From b21d12949dc4c1245716493dd86e305dd863da37 Mon Sep 17 00:00:00 2001 From: Guanghua Yu <742925032@qq.com> Date: Mon, 1 Nov 2021 15:24:45 +0800 Subject: [PATCH] Update PicoDet docs (#4407) --- configs/picodet/README.md | 55 ++++++++++++--------- configs/picodet/picodet_l_320_coco.yml | 11 +++-- configs/picodet/picodet_l_416_coco.yml | 11 +++-- configs/picodet/picodet_l_640_coco.yml | 11 +++-- deploy/lite/README.md | 35 ++++++------- ppdet/modeling/assigners/simota_assigner.py | 8 +-- 6 files changed, 74 insertions(+), 57 deletions(-) diff --git a/configs/picodet/README.md b/configs/picodet/README.md index c52d75586..2a75d7f68 100644 --- a/configs/picodet/README.md +++ b/configs/picodet/README.md @@ -1,12 +1,12 @@ -# PicoDet +# PP-PicoDet ![](../../docs/images/picedet_demo.jpeg) ## Introduction -We developed a series of lightweight models, which named `PicoDet`. Because of its excellent performance, it is very suitable for deployment on mobile or CPU. +We developed a series of lightweight models, which named `PP-PicoDet`. Because of its excellent performance, it is very suitable for deployment on mobile or CPU. - 🌟 Higher mAP: the **first** object detectors that surpass mAP(0.5:0.95) **30+** within 1M parameters when the input size is 416. -- 🚀 Faster latency: 129FPS on mobile ARM CPU. +- 🚀 Faster latency: 150FPS on mobile ARM CPU. - 😊 Deploy friendly: support PaddleLite/MNN/NCNN/OpenVINO and provide C++/Python/Android implementation. - 😍 Advanced algorithm: use the most advanced algorithms and innovate, such as ESNet, CSP-PAN, SimOTA with VFL, etc. @@ -17,40 +17,35 @@ We developed a series of lightweight models, which named `PicoDet`. Because of i ## Requirements - PaddlePaddle >= 2.1.2 -- PaddleSlim >= 2.1.1 ## Benchmark -| Model | Input size | mAPval
0.5:0.95 | mAPval
0.5 | Params
(M) | FLOPS
(G) | Latency*
(ms) | Latency#
(ms) | download | config | +| Model | Input size | mAPval
0.5:0.95 | mAPval
0.5 | Params
(M) | FLOPS
(G) | Latency[NCNN](#latency)
(ms) | Latency[Lite](#latency)
(ms) | download | config | | :-------- | :--------: | :---------------------: | :----------------: | :----------------: | :---------------: | :-----------------------------: | :-----------------------------: | :---------------------------------------------------------------------------------------------------------------------------------------------------------: | :------------------------------------------------------------------------------------------------------------ | | PicoDet-S | 320*320 | 27.1 | 41.4 | 0.99 | 0.73 | 8.13 | **6.65** | [model](https://paddledet.bj.bcebos.com/models/picodet_s_320_coco.pdparams) | [log](https://paddledet.bj.bcebos.com/logs/train_picodet_s_320_coco.log) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/picodet/picodet_s_320_coco.yml) | | PicoDet-S | 416*416 | 30.6 | 45.5 | 0.99 | 1.24 | 12.37 | **9.82** | [model](https://paddledet.bj.bcebos.com/models/picodet_s_416_coco.pdparams) | [log](https://paddledet.bj.bcebos.com/logs/train_picodet_s_416_coco.log) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/picodet/picodet_s_416_coco.yml) | | PicoDet-M | 320*320 | 30.9 | 45.7 | 2.15 | 1.48 | 11.27 | **9.61** | [model](https://paddledet.bj.bcebos.com/models/picodet_m_320_coco.pdparams) | [log](https://paddledet.bj.bcebos.com/logs/train_picodet_m_320_coco.log) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/picodet/picodet_m_320_coco.yml) | | PicoDet-M | 416*416 | 34.3 | 49.8 | 2.15 | 2.50 | 17.39 | **15.88** | [model](https://paddledet.bj.bcebos.com/models/picodet_m_416_coco.pdparams) | [log](https://paddledet.bj.bcebos.com/logs/train_picodet_m_416_coco.log) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/picodet/picodet_m_416_coco.yml) | -| PicoDet-L | 320*320 | 32.6 | 47.9 | 3.24 | 2.18 | 15.26 | **13.42** | [model](https://paddledet.bj.bcebos.com/models/picodet_l_320_coco.pdparams) | [log](https://paddledet.bj.bcebos.com/logs/train_picodet_l_320_coco.log) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/picodet/picodet_l_320_coco.yml) | -| PicoDet-L | 416*416 | 35.9 | 51.7 | 3.24 | 3.69 | 23.36 | **21.85** | [model](https://paddledet.bj.bcebos.com/models/picodet_l_416_coco.pdparams) | [log](https://paddledet.bj.bcebos.com/logs/train_picodet_l_416_coco.log) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/picodet/picodet_l_416_coco.yml) | -| PicoDet-L | 640*640 | 40.3 | 57.1 | 3.24 | 8.74 | 54.11 | **50.55** | [model](https://paddledet.bj.bcebos.com/models/picodet_l_640_coco.pdparams) | [log](https://paddledet.bj.bcebos.com/logs/train_picodet_l_640_coco.log) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/picodet/picodet_l_640_coco.yml) | +| PicoDet-L | 320*320 | 32.9 | 48.2 | 3.30 | 2.23 | 15.26 | **13.42** | [model](https://paddledet.bj.bcebos.com/models/picodet_l_320_coco.pdparams) | [log](https://paddledet.bj.bcebos.com/logs/train_picodet_l_320_coco.log) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/picodet/picodet_l_320_coco.yml) | +| PicoDet-L | 416*416 | 36.6 | 52.5 | 3.30 | 3.76 | 23.36 | **21.85** | [model](https://paddledet.bj.bcebos.com/models/picodet_l_416_coco.pdparams) | [log](https://paddledet.bj.bcebos.com/logs/train_picodet_l_416_coco.log) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/picodet/picodet_l_416_coco.yml) | +| PicoDet-L | 640*640 | 40.9 | 57.6 | 3.30 | 8.91 | 54.11 | **50.55** | [model](https://paddledet.bj.bcebos.com/models/picodet_l_640_coco.pdparams) | [log](https://paddledet.bj.bcebos.com/logs/train_picodet_l_640_coco.log) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/picodet/picodet_l_640_coco.yml) | -**Attetnion:** * represents NCNN inference speed, # represents Paddle-Lite inference speed. +#### More config -
-Table Notes (click to expand) - -- PicoDet inference speed is tested on SD 888(1*X1+3*A78+4*A55) with 4 threads by arm8 and with FP16. -- PicoDet is trained on COCO train2017 dataset and evaluated on COCO val2017. -- PicoDet used 4 or 8 GPUs for training and all checkpoints are trained to 300 epochs with default settings and hyperparameters. - -
- -## More config - -| Model | Input size | mAPval
0.5:0.95 | mAPval
0.5 | Params
(M) | FLOPS
(G) | Latency*
(ms) | Latency#
(ms) | download | config | +| Model | Input size | mAPval
0.5:0.95 | mAPval
0.5 | Params
(M) | FLOPS
(G) | Latency[NCNN](#latency)
(ms) | Latency[Lite](#latency)
(ms) | download | config | | :--------------------------- | :--------: | :---------------------: | :----------------: | :----------------: | :---------------: | :-----------------------------: | :-----------------------------: | :-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------: | :------------------------------------------------------------------------------------------------------------------------------------------- | | PicoDet-Shufflenetv2 1x | 416*416 | 30.0 | 44.6 | 1.17 | 1.53 | 15.06 | **10.63** | [model](https://paddledet.bj.bcebos.com/models/picodet_shufflenetv2_1x_416_coco.pdparams) | [log](https://paddledet.bj.bcebos.com/logs/train_picodet_shufflenetv2_1x_416_coco.log) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/picodet/more_config/picodet_shufflenetv2_1x_416_coco.yml) | | PicoDet-MobileNetv3-large 1x | 416*416 | 35.6 | 52.0 | 3.55 | 2.80 | 20.71 | **17.88** | [model](https://paddledet.bj.bcebos.com/models/picodet_mobilenetv3_large_1x_416_coco.pdparams) | [log](https://paddledet.bj.bcebos.com/logs/train_picodet_mobilenetv3_large_1x_416_coco.log) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/picodet/more_config/picodet_mobilenetv3_large_1x_416_coco.yml) | | PicoDet-LCNet 1.5x | 416*416 | 36.3 | 52.2 | 3.10 | 3.85 | 21.29 | **20.8** | [model](https://paddledet.bj.bcebos.com/models/picodet_lcnet_1_5x_416_coco.pdparams) | [log](https://paddledet.bj.bcebos.com/logs/train_picodet_lcnet_1_5x_416_coco.log) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/picodet/more_config/picodet_lcnet_1_5x_416_coco.yml) | -**Attetnion:** * represents NCNN inference speed, # represents Paddle-Lite inference speed. +
+Table Notes: + +- Latency: All our models test on `Qualcomm Snapdragon 865(4\*A77+4\*A55)` with 4 threads by arm8 and with FP16. In the above table, test latency on [NCNN](https://github.com/Tencent/ncnn) and `Lite`->[Paddle-Lite](https://github.com/PaddlePaddle/Paddle-Lite). And testing latency with code:[MobileDetBenchmark](https://github.com/JiweiMaster/MobileDetBenchmark). +- PicoDet is trained on COCO train2017 dataset and evaluated on COCO val2017. +- PicoDet used 4 or 8 GPUs for training and all checkpoints are trained with default settings and hyperparameters. + +
## Deployment @@ -134,6 +129,20 @@ paddle2onnx --model_dir output_inference/picodet_s_320_coco/ \ ### quantization +
+Requirements: + +- PaddlePaddle >= 2.2.0rc0 +- PaddleSlim >= 2.2.0rc0 + +**Install:** + +```shell +pip install paddleslim==2.2.0rc0 +``` + +
+
Quant aware (click to expand) @@ -156,6 +165,8 @@ python tools/post_quant.py -c configs/picodet/picodet_s_320_coco.yml \ --slim_config configs/slim/post_quant/picodet_s_ptq.yml ``` +- Notes: Now the accuracy of post quant is abnormal and it is being debugged. +
## Cite PiocDet diff --git a/configs/picodet/picodet_l_320_coco.yml b/configs/picodet/picodet_l_320_coco.yml index ea9aef722..ab3690b1e 100644 --- a/configs/picodet/picodet_l_320_coco.yml +++ b/configs/picodet/picodet_l_320_coco.yml @@ -20,16 +20,19 @@ ESNet: act: hard_swish channel_ratio: [0.875, 0.5, 1.0, 0.625, 0.5, 0.75, 0.625, 0.625, 0.5, 0.625, 1.0, 0.625, 0.75] +CSPPAN: + out_channels: 160 + PicoHead: conv_feat: name: PicoFeat - feat_in: 128 - feat_out: 128 + feat_in: 160 + feat_out: 160 num_convs: 4 num_fpn_stride: 4 norm_type: bn - share_cls_reg: False - feat_in_chan: 128 + share_cls_reg: True + feat_in_chan: 160 TrainReader: batch_size: 56 diff --git a/configs/picodet/picodet_l_416_coco.yml b/configs/picodet/picodet_l_416_coco.yml index 1a867d8b6..1492c2627 100644 --- a/configs/picodet/picodet_l_416_coco.yml +++ b/configs/picodet/picodet_l_416_coco.yml @@ -20,16 +20,19 @@ ESNet: act: hard_swish channel_ratio: [0.875, 0.5, 1.0, 0.625, 0.5, 0.75, 0.625, 0.625, 0.5, 0.625, 1.0, 0.625, 0.75] +CSPPAN: + out_channels: 160 + PicoHead: conv_feat: name: PicoFeat - feat_in: 128 - feat_out: 128 + feat_in: 160 + feat_out: 160 num_convs: 4 num_fpn_stride: 4 norm_type: bn - share_cls_reg: False - feat_in_chan: 128 + share_cls_reg: True + feat_in_chan: 160 TrainReader: batch_size: 48 diff --git a/configs/picodet/picodet_l_640_coco.yml b/configs/picodet/picodet_l_640_coco.yml index 1900aa552..4f6614c38 100644 --- a/configs/picodet/picodet_l_640_coco.yml +++ b/configs/picodet/picodet_l_640_coco.yml @@ -20,16 +20,19 @@ ESNet: act: hard_swish channel_ratio: [0.875, 0.5, 1.0, 0.625, 0.5, 0.75, 0.625, 0.625, 0.5, 0.625, 1.0, 0.625, 0.75] +CSPPAN: + out_channels: 160 + PicoHead: conv_feat: name: PicoFeat - feat_in: 128 - feat_out: 128 + feat_in: 160 + feat_out: 160 num_convs: 4 num_fpn_stride: 4 norm_type: bn - share_cls_reg: False - feat_in_chan: 128 + share_cls_reg: True + feat_in_chan: 160 TrainReader: batch_size: 32 diff --git a/deploy/lite/README.md b/deploy/lite/README.md index 5209206ff..dd3ddc5c9 100644 --- a/deploy/lite/README.md +++ b/deploy/lite/README.md @@ -24,7 +24,7 @@ Paddle Lite是飞桨轻量化推理引擎，为手机、IOT端提供高效推理 1. [**建议**]直接下载，预测库下载链接如下： |平台|预测库下载链接| |-|-| - |Android|[arm7](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.9.1/inference_lite_lib.android.armv7.clang.c++_static.with_extra.with_cv.tar.gz) / [arm8](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.9.1/inference_lite_lib.android.armv8.clang.c++_static.with_extra.with_cv.tar.gz)| + |Android|[arm7](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.10-rc/inference_lite_lib.android.armv7.clang.c++_static.with_extra.with_cv.tar.gz) / [arm8](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.10-rc/inference_lite_lib.android.armv8.clang.c++_static.with_extra.with_cv.tar.gz)| **注意**：1. 如果是从 Paddle-Lite [官方文档](https://paddle-lite.readthedocs.io/zh/latest/quick_start/release_lib.html#android-toolchain-gcc)下载的预测库，注意选择`with_extra=ON，with_cv=ON`的下载链接。2. 目前只提供Android端demo，IOS端demo可以参考[Paddle-Lite IOS demo](https://github.com/PaddlePaddle/Paddle-Lite-Demo/tree/master/PaddleLite-ios-demo) @@ -77,13 +77,13 @@ Paddle-Lite 提供了多种策略来自动优化原始的模型，其中包括 **注意**：如果已经准备好了 `.nb` 结尾的模型文件，可以跳过此步骤。 #### 2.1.1 安装paddle_lite_opt工具 -安装paddle_lite_opt工具有如下两种方法： +安装`paddle_lite_opt`工具有如下两种方法： 1. [**建议**]pip安装paddlelite并进行转换 ```shell - pip install paddlelite + pip install paddlelite==2.10rc ``` -2. 源码编译Paddle-Lite生成opt工具 +2. 源码编译Paddle-Lite生成`paddle_lite_opt`工具 模型优化需要Paddle-Lite的`opt`可执行文件，可以通过编译Paddle-Lite源码获得，编译步骤如下： ```shell @@ -120,23 +120,24 @@ Paddle-Lite 提供了多种策略来自动优化原始的模型，其中包括 #### 2.1.3 转换示例 -下面以PaddleDetection中的 `ppyolo` 模型为例，介绍使用`paddle_lite_opt`完成预训练模型到inference模型，再到Paddle-Lite优化模型的转换。 +下面以PaddleDetection中的 `PicoDet` 模型为例，介绍使用`paddle_lite_opt`完成预训练模型到inference模型，再到Paddle-Lite优化模型的转换。 ```shell # 进入PaddleDetection根目录 cd PaddleDetection_root_path # 将预训练模型导出为inference模型 -python tools/export_model.py -c configs/ppyolo/ppyolo_tiny_650e_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyolo_tiny_650e_coco.pdparams +python tools/export_model.py -c configs/picodet/picodet_s_320_coco.yml \ + -o weights=https://paddledet.bj.bcebos.com/models/picodet_s_320_coco.pdparams --output_dir=output_inference # 将inference模型转化为Paddle-Lite优化模型 -paddle_lite_opt --valid_targets=arm --model_file=output_inference/ppyolo_tiny_650e_coco/model.pdmodel --param_file=output_inference/ppyolo_tiny_650e_coco/model.pdiparams --optimize_out=output_inference/ppyolo_tiny_650e_coco/model +paddle_lite_opt --valid_targets=arm --model_file=output_inference/picodet_s_320_coco/model.pdmodel --param_file=output_inference/picodet_s_320_coco/model.pdiparams --optimize_out=output_inference/picodet_s_320_coco/model # 将inference模型配置转化为json格式 -python deploy/lite/convert_yml_to_json.py output_inference/ppyolo_tiny_650e_coco/infer_cfg.yml +python deploy/lite/convert_yml_to_json.py output_inference/picodet_s_320_coco/infer_cfg.yml ``` -最终在output_inference/ppyolo_tiny_650e_coco/文件夹下生成`ppyolo_tiny.nb` 和 `infer_cfg.json`的文件。 +最终在output_inference/picodet_s_320_coco/文件夹下生成`model.nb` 和 `infer_cfg.json`的文件。 **注意**：`--optimize_out` 参数为优化后模型的保存路径，无需加后缀`.nb`；`--model_file` 参数为模型结构信息文件的路径，`--param_file` 参数为模型权重信息文件的路径，请注意文件名。 @@ -185,7 +186,7 @@ cp -r Makefile src/ include/ *runtime_config.json $inference_lite_path/demo/cxx/ cd $inference_lite_path/demo/cxx/lite # 执行编译，等待完成后得到可执行文件main -make ARM_ABI = arm8 +make ARM_ABI=arm8 #如果是arm7，则执行 make ARM_ABI = arm7 (或者在Makefile中修改该项) ``` @@ -200,10 +201,10 @@ mkdir model_det mkdir model_keypoint # 将优化后的模型、预测库文件、测试图像放置在预测库中的demo/cxx/detection文件夹下 -cp {PadddleDetection_Root}/output_inference/ppyolo_tiny_650e_coco/model.nb ./model_det/ -cp {PadddleDetection_Root}/output_inference/ppyolo_tiny_650e_coco/infer_cfg.json ./model_det/ +cp {PadddleDetection_Root}/output_inference/picodet_s_320_coco/model.nb ./model_det/ +cp {PadddleDetection_Root}/output_inference/picodet_s_320_coco/infer_cfg.json ./model_det/ -# 如果需要关键点模型，则只需一下操作 +# 如果需要关键点模型，则只需操作： cp {PadddleDetection_Root}/output_inference/hrnet_w32_256x192/model.nb ./model_keypoint/ cp {PadddleDetection_Root}/output_inference/hrnet_w32_256x192/infer_cfg.json ./model_keypoint/ @@ -219,10 +220,10 @@ cp ../../../cxx/lib/libpaddle_light_api_shared.so ./ ``` deploy/ |-- model_det/ -| |--mdoel.nb 优化后的检测模型文件 +| |--model.nb 优化后的检测模型文件 | |--infer_cfg.json 检测器模型配置文件 |-- model_keypoint/ -| |--mdoel.nb 优化后的关键点模型文件 +| |--model.nb 优化后的关键点模型文件 | |--infer_cfg.json 关键点模型配置文件 |-- main 生成的移动端执行文件 |-- det_runtime_config.json 目标检测执行时参数配置文件 @@ -240,7 +241,7 @@ deploy/ "threshold_det": 0.5, #检测器输出阈值 "image_file": "demo.jpg", #测试图片 "image_dir": "", #测试图片文件夹 - "run_benchmark": false, #性能测试开关 + "run_benchmark": true, #性能测试开关 "cpu_threads": 4 #线程数 } ``` @@ -253,7 +254,7 @@ deploy/ "threshold_keypoint": 0.5, #关键点输出阈值 "image_file": "demo.jpg", #测试图片 "image_dir": "", #测试图片文件夹 - "run_benchmark": false, #性能测试开关 + "run_benchmark": true, #性能测试开关 "cpu_threads": 4 #线程数 } ``` diff --git a/ppdet/modeling/assigners/simota_assigner.py b/ppdet/modeling/assigners/simota_assigner.py index c0f337e46..34e26eee6 100644 --- a/ppdet/modeling/assigners/simota_assigner.py +++ b/ppdet/modeling/assigners/simota_assigner.py @@ -129,15 +129,11 @@ class SimOTAAssigner(object): valid_mask[valid_mask.copy()] = fg_mask_inboxes matched_gt_inds = matching_matrix[fg_mask_inboxes, :].argmax(1) - matched_pred_ious = (matching_matrix * - pairwise_ious.numpy()).sum(1)[fg_mask_inboxes] - matched_pred_ious = paddle.to_tensor( - matched_pred_ious, place=pairwise_ious.place) matched_gt_inds = paddle.to_tensor( matched_gt_inds, place=pairwise_ious.place) - return matched_pred_ious, matched_gt_inds, valid_mask + return matched_gt_inds, valid_mask def get_sample(self, assign_gt_inds, gt_bboxes): pos_inds = np.unique(np.nonzero(assign_gt_inds > 0)[0]) @@ -231,7 +227,7 @@ class SimOTAAssigner(object): paddle.logical_not(is_in_boxes_and_center).cast('float32') * INF ) - matched_pred_ious, matched_gt_inds, valid_mask = \ + matched_gt_inds, valid_mask = \ self.dynamic_k_matching( cost_matrix, pairwise_ious, num_gt, valid_mask.numpy()) -- GitLab