + + +def soft_nms(boxes_in, Nt_thres=0.3, threshold=0.8, sigma=0.5, method=2): + """ + soft_nms + :para boxes_in, N x 9 (coords + score) + :para threshould, eliminate cases min score(0.001) + :para Nt_thres, iou_threshi + :para sigma, gaussian weght + :method, linear or gaussian + """ + boxes = boxes_in.copy() + N = boxes.shape[0] + if N is None or N < 1: + return np.array([]) + pos, maxpos = 0, 0 + weight = 0.0 + inds = np.arange(N) + tbox, sbox = boxes[0].copy(), boxes[0].copy() + for i in range(N): + maxscore = boxes[i, 8] + maxpos = i + tbox = boxes[i].copy() + ti = inds[i] + pos = i + 1 + #get max box + while pos < N: + if maxscore < boxes[pos, 8]: + maxscore = boxes[pos, 8] + maxpos = pos + pos = pos + 1 + #add max box as a detection + boxes[i, :] = boxes[maxpos, :] + inds[i] = inds[maxpos] + #swap + boxes[maxpos, :] = tbox + inds[maxpos] = ti + tbox = boxes[i].copy() + pos = i + 1 + #NMS iteration + while pos < N: + sbox = boxes[pos].copy() + ts_iou_val = intersection(tbox, sbox) + if ts_iou_val > 0: + if method == 1: + if ts_iou_val > Nt_thres: + weight = 1 - ts_iou_val + else: + weight = 1 + elif method == 2: + weight = np.exp(-1.0 * ts_iou_val**2 / sigma) + else: + if ts_iou_val > Nt_thres: + weight = 0 + else: + weight = 1 + boxes[pos, 8] = weight * boxes[pos, 8] + #if box score falls below thresold, discard the box by + #swaping last box update N + if boxes[pos, 8] < threshold: + boxes[pos, :] = boxes[N - 1, :] + inds[pos] = inds[N - 1] + N = N - 1 + pos = pos - 1 + pos = pos + 1 + + return boxes[:N] + + +def nms_locality(polys, thres=0.3): + """ + locality aware nms of EAST + :param polys: a N*9 numpy array. first 8 coordinates, then prob + :return: boxes after nms + """ + S = [] + p = None + for g in polys: + if p is not None and intersection(g, p) > thres: + p = weighted_merge(g, p) + else: + if p is not None: + S.append(p) + p = g + if p is not None: + S.append(p) + + if len(S) == 0: + return np.array([]) + return standard_nms(np.array(S), thres) + + +if __name__ == '__main__': + # 343,350,448,135,474,143,369,359 + print( + Polygon(np.array([[343, 350], [448, 135], [474, 143], [369, 359]])) + .area) \ No newline at end of file diff --git a/ppocr/postprocess/rec_postprocess.py b/ppocr/postprocess/rec_postprocess.py index 6943f84f87913b5e4093c4ced2beebc8de98da73..a18e101bf428930246d04b1553113184cdf2753d 100644 --- a/ppocr/postprocess/rec_postprocess.py +++ b/ppocr/postprocess/rec_postprocess.py @@ -27,7 +27,7 @@ class BaseRecLabelDecode(object): 'ch', 'en', 'en_sensitive', 'french', 'german', 'japan', 'korean' ] assert character_type in support_character_type, "Only {} are supported now but get {}".format( - support_character_type, self.character_str) + support_character_type, character_type) if character_type == "en": self.character_str = "0123456789abcdefghijklmnopqrstuvwxyz" diff --git a/ppocr/postprocess/sast_postprocess.py b/ppocr/postprocess/sast_postprocess.py new file mode 100644 index 0000000000000000000000000000000000000000..03b0e8f17d60a8863a2d1d5900e01dcd4874d5a9 --- /dev/null +++ b/ppocr/postprocess/sast_postprocess.py @@ -0,0 +1,295 @@ +# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import os +import sys +__dir__ = os.path.dirname(__file__) +sys.path.append(__dir__) +sys.path.append(os.path.join(__dir__, '..')) + +import numpy as np +from .locality_aware_nms import nms_locality +# import lanms +import cv2 +import time + + +class SASTPostProcess(object): + """ + The post process for SAST. + """ + + def __init__(self, + score_thresh=0.5, + nms_thresh=0.2, + sample_pts_num=2, + shrink_ratio_of_width=0.3, + expand_scale=1.0, + tcl_map_thresh=0.5, + **kwargs): + + self.score_thresh = score_thresh + self.nms_thresh = nms_thresh + self.sample_pts_num = sample_pts_num + self.shrink_ratio_of_width = shrink_ratio_of_width + self.expand_scale = expand_scale + self.tcl_map_thresh = tcl_map_thresh + + # c++ la-nms is faster, but only support python 3.5 + self.is_python35 = False + if sys.version_info.major == 3 and sys.version_info.minor == 5: + self.is_python35 = True + + def point_pair2poly(self, point_pair_list): + """ + Transfer vertical point_pairs into poly point in clockwise. + """ + # constract poly + point_num = len(point_pair_list) * 2 + point_list = [0] * point_num + for idx, point_pair in enumerate(point_pair_list): + point_list[idx] = point_pair[0] + point_list[point_num - 1 - idx] = point_pair[1] + return np.array(point_list).reshape(-1, 2) + + def shrink_quad_along_width(self, quad, begin_width_ratio=0., end_width_ratio=1.): + """ + Generate shrink_quad_along_width. + """ + ratio_pair = np.array([[begin_width_ratio], [end_width_ratio]], dtype=np.float32) + p0_1 = quad[0] + (quad[1] - quad[0]) * ratio_pair + p3_2 = quad[3] + (quad[2] - quad[3]) * ratio_pair + return np.array([p0_1[0], p0_1[1], p3_2[1], p3_2[0]]) + + def expand_poly_along_width(self, poly, shrink_ratio_of_width=0.3): + """ + expand poly along width. + """ + point_num = poly.shape[0] + left_quad = np.array([poly[0], poly[1], poly[-2], poly[-1]], dtype=np.float32) + left_ratio = -shrink_ratio_of_width * np.linalg.norm(left_quad[0] - left_quad[3]) / \ + (np.linalg.norm(left_quad[0] - left_quad[1]) + 1e-6) + left_quad_expand = self.shrink_quad_along_width(left_quad, left_ratio, 1.0) + right_quad = np.array([poly[point_num // 2 - 2], poly[point_num // 2 - 1], + poly[point_num // 2], poly[point_num // 2 + 1]], dtype=np.float32) + right_ratio = 1.0 + \ + shrink_ratio_of_width * np.linalg.norm(right_quad[0] - right_quad[3]) / \ + (np.linalg.norm(right_quad[0] - right_quad[1]) + 1e-6) + right_quad_expand = self.shrink_quad_along_width(right_quad, 0.0, right_ratio) + poly[0] = left_quad_expand[0] + poly[-1] = left_quad_expand[-1] + poly[point_num // 2 - 1] = right_quad_expand[1] + poly[point_num // 2] = right_quad_expand[2] + return poly + + def restore_quad(self, tcl_map, tcl_map_thresh, tvo_map): + """Restore quad.""" + xy_text = np.argwhere(tcl_map[:, :, 0] > tcl_map_thresh) + xy_text = xy_text[:, ::-1] # (n, 2) + + # Sort the text boxes via the y axis + xy_text = xy_text[np.argsort(xy_text[:, 1])] + + scores = tcl_map[xy_text[:, 1], xy_text[:, 0], 0] + scores = scores[:, np.newaxis] + + # Restore + point_num = int(tvo_map.shape[-1] / 2) + assert point_num == 4 + tvo_map = tvo_map[xy_text[:, 1], xy_text[:, 0], :] + xy_text_tile = np.tile(xy_text, (1, point_num)) # (n, point_num * 2) + quads = xy_text_tile - tvo_map + + return scores, quads, xy_text + + def quad_area(self, quad): + """ + compute area of a quad. + """ + edge = [ + (quad[1][0] - quad[0][0]) * (quad[1][1] + quad[0][1]), + (quad[2][0] - quad[1][0]) * (quad[2][1] + quad[1][1]), + (quad[3][0] - quad[2][0]) * (quad[3][1] + quad[2][1]), + (quad[0][0] - quad[3][0]) * (quad[0][1] + quad[3][1]) + ] + return np.sum(edge) / 2. + + def nms(self, dets): + if self.is_python35: + import lanms + dets = lanms.merge_quadrangle_n9(dets, self.nms_thresh) + else: + dets = nms_locality(dets, self.nms_thresh) + return dets + + def cluster_by_quads_tco(self, tcl_map, tcl_map_thresh, quads, tco_map): + """ + Cluster pixels in tcl_map based on quads. + """ + instance_count = quads.shape[0] + 1 # contain background + instance_label_map = np.zeros(tcl_map.shape[:2], dtype=np.int32) + if instance_count == 1: + return instance_count, instance_label_map + + # predict text center + xy_text = np.argwhere(tcl_map[:, :, 0] > tcl_map_thresh) + n = xy_text.shape[0] + xy_text = xy_text[:, ::-1] # (n, 2) + tco = tco_map[xy_text[:, 1], xy_text[:, 0], :] # (n, 2) + pred_tc = xy_text - tco + + # get gt text center + m = quads.shape[0] + gt_tc = np.mean(quads, axis=1) # (m, 2) + + pred_tc_tile = np.tile(pred_tc[:, np.newaxis, :], (1, m, 1)) # (n, m, 2) + gt_tc_tile = np.tile(gt_tc[np.newaxis, :, :], (n, 1, 1)) # (n, m, 2) + dist_mat = np.linalg.norm(pred_tc_tile - gt_tc_tile, axis=2) # (n, m) + xy_text_assign = np.argmin(dist_mat, axis=1) + 1 # (n,) + + instance_label_map[xy_text[:, 1], xy_text[:, 0]] = xy_text_assign + return instance_count, instance_label_map + + def estimate_sample_pts_num(self, quad, xy_text): + """ + Estimate sample points number. + """ + eh = (np.linalg.norm(quad[0] - quad[3]) + np.linalg.norm(quad[1] - quad[2])) / 2.0 + ew = (np.linalg.norm(quad[0] - quad[1]) + np.linalg.norm(quad[2] - quad[3])) / 2.0 + + dense_sample_pts_num = max(2, int(ew)) + dense_xy_center_line = xy_text[np.linspace(0, xy_text.shape[0] - 1, dense_sample_pts_num, + endpoint=True, dtype=np.float32).astype(np.int32)] + + dense_xy_center_line_diff = dense_xy_center_line[1:] - dense_xy_center_line[:-1] + estimate_arc_len = np.sum(np.linalg.norm(dense_xy_center_line_diff, axis=1)) + + sample_pts_num = max(2, int(estimate_arc_len / eh)) + return sample_pts_num + + def detect_sast(self, tcl_map, tvo_map, tbo_map, tco_map, ratio_w, ratio_h, src_w, src_h, + shrink_ratio_of_width=0.3, tcl_map_thresh=0.5, offset_expand=1.0, out_strid=4.0): + """ + first resize the tcl_map, tvo_map and tbo_map to the input_size, then restore the polys + """ + # restore quad + scores, quads, xy_text = self.restore_quad(tcl_map, tcl_map_thresh, tvo_map) + dets = np.hstack((quads, scores)).astype(np.float32, copy=False) + dets = self.nms(dets) + if dets.shape[0] == 0: + return [] + quads = dets[:, :-1].reshape(-1, 4, 2) + + # Compute quad area + quad_areas = [] + for quad in quads: + quad_areas.append(-self.quad_area(quad)) + + # instance segmentation + # instance_count, instance_label_map = cv2.connectedComponents(tcl_map.astype(np.uint8), connectivity=8) + instance_count, instance_label_map = self.cluster_by_quads_tco(tcl_map, tcl_map_thresh, quads, tco_map) + + # restore single poly with tcl instance. + poly_list = [] + for instance_idx in range(1, instance_count): + xy_text = np.argwhere(instance_label_map == instance_idx)[:, ::-1] + quad = quads[instance_idx - 1] + q_area = quad_areas[instance_idx - 1] + if q_area < 5: + continue + + # + len1 = float(np.linalg.norm(quad[0] -quad[1])) + len2 = float(np.linalg.norm(quad[1] -quad[2])) + min_len = min(len1, len2) + if min_len < 3: + continue + + # filter small CC + if xy_text.shape[0] <= 0: + continue + + # filter low confidence instance + xy_text_scores = tcl_map[xy_text[:, 1], xy_text[:, 0], 0] + if np.sum(xy_text_scores) / quad_areas[instance_idx - 1] < 0.1: + # if np.sum(xy_text_scores) / quad_areas[instance_idx - 1] < 0.05: + continue + + # sort xy_text + left_center_pt = np.array([[(quad[0, 0] + quad[-1, 0]) / 2.0, + (quad[0, 1] + quad[-1, 1]) / 2.0]]) # (1, 2) + right_center_pt = np.array([[(quad[1, 0] + quad[2, 0]) / 2.0, + (quad[1, 1] + quad[2, 1]) / 2.0]]) # (1, 2) + proj_unit_vec = (right_center_pt - left_center_pt) / \ + (np.linalg.norm(right_center_pt - left_center_pt) + 1e-6) + proj_value = np.sum(xy_text * proj_unit_vec, axis=1) + xy_text = xy_text[np.argsort(proj_value)] + + # Sample pts in tcl map + if self.sample_pts_num == 0: + sample_pts_num = self.estimate_sample_pts_num(quad, xy_text) + else: + sample_pts_num = self.sample_pts_num + xy_center_line = xy_text[np.linspace(0, xy_text.shape[0] - 1, sample_pts_num, + endpoint=True, dtype=np.float32).astype(np.int32)] + + point_pair_list = [] + for x, y in xy_center_line: + # get corresponding offset + offset = tbo_map[y, x, :].reshape(2, 2) + if offset_expand != 1.0: + offset_length = np.linalg.norm(offset, axis=1, keepdims=True) + expand_length = np.clip(offset_length * (offset_expand - 1), a_min=0.5, a_max=3.0) + offset_detal = offset / offset_length * expand_length + offset = offset + offset_detal + # original point + ori_yx = np.array([y, x], dtype=np.float32) + point_pair = (ori_yx + offset)[:, ::-1]* out_strid / np.array([ratio_w, ratio_h]).reshape(-1, 2) + point_pair_list.append(point_pair) + + # ndarry: (x, 2), expand poly along width + detected_poly = self.point_pair2poly(point_pair_list) + detected_poly = self.expand_poly_along_width(detected_poly, shrink_ratio_of_width) + detected_poly[:, 0] = np.clip(detected_poly[:, 0], a_min=0, a_max=src_w) + detected_poly[:, 1] = np.clip(detected_poly[:, 1], a_min=0, a_max=src_h) + poly_list.append(detected_poly) + + return poly_list + + def __call__(self, outs_dict, shape_list): + score_list = outs_dict['f_score'] + border_list = outs_dict['f_border'] + tvo_list = outs_dict['f_tvo'] + tco_list = outs_dict['f_tco'] + + img_num = len(shape_list) + poly_lists = [] + for ino in range(img_num): + p_score = score_list[ino].transpose((1,2,0)).numpy() + p_border = border_list[ino].transpose((1,2,0)).numpy() + p_tvo = tvo_list[ino].transpose((1,2,0)).numpy() + p_tco = tco_list[ino].transpose((1,2,0)).numpy() + src_h, src_w, ratio_h, ratio_w = shape_list[ino] + + poly_list = self.detect_sast(p_score, p_tvo, p_border, p_tco, ratio_w, ratio_h, src_w, src_h, + shrink_ratio_of_width=self.shrink_ratio_of_width, + tcl_map_thresh=self.tcl_map_thresh, offset_expand=self.expand_scale) + poly_lists.append({'points': np.array(poly_list)}) + + return poly_lists + diff --git a/ppocr/utils/character.py b/ppocr/utils/character.py deleted file mode 100755 index b4b2021e02c9905623fd9fad5c9673543569c1c2..0000000000000000000000000000000000000000 --- a/ppocr/utils/character.py +++ /dev/null @@ -1,214 +0,0 @@ -# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import numpy as np -import string -import re -from .check import check_config_params -import sys - - -class CharacterOps(object): - """ Convert between text-label and text-index """ - - def __init__(self, config): - self.character_type = config['character_type'] - self.loss_type = config['loss_type'] - self.max_text_len = config['max_text_length'] - if self.character_type == "en": - self.character_str = "0123456789abcdefghijklmnopqrstuvwxyz" - dict_character = list(self.character_str) - elif self.character_type == "ch": - character_dict_path = config['character_dict_path'] - add_space = False - if 'use_space_char' in config: - add_space = config['use_space_char'] - self.character_str = "" - with open(character_dict_path, "rb") as fin: - lines = fin.readlines() - for line in lines: - line = line.decode('utf-8').strip("\n").strip("\r\n") - self.character_str += line - if add_space: - self.character_str += " " - dict_character = list(self.character_str) - elif self.character_type == "en_sensitive": - # same with ASTER setting (use 94 char). - self.character_str = string.printable[:-6] - dict_character = list(self.character_str) - else: - self.character_str = None - assert self.character_str is not None, \ - "Nonsupport type of the character: {}".format(self.character_str) - self.beg_str = "sos" - self.end_str = "eos" - if self.loss_type == "attention": - dict_character = [self.beg_str, self.end_str] + dict_character - elif self.loss_type == "srn": - dict_character = dict_character + [self.beg_str, self.end_str] - self.dict = {} - for i, char in enumerate(dict_character): - self.dict[char] = i - self.character = dict_character - - def encode(self, text): - """convert text-label into text-index. - input: - text: text labels of each image. [batch_size] - - output: - text: concatenated text index for CTCLoss. - [sum(text_lengths)] = [text_index_0 + text_index_1 + ... + text_index_(n - 1)] - length: length of each text. [batch_size] - """ - if self.character_type == "en": - text = text.lower() - - text_list = [] - for char in text: - if char not in self.dict: - continue - text_list.append(self.dict[char]) - text = np.array(text_list) - return text - - def decode(self, text_index, is_remove_duplicate=False): - """ convert text-index into text-label. """ - char_list = [] - char_num = self.get_char_num() - - if self.loss_type == "attention": - beg_idx = self.get_beg_end_flag_idx("beg") - end_idx = self.get_beg_end_flag_idx("end") - ignored_tokens = [beg_idx, end_idx] - else: - ignored_tokens = [char_num] - - for idx in range(len(text_index)): - if text_index[idx] in ignored_tokens: - continue - if is_remove_duplicate: - if idx > 0 and text_index[idx - 1] == text_index[idx]: - continue - char_list.append(self.character[int(text_index[idx])]) - text = ''.join(char_list) - return text - - def get_char_num(self): - return len(self.character) - - def get_beg_end_flag_idx(self, beg_or_end): - if self.loss_type == "attention": - if beg_or_end == "beg": - idx = np.array(self.dict[self.beg_str]) - elif beg_or_end == "end": - idx = np.array(self.dict[self.end_str]) - else: - assert False, "Unsupport type %s in get_beg_end_flag_idx"\ - % beg_or_end - return idx - else: - err = "error in get_beg_end_flag_idx when using the loss %s"\ - % (self.loss_type) - assert False, err - - -def cal_predicts_accuracy(char_ops, - preds, - preds_lod, - labels, - labels_lod, - is_remove_duplicate=False): - acc_num = 0 - img_num = 0 - for ino in range(len(labels_lod) - 1): - beg_no = preds_lod[ino] - end_no = preds_lod[ino + 1] - preds_text = preds[beg_no:end_no].reshape(-1) - preds_text = char_ops.decode(preds_text, is_remove_duplicate) - - beg_no = labels_lod[ino] - end_no = labels_lod[ino + 1] - labels_text = labels[beg_no:end_no].reshape(-1) - labels_text = char_ops.decode(labels_text, is_remove_duplicate) - img_num += 1 - - if preds_text == labels_text: - acc_num += 1 - acc = acc_num * 1.0 / img_num - return acc, acc_num, img_num - - -def cal_predicts_accuracy_srn(char_ops, - preds, - labels, - max_text_len, - is_debug=False): - acc_num = 0 - img_num = 0 - - char_num = char_ops.get_char_num() - - total_len = preds.shape[0] - img_num = int(total_len / max_text_len) - for i in range(img_num): - cur_label = [] - cur_pred = [] - for j in range(max_text_len): - if labels[j + i * max_text_len] != int(char_num-1): #0 - cur_label.append(labels[j + i * max_text_len][0]) - else: - break - - for j in range(max_text_len + 1): - if j < len(cur_label) and preds[j + i * max_text_len][ - 0] != cur_label[j]: - break - elif j == len(cur_label) and j == max_text_len: - acc_num += 1 - break - elif j == len(cur_label) and preds[j + i * max_text_len][0] == int(char_num-1): - acc_num += 1 - break - acc = acc_num * 1.0 / img_num - return acc, acc_num, img_num - - -def convert_rec_attention_infer_res(preds): - img_num = preds.shape[0] - target_lod = [0] - convert_ids = [] - for ino in range(img_num): - end_pos = np.where(preds[ino, :] == 1)[0] - if len(end_pos) <= 1: - text_list = preds[ino, 1:] - else: - text_list = preds[ino, 1:end_pos[1]] - target_lod.append(target_lod[ino] + len(text_list)) - convert_ids = convert_ids + list(text_list) - convert_ids = np.array(convert_ids) - convert_ids = convert_ids.reshape((-1, 1)) - return convert_ids, target_lod - - -def convert_rec_label_to_lod(ori_labels): - img_num = len(ori_labels) - target_lod = [0] - convert_ids = [] - for ino in range(img_num): - target_lod.append(target_lod[ino] + len(ori_labels[ino])) - convert_ids = convert_ids + list(ori_labels[ino]) - convert_ids = np.array(convert_ids) - convert_ids = convert_ids.reshape((-1, 1)) - return convert_ids, target_lod diff --git a/ppocr/utils/check.py b/ppocr/utils/check.py deleted file mode 100755 index 3a0b14061e2058464c4942e6cf4f891a2a77ba69..0000000000000000000000000000000000000000 --- a/ppocr/utils/check.py +++ /dev/null @@ -1,31 +0,0 @@ -# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from __future__ import absolute_import -from __future__ import division -from __future__ import print_function -from __future__ import unicode_literals - -import sys - -import logging -logger = logging.getLogger(__name__) - - -def check_config_params(config, config_name, params): - for param in params: - if param not in config: - err = "param %s didn't find in %s!" % (param, config_name) - assert False, err - return diff --git a/tools/infer/utility.py b/tools/infer/utility.py index ee1f954dcc4b6518cfe454a86650b397b9db449e..fabc33dc67265cd304294bab15ecd6d242a3add6 100755 --- a/tools/infer/utility.py +++ b/tools/infer/utility.py @@ -230,10 +230,10 @@ def draw_ocr_box_txt(image, box[2][1], box[3][0], box[3][1] ], outline=color) - box_height = math.sqrt((box[0][0] - box[3][0]) ** 2 + (box[0][1] - box[3][ - 1]) ** 2) - box_width = math.sqrt((box[0][0] - box[1][0]) ** 2 + (box[0][1] - box[1][ - 1]) ** 2) + box_height = math.sqrt((box[0][0] - box[3][0])**2 + (box[0][1] - box[3][ + 1])**2) + box_width = math.sqrt((box[0][0] - box[1][0])**2 + (box[0][1] - box[1][ + 1])**2) if box_height > 2 * box_width: font_size = max(int(box_width * 0.9), 10) font = ImageFont.truetype(font_path, font_size, encoding="utf-8") @@ -260,7 +260,6 @@ def str_count(s): Count the number of Chinese characters, a single English character and a single number equal to half the length of Chinese characters. - args: s(string): the input of string return(int): @@ -295,7 +294,6 @@ def text_visual(texts, img_w(int): the width of blank img font_path: the path of font which is used to draw text return(array): - """ if scores is not None: assert len(texts) == len(