fix bug in Faster RCNN eval & refine code (#1737)

* fix bug in eval faster & refine code

fluid/PaddleCV/rcnn/config.py

@@ -219,12 +219,6 @@ _C.pixel_means = [102.9801, 115.9465, 122.7717]
 # clip box to prevent overflowing
 _C.bbox_clip = np.log(1000. / 16.)
 
-# dataset path
-_C.train_file_list = 'annotations/instances_train2017.json'
-_C.train_data_dir = 'train2017'
-_C.val_file_list = 'annotations/instances_val2017.json'
-_C.val_data_dir = 'val2017'
-
 
 def merge_cfg_from_args(args, mode):
     """Merge config keys, values in args into the global config."""

fluid/PaddleCV/rcnn/eval_coco_map.py

@@ -29,20 +29,20 @@ import json
 from pycocotools.coco import COCO
 from pycocotools.cocoeval import COCOeval, Params
 from config import cfg
+from roidbs import DatasetPath
 
 
 def eval():
-    if '2014' in cfg.dataset:
-        test_list = 'annotations/instances_val2014.json'
-    elif '2017' in cfg.dataset:
-        test_list = 'annotations/instances_val2017.json'
+
+    data_path = DatasetPath('val')
+    test_list = data_path.get_file_list()
 
     image_shape = [3, cfg.TEST.max_size, cfg.TEST.max_size]
     class_nums = cfg.class_num
     devices = os.getenv("CUDA_VISIBLE_DEVICES") or ""
     devices_num = len(devices.split(","))
     total_batch_size = devices_num * cfg.TRAIN.im_per_batch
-    cocoGt = COCO(os.path.join(cfg.data_dir, test_list))
+    cocoGt = COCO(test_list)
     num_id_to_cat_id_map = {i + 1: v for i, v in enumerate(cocoGt.getCatIds())}
     category_ids = cocoGt.getCatIds()
     label_list = {
@@ -51,14 +51,13 @@ def eval():
     }
     label_list[0] = ['background']
 
-    model = model_builder.FasterRCNN(
+    model = model_builder.RCNN(
         add_conv_body_func=resnet.add_ResNet50_conv4_body,
         add_roi_box_head_func=resnet.add_ResNet_roi_conv5_head,
         use_pyreader=False,
         is_train=False)
     model.build_model(image_shape)
-    rpn_rois, confs, locs = model.eval_bbox_out()
-    pred_boxes = model.eval()
+    pred_boxes = model.eval_bbox_out()
     if cfg.MASK_ON:
         masks = model.eval_mask_out()
     place = fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace()
@@ -77,9 +76,10 @@ def eval():
     dts_res = []
     segms_res = []
     if cfg.MASK_ON:
-        fetch_list = [rpn_rois, confs, locs, pred_boxes, masks]
+        fetch_list = [pred_boxes, masks]
     else:
-        fetch_list = [rpn_rois, confs, locs]
+        fetch_list = [pred_boxes]
+    eval_start = time.time()
     for batch_id, batch_data in enumerate(test_reader()):
         start = time.time()
         im_info = []
@@ -89,12 +89,9 @@ def eval():
                          feed=feeder.feed(batch_data),
                          return_numpy=False)
 
-        rpn_rois_v = result[0]
-        confs_v = result[1]
-        locs_v = result[2]
+        pred_boxes_v = result[0]
         if cfg.MASK_ON:
-            pred_boxes_v = result[3]
-            masks_v = result[4]
+            masks_v = result[1]
 
         new_lod = pred_boxes_v.lod()
         nmsed_out = pred_boxes_v
@@ -108,6 +105,9 @@ def eval():
                                        batch_data, num_id_to_cat_id_map)
         end = time.time()
         print('batch id: {}, time: {}'.format(batch_id, end - start))
+    eval_end = time.time()
+    total_time = eval_end - eval_start
+    print('average time of eval is: {}'.format(total_time / (batch_id + 1)))
     with open("detection_bbox_result.json", 'w') as outfile:
         json.dump(dts_res, outfile)
     print("start evaluate bbox using coco api")

fluid/PaddleCV/rcnn/infer.py

@@ -12,16 +12,15 @@ import json
 from pycocotools.coco import COCO
 from pycocotools.cocoeval import COCOeval, Params
 from config import cfg
+from roidbs import DatasetPath
 
 
 def infer():
 
-    if '2014' in cfg.dataset:
-        test_list = 'annotations/instances_val2014.json'
-    elif '2017' in cfg.dataset:
-        test_list = 'annotations/instances_val2017.json'
+    data_path = DatasetPath('val')
+    test_list = data_path.get_file_list()
 
-    cocoGt = COCO(os.path.join(cfg.data_dir, test_list))
+    cocoGt = COCO(test_list)
     num_id_to_cat_id_map = {i + 1: v for i, v in enumerate(cocoGt.getCatIds())}
     category_ids = cocoGt.getCatIds()
     label_list = {
@@ -32,14 +31,13 @@ def infer():
     image_shape = [3, cfg.TEST.max_size, cfg.TEST.max_size]
     class_nums = cfg.class_num
 
-    model = model_builder.FasterRCNN(
+    model = model_builder.RCNN(
         add_conv_body_func=resnet.add_ResNet50_conv4_body,
         add_roi_box_head_func=resnet.add_ResNet_roi_conv5_head,
         use_pyreader=False,
         is_train=False)
     model.build_model(image_shape)
-    rpn_rois, confs, locs = model.eval_bbox_out()
-    pred_boxes = model.eval()
+    pred_boxes = model.eval_bbox_out()
     if cfg.MASK_ON:
         masks = model.eval_mask_out()
     place = fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace()
@@ -56,20 +54,17 @@ def infer():
     dts_res = []
     segms_res = []
     if cfg.MASK_ON:
-        fetch_list = [rpn_rois, confs, locs, pred_boxes, masks]
+        fetch_list = [pred_boxes, masks]
     else:
-        fetch_list = [rpn_rois, confs, locs]
+        fetch_list = [pred_boxes]
     data = next(infer_reader())
     im_info = [data[0][1]]
     result = exe.run(fetch_list=[v.name for v in fetch_list],
                      feed=feeder.feed(data),
                      return_numpy=False)
-    rpn_rois_v = result[0]
-    confs_v = result[1]
-    locs_v = result[2]
+    pred_boxes_v = result[0]
     if cfg.MASK_ON:
-        pred_boxes_v = result[3]
-        masks_v = result[4]
+        masks_v = result[1]
     new_lod = pred_boxes_v.lod()
     nmsed_out = pred_boxes_v
     path = os.path.join(cfg.image_path, cfg.image_name)

fluid/PaddleCV/rcnn/models/model_builder.py

@@ -23,7 +23,7 @@ import cPickle as cp
 import numpy as np
 
 
-class FasterRCNN(object):
+class RCNN(object):
     def __init__(self,
                  add_conv_body_func=None,
                  add_roi_box_head_func=None,
@@ -64,14 +64,10 @@ class FasterRCNN(object):
         rloss = [loss] + losses
         return rloss, rkeys
 
-    def eval_bbox_out(self):
-        cls_prob = fluid.layers.softmax(self.cls_score, use_cudnn=False)
-        return [self.rpn_rois, cls_prob, self.bbox_pred]
-
     def eval_mask_out(self):
         return self.mask_fcn_logits
 
-    def eval(self):
+    def eval_bbox_out(self):
         return self.pred_result
 
     def build_input(self, image_shape):
@@ -131,6 +127,33 @@ class FasterRCNN(object):
             self.im_id, self.gt_masks
         ]
 
+    def eval_bbox(self):
+        self.im_scale = fluid.layers.slice(
+            self.im_info, [1], starts=[2], ends=[3])
+        im_scale_lod = fluid.layers.sequence_expand(self.im_scale,
+                                                    self.rpn_rois)
+        boxes = self.rpn_rois / im_scale_lod
+        cls_prob = fluid.layers.softmax(self.cls_score, use_cudnn=False)
+        bbox_pred_reshape = fluid.layers.reshape(self.bbox_pred,
+                                                 (-1, cfg.class_num, 4))
+        decoded_box = fluid.layers.box_coder(
+            prior_box=boxes,
+            prior_box_var=cfg.bbox_reg_weights,
+            target_box=bbox_pred_reshape,
+            code_type='decode_center_size',
+            box_normalized=False,
+            axis=1)
+        cliped_box = fluid.layers.box_clip(
+            input=decoded_box, im_info=self.im_info)
+        self.pred_result = fluid.layers.multiclass_nms(
+            bboxes=cliped_box,
+            scores=cls_prob,
+            score_threshold=cfg.TEST.score_thresh,
+            nms_top_k=-1,
+            nms_threshold=cfg.TEST.nms_thresh,
+            keep_top_k=cfg.TEST.detections_per_im,
+            normalized=False)
+
     def rpn_heads(self, rpn_input):
         # RPN hidden representation
         dim_out = rpn_input.shape[1]
@@ -275,6 +298,8 @@ class FasterRCNN(object):
                                              name='bbox_pred_b',
                                              learning_rate=2.,
                                              regularizer=L2Decay(0.)))
+        if not self.is_train:
+            self.eval_bbox()
 
     def SuffixNet(self, conv5):
         mask_out = fluid.layers.conv2d_transpose(
@@ -313,30 +338,7 @@ class FasterRCNN(object):
                                         self.roi_has_mask_int32)
             self.mask_fcn_logits = self.SuffixNet(conv5)
         else:
-            im_scale = fluid.layers.slice(
-                self.im_info, [1], starts=[2], ends=[3])
-            im_scale_lod = fluid.layers.sequence_expand(im_scale, self.rpn_rois)
-            boxes = self.rpn_rois / im_scale_lod
-            cls_prob = fluid.layers.softmax(self.cls_score, use_cudnn=False)
-            bbox_pred_reshape = fluid.layers.reshape(self.bbox_pred,
-                                                     (-1, cfg.class_num, 4))
-            decoded_box = fluid.layers.box_coder(
-                prior_box=boxes,
-                prior_box_var=cfg.bbox_reg_weights,
-                target_box=bbox_pred_reshape,
-                code_type='decode_center_size',
-                box_normalized=False,
-                axis=1)
-            cliped_box = fluid.layers.box_clip(
-                input=decoded_box, im_info=self.im_info)
-            self.pred_result = fluid.layers.multiclass_nms(
-                bboxes=cliped_box,
-                scores=cls_prob,
-                score_threshold=cfg.TEST.score_thresh,
-                nms_top_k=-1,
-                nms_threshold=cfg.TEST.nms_thresh,
-                keep_top_k=cfg.TEST.detections_per_im,
-                normalized=False)
+            self.eval_bbox()
             pred_res_shape = fluid.layers.shape(self.pred_result)
             shape = fluid.layers.reduce_prod(pred_res_shape)
             shape = fluid.layers.reshape(shape, [1, 1])
@@ -351,7 +353,7 @@ class FasterRCNN(object):
                 pred_res = ie.input(self.pred_result)
                 pred_boxes = fluid.layers.slice(
                     pred_res, [1], starts=[2], ends=[6])
-                im_scale_lod = fluid.layers.sequence_expand(im_scale,
+                im_scale_lod = fluid.layers.sequence_expand(self.im_scale,
                                                             pred_boxes)
                 mask_rois = pred_boxes * im_scale_lod
                 conv5 = self.add_roi_box_head_func(mask_input, mask_rois)

fluid/PaddleCV/rcnn/profile.py

@@ -37,7 +37,7 @@ def train():
     devices = os.getenv("CUDA_VISIBLE_DEVICES") or ""
     devices_num = len(devices.split(","))
     total_batch_size = devices_num * cfg.TRAIN.im_per_batch
-    model = model_builder.FasterRCNN(
+    model = model_builder.RCNN(
         add_conv_body_func=resnet.add_ResNet50_conv4_body,
         add_roi_box_head_func=resnet.add_ResNet_roi_conv5_head,
         use_pyreader=cfg.use_pyreader,

fluid/PaddleCV/rcnn/reader.py

@@ -36,7 +36,7 @@ def roidb_reader(roidb, mode):
     im_height = np.round(roidb['height'] * im_scales)
     im_width = np.round(roidb['width'] * im_scales)
     im_info = np.array([im_height, im_width, im_scales], dtype=np.float32)
-    if mode == 'test' or mode == 'infer':
+    if mode == 'val' or mode == 'infer':
         return im, im_info, im_id
 
     gt_boxes = roidb['gt_boxes'].astype('float32')
@@ -74,31 +74,12 @@ def coco(mode,
          total_batch_size=None,
          padding_total=False,
          shuffle=False):
-    if 'coco2014' in cfg.dataset:
-        cfg.train_file_list = 'annotations/instances_train2014.json'
-        cfg.train_data_dir = 'train2014'
-        cfg.val_file_list = 'annotations/instances_val2014.json'
-        cfg.val_data_dir = 'val2014'
-    elif 'coco2017' in cfg.dataset:
-        cfg.train_file_list = 'annotations/instances_train2017.json'
-        cfg.train_data_dir = 'train2017'
-        cfg.val_file_list = 'annotations/instances_val2017.json'
-        cfg.val_data_dir = 'val2017'
-    else:
-        raise NotImplementedError('Dataset {} not supported'.format(
-            cfg.dataset))
     cfg.mean_value = np.array(cfg.pixel_means)[np.newaxis,
                                                np.newaxis, :].astype('float32')
     total_batch_size = total_batch_size if total_batch_size else batch_size
     if mode != 'infer':
         assert total_batch_size % batch_size == 0
-    if mode == 'train':
-        cfg.train_file_list = os.path.join(cfg.data_dir, cfg.train_file_list)
-        cfg.train_data_dir = os.path.join(cfg.data_dir, cfg.train_data_dir)
-    elif mode == 'test' or mode == 'infer':
-        cfg.val_file_list = os.path.join(cfg.data_dir, cfg.val_file_list)
-        cfg.val_data_dir = os.path.join(cfg.data_dir, cfg.val_data_dir)
-    json_dataset = JsonDataset(train=(mode == 'train'))
+    json_dataset = JsonDataset(mode)
     roidbs = json_dataset.get_roidb()
 
     print("{} on {} with {} roidbs".format(mode, cfg.dataset, len(roidbs)))
@@ -166,7 +147,7 @@ def coco(mode,
                 iter_id = count // device_num
                 if iter_id >= cfg.max_iter:
                     return
-        elif mode == "test":
+        elif mode == "val":
             batch_out = []
             for roidb in roidbs:
                 im, im_info, im_id = roidb_reader(roidb, mode)
@@ -194,7 +175,7 @@ def train(batch_size, total_batch_size=None, padding_total=False, shuffle=True):
 
 
 def test(batch_size, total_batch_size=None, padding_total=False):
-    return coco('test', batch_size, total_batch_size, shuffle=False)
+    return coco('val', batch_size, total_batch_size, shuffle=False)
 
 
 def infer():

fluid/PaddleCV/rcnn/roidbs.py

@@ -42,19 +42,33 @@ from config import cfg
 logger = logging.getLogger(__name__)
 
 
+class DatasetPath(object):
+    def __init__(self, mode):
+        self.mode = mode
+        mode_name = 'train' if mode == 'train' else 'val'
+        if cfg.dataset != 'coco2014' and cfg.dataset != 'coco2017':
+            raise NotImplementedError('Dataset {} not supported'.format(
+                cfg.dataset))
+        self.sub_name = mode_name + cfg.dataset[-4:]
+
+    def get_data_dir(self):
+        return os.path.join(cfg.data_dir, self.sub_name)
+
+    def get_file_list(self):
+        sfile_list = 'annotations/instances_' + self.sub_name + '.json'
+        return os.path.join(cfg.data_dir, sfile_list)
+
+
 class JsonDataset(object):
     """A class representing a COCO json dataset."""
 
-    def __init__(self, train=False):
+    def __init__(self, mode):
         print('Creating: {}'.format(cfg.dataset))
         self.name = cfg.dataset
-        self.is_train = train
-        if self.is_train:
-            data_dir = cfg.train_data_dir
-            file_list = cfg.train_file_list
-        else:
-            data_dir = cfg.val_data_dir
-            file_list = cfg.val_file_list
+        self.is_train = mode == 'train'
+        data_path = DatasetPath(mode)
+        data_dir = data_path.get_data_dir()
+        file_list = data_path.get_file_list()
         self.image_directory = data_dir
         self.COCO = COCO(file_list)
         # Set up dataset classes

fluid/PaddleCV/rcnn/train.py

@@ -36,7 +36,7 @@ def train():
     learning_rate = cfg.learning_rate
     image_shape = [3, cfg.TRAIN.max_size, cfg.TRAIN.max_size]
 
-    if cfg.debug or cfg.enable_ce:
+    if cfg.enable_ce:
         fluid.default_startup_program().random_seed = 1000
         fluid.default_main_program().random_seed = 1000
         import random
@@ -50,7 +50,7 @@ def train():
     use_random = True
     if cfg.enable_ce:
         use_random = False
-    model = model_builder.FasterRCNN(
+    model = model_builder.RCNN(
         add_conv_body_func=resnet.add_ResNet50_conv4_body,
         add_roi_box_head_func=resnet.add_ResNet_roi_conv5_head,
         use_pyreader=cfg.use_pyreader,
@@ -131,8 +131,8 @@ def train():
                 stats = {k: np.array(v).mean() for k, v in zip(keys, outs[:-1])}
                 train_stats.update(stats)
                 logs = train_stats.log()
-                strs = '{}, lr: {:.5f}, {}, time: {:.3f}'.format(
-                    now_time(),
+                strs = '{}, iter: {}, lr: {:.5f}, {}, time: {:.3f}'.format(
+                    now_time(), iter_id,
                     np.mean(outs[-1]), logs, start_time - prev_start_time)
                 print(strs)
                 sys.stdout.flush()
@@ -164,8 +164,8 @@ def train():
             stats = {k: np.array(v).mean() for k, v in zip(keys, outs[:-1])}
             train_stats.update(stats)
             logs = train_stats.log()
-            strs = '{}, lr: {:.5f}, {}, time: {:.3f}'.format(
-                now_time(),
+            strs = '{}, iter: {}, lr: {:.5f}, {}, time: {:.3f}'.format(
+                now_time(), iter_id,
                 np.mean(outs[-1]), logs, start_time - prev_start_time)
             print(strs)
             sys.stdout.flush()

fluid/PaddleCV/rcnn/utility.py

@@ -148,7 +148,7 @@ def parse_args():
     add_arg('variance',         float,  [1.,1.,1.,1.],    "The variance of anchors.")
     add_arg('rpn_stride',       float,  [16.,16.],    "Stride of the feature map that RPN is attached.")
     add_arg('rpn_nms_thresh',    float,   0.7,          "NMS threshold used on RPN proposals")
-    # TRAIN TEST INFER
+    # TRAIN VAL INFER
     add_arg('MASK_ON', bool, False, "Option for different models. If False, choose faster_rcnn. If True, choose mask_rcnn")
     add_arg('im_per_batch',       int,   1,        "Minibatch size.")
     add_arg('max_size',         int,   1333,    "The resized image height.")
@@ -158,7 +158,6 @@ def parse_args():
     add_arg('nms_thresh',    float, 0.5,    "NMS threshold.")
     add_arg('score_thresh',    float, 0.05,    "score threshold for NMS.")
     add_arg('snapshot_stride',  int,    10000,    "save model every snapshot stride.")
-    add_arg('debug',            bool,   False,   "Debug mode")
     # SINGLE EVAL AND DRAW
     add_arg('draw_threshold',  float, 0.8,    "Confidence threshold to draw bbox.")
     add_arg('image_path',       str,   'dataset/coco/val2017',  "The image path used to inference and visualize.")
@@ -172,5 +171,5 @@ def parse_args():
     if 'train' in file_name or 'profile' in file_name:
         merge_cfg_from_args(args, 'train')
     else:
-        merge_cfg_from_args(args, 'test')
+        merge_cfg_from_args(args, 'val')
     return args