Merge branch 'PaddlePaddle:develop' into develop

README_ch.md

@@ -8,7 +8,7 @@
 
 **近期更新**
 
-- 2021.07.08、07.27 添加26个[FAQ](docs/zh_CN/faq_series/faq_2021_s2.md)
+- 2021.08.11 更新7个[FAQ](docs/zh_CN/faq_series/faq_2021_s2.md)。
 - 2021.06.29 添加Swin-transformer系列模型，ImageNet1k数据集上Top1 acc最高精度可达87.2%；支持训练预测评估与whl包部署，预训练模型可以从[这里](docs/zh_CN/models/models_intro.md)下载。
 - 2021.06.22,23,24 PaddleClas官方研发团队带来技术深入解读三日直播课。课程回放：[https://aistudio.baidu.com/aistudio/course/introduce/24519](https://aistudio.baidu.com/aistudio/course/introduce/24519)
 - 2021.06.16 PaddleClas v2.2版本升级，集成Metric learning，向量检索等组件。新增商品识别、动漫人物识别、车辆识别和logo识别等4个图像识别应用。新增LeViT、Twins、TNT、DLA、HarDNet、RedNet系列30个预训练模型。

docs/zh_CN/application/mainbody_detection.md

@@ -167,4 +167,22 @@ python tools/export_model.py -c configs/ppyolo/ppyolov2_r50vd_dcn_365e_coco.yml
 
 更多模型导出教程，请参考：[EXPORT_MODEL](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.1/deploy/EXPORT_MODEL.md)
 
-导出模型之后，在主体检测与识别任务中，就可以将检测模型的路径更改为该inference模型路径，完成预测。图像识别快速体验可以参考：[图像识别快速开始教程](../tutorials/quick_start_recognition.md)。
+最终，目录`inference/ppyolov2_r50vd_dcn_365e_coco`中包含`inference.pdiparams`, `inference.pdiparams.info` 以及 `inference.pdmodel` 文件，其中`inference.pdiparams`为保存的inference模型权重文件，`inference.pdmodel`为保存的inference模型结构文件。
+
+
+导出模型之后，在主体检测与识别任务中，就可以将检测模型的路径更改为该inference模型路径，完成预测。
+
+以商品识别为例，其配置文件为[inference_product.yaml](../../../deploy/configs/inference_product.yaml)，修改其中的`Global.det_inference_model_dir`字段为导出的主体检测inference模型目录，参考[图像识别快速开始教程](../tutorials/quick_start_recognition.md)，即可完成商品检测与识别过程。
+
+
+### FAQ
+
+#### Q：可以使用其他的主体检测模型结构吗？
+
+* A：可以的，但是目前的检测预处理过程仅适配yolo系列的预处理，因此在使用的时候，建议优先使用yolo系列的模型进行训练，如果希望使用faster rcnn等其他系列的模型，需要按照PaddleDetection的数据预处理，修改下预处理逻辑，这块如果您有需求或者有问题的话，欢迎提issue或者在群里反馈。
+
+#### Q：可以修改主体检测的预测尺度吗？
+
+* A：可以的，但是需要注意2个地方
+  * PaddleClas中提供的主体检测模型是基于640x640的分辨率去训练的，因此预测的时候也是默认使用640x640的分辨率进行预测，使用其他分辨率预测的话，精度会有所降低。
+  * 在模型导出的时候，建议也修改下模型导出的分辨率，保持模型导出、模型预测的分辨率一致。

docs/zh_CN/tutorials/getting_started.md

@@ -244,7 +244,7 @@ python3 python/predict_cls.py \
     -c configs/inference_cls.yaml \
     -o Global.infer_imgs=../dataset/flowers102/jpg/image_00001.jpg \
     -o Global.inference_model_dir=../inference/ \
-    -o PostProcess.class_id_map_file=None
+    -o PostProcess.Topk.class_id_map_file=None
 
 
 其中：

docs/zh_CN/tutorials/quick_start_professional.md

@@ -128,7 +128,7 @@ python3 -m paddle.distributed.launch \
 PaddleClas包含了自研的SSLD知识蒸馏方案，具体的内容可以参考[知识蒸馏章节](../advanced_tutorials/distillation/distillation.md), 本小节将尝试使用知识蒸馏技术对MobileNetV3_large_x1_0模型进行训练，使用`2.1.2小节`训练得到的ResNet50_vd模型作为蒸馏所用的教师模型，首先将`2.1.2小节`训练得到的ResNet50_vd模型保存到指定目录，脚本如下。
 
 ```shell
-mkdir pretrained 
+mkdir pretrained
 cp -r output_CIFAR/ResNet50_vd/best_model.pdparams  ./pretrained/
 ```
 
@@ -256,5 +256,5 @@ PreProcess:
 python3 python/predict_cls.py \
     -c configs/inference_cls.yaml \
     -o Global.infer_imgs=../dataset/CIFAR100/test/0/0001.png \
-    -o PostProcess.class_id_map_file=None
+    -o PostProcess.Topk.class_id_map_file=None
 ```

ppcls/arch/backbone/base/theseus_layer.py

@@ -12,15 +12,9 @@ class Identity(nn.Layer):
 
 
 class TheseusLayer(nn.Layer):
-    def __init__(self, *args, return_patterns=None, **kwargs):
+    def __init__(self, *args, **kwargs):
         super(TheseusLayer, self).__init__()
-        self.res_dict = None
-        if return_patterns is not None:
-            self._update_res(return_patterns)
-
-    def forward(self, *input, res_dict=None, **kwargs):
-        if res_dict is not None:
-            self.res_dict = res_dict
+        self.res_dict = {}
 
     # stop doesn't work when stop layer has a parallel branch.
     def stop_after(self, stop_layer_name: str):
@@ -38,33 +32,43 @@ class TheseusLayer(nn.Layer):
                     stop_layer_name)
         return after_stop
 
-    def _update_res(self, return_layers):
+    def update_res(self, return_patterns):
+        if not return_patterns or isinstance(self, WrapLayer):
+            return
+        for layer_i in self._sub_layers:
+            layer_name = self._sub_layers[layer_i].full_name()
+            if isinstance(self._sub_layers[layer_i], (nn.Sequential, nn.LayerList)):
+                self._sub_layers[layer_i] = wrap_theseus(self._sub_layers[layer_i])
+                self._sub_layers[layer_i].res_dict = self.res_dict
+                self._sub_layers[layer_i].update_res(return_patterns)
+            else:
+                for return_pattern in return_patterns:
+                    if re.match(return_pattern, layer_name):
+                        if not isinstance(self._sub_layers[layer_i], TheseusLayer):
+                            self._sub_layers[layer_i] = wrap_theseus(self._sub_layers[layer_i])
+                        self._sub_layers[layer_i].register_forward_post_hook(
+                            self._sub_layers[layer_i]._save_sub_res_hook)
+                        self._sub_layers[layer_i].res_dict = self.res_dict
+                if isinstance(self._sub_layers[layer_i], TheseusLayer):
+                    self._sub_layers[layer_i].res_dict = self.res_dict
+                    self._sub_layers[layer_i].update_res(return_patterns)
+
+    def _save_sub_res_hook(self, layer, input, output):
+        self.res_dict[layer.full_name()] = output
+
+    def replace_sub(self, layer_name_pattern, replace_function, recursive=True):
         for layer_i in self._sub_layers:
             layer_name = self._sub_layers[layer_i].full_name()
-            for return_pattern in return_layers:
-                if return_layers is not None and re.match(return_pattern,
-                                                          layer_name):
-                    self._sub_layers[layer_i].register_forward_post_hook(
-                        self._save_sub_res_hook)
-
-    def replace_sub(self, layer_name_pattern, replace_function,
-                    recursive=True):
-        for k in self._sub_layers.keys():
-            layer_name = self._sub_layers[k].full_name()
             if re.match(layer_name_pattern, layer_name):
-                self._sub_layers[k] = replace_function(self._sub_layers[k])
+                self._sub_layers[layer_i] = replace_function(self._sub_layers[layer_i])
             if recursive:
-                if isinstance(self._sub_layers[k], TheseusLayer):
-                    self._sub_layers[k].replace_sub(
+                if isinstance(self._sub_layers[layer_i], TheseusLayer):
+                    self._sub_layers[layer_i].replace_sub(
                         layer_name_pattern, replace_function, recursive)
-                elif isinstance(self._sub_layers[k],
-                                nn.Sequential) or isinstance(
-                                    self._sub_layers[k], nn.LayerList):
-                    for kk in self._sub_layers[k]._sub_layers.keys():
-                        self._sub_layers[k]._sub_layers[kk].replace_sub(
+                elif isinstance(self._sub_layers[layer_i], (nn.Sequential, nn.LayerList)):
+                    for layer_j in self._sub_layers[layer_i]._sub_layers:
+                        self._sub_layers[layer_i]._sub_layers[layer_j].replace_sub(
                             layer_name_pattern, replace_function, recursive)
-                else:
-                    pass
 
     '''
     example of replace function:
@@ -78,3 +82,40 @@ class TheseusLayer(nn.Layer):
         return new_conv
 
         '''
+
+
+class WrapLayer(TheseusLayer):
+    def __init__(self, sub_layer):
+        super(WrapLayer, self).__init__()
+        self.sub_layer = sub_layer
+        self.name = sub_layer.full_name()
+
+    def full_name(self):
+        return self.name
+
+    def forward(self, *inputs, **kwargs):
+        return self.sub_layer(*inputs, **kwargs)
+
+    def update_res(self, return_patterns):
+        if not return_patterns or not isinstance(self.sub_layer, (nn.Sequential, nn.LayerList)):
+            return
+        for layer_i in self.sub_layer._sub_layers:
+            if isinstance(self.sub_layer._sub_layers[layer_i], (nn.Sequential, nn.LayerList)):
+                self.sub_layer._sub_layers[layer_i] = wrap_theseus(self.sub_layer._sub_layers[layer_i])
+                self.sub_layer._sub_layers[layer_i].res_dict = self.res_dict
+                self.sub_layer._sub_layers[layer_i].update_res(return_patterns)
+
+            layer_name = self.sub_layer._sub_layers[layer_i].full_name()
+            for return_pattern in return_patterns:
+                if re.match(return_pattern, layer_name):
+                    self.sub_layer._sub_layers[layer_i].res_dict = self.res_dict
+                    self.sub_layer._sub_layers[layer_i].register_forward_post_hook(
+                        self._sub_layers[layer_i]._save_sub_res_hook)
+
+            if isinstance(self.sub_layer._sub_layers[layer_i], TheseusLayer):
+                self.sub_layer._sub_layers[layer_i].update_res(return_patterns)
+
+
+def wrap_theseus(sub_layer):
+    wrapped_layer = WrapLayer(sub_layer)
+    return wrapped_layer

ppcls/arch/backbone/legendary_models/vgg.py

@@ -111,7 +111,7 @@ class VGGNet(TheseusLayer):
         model: nn.Layer. Specific VGG model depends on args.
     """
 
-    def __init__(self, config, stop_grad_layers=0, class_num=1000):
+    def __init__(self, config, stop_grad_layers=0, class_num=1000, return_patterns=None):
         super().__init__()
 
         self.stop_grad_layers = stop_grad_layers
@@ -138,7 +138,7 @@ class VGGNet(TheseusLayer):
         self.fc2 = Linear(4096, 4096)
         self.fc3 = Linear(4096, class_num)
 
-    def forward(self, inputs):
+    def forward(self, inputs, res_dict=None):
         x = self.conv_block_1(inputs)
         x = self.conv_block_2(x)
         x = self.conv_block_3(x)
@@ -152,6 +152,9 @@ class VGGNet(TheseusLayer):
         x = self.relu(x)
         x = self.drop(x)
         x = self.fc3(x)
+        if self.res_dict and res_dict is not None:
+            for res_key in list(self.res_dict):
+                res_dict[res_key] = self.res_dict.pop(res_key)
         return x
 
 

ppcls/arch/backbone/model_zoo/resnext101_wsl.py

@@ -12,7 +12,7 @@ MODEL_URLS = {
     "ResNeXt101_32x8d_wsl":
     "https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ResNeXt101_32x8d_wsl_pretrained.pdparams",
     "ResNeXt101_32x16d_wsl":
-    "https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ResNeXt101_32x816_wsl_pretrained.pdparams",
+    "https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ResNeXt101_32x16_wsl_pretrained.pdparams",
     "ResNeXt101_32x32d_wsl":
     "https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ResNeXt101_32x32d_wsl_pretrained.pdparams",
     "ResNeXt101_32x48d_wsl":

ppcls/configs/ImageNet/GhostNet/GhostNet_x0_5.yaml

@@ -24,6 +24,7 @@ Loss:
   Train:
     - CELoss:
         weight: 1.0
+        epsilon: 0.1
   Eval:
     - CELoss:
         weight: 1.0
@@ -35,9 +36,10 @@ Optimizer:
   lr:
     name: Cosine
     learning_rate: 0.8
+    warmup_epoch: 5
   regularizer:
     name: 'L2'
-    coeff: 0.0004
+    coeff: 0.00004
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/GhostNet/GhostNet_x1_0.yaml

@@ -24,6 +24,7 @@ Loss:
   Train:
     - CELoss:
         weight: 1.0
+        epsilon: 0.1
   Eval:
     - CELoss:
         weight: 1.0
@@ -35,9 +36,10 @@ Optimizer:
   lr:
     name: Cosine
     learning_rate: 0.8
+    warmup_epoch: 5
   regularizer:
     name: 'L2'
-    coeff: 0.0004
+    coeff: 0.00004
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/GhostNet/GhostNet_x1_3.yaml

@@ -24,6 +24,7 @@ Loss:
   Train:
     - CELoss:
         weight: 1.0
+        epsilon: 0.1
   Eval:
     - CELoss:
         weight: 1.0
@@ -35,9 +36,10 @@ Optimizer:
   lr:
     name: Cosine
     learning_rate: 0.8
+    warmup_epoch: 5
   regularizer:
     name: 'L2'
-    coeff: 0.0004
+    coeff: 0.00004
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/MobileNetV1/MobileNetV1.yaml

@@ -41,7 +41,7 @@ Optimizer:
     values: [0.1, 0.01, 0.001, 0.0001]
   regularizer:
     name: 'L2'
-    coeff: 0.0003
+    coeff: 0.00003
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/MobileNetV1/MobileNetV1_x0_25.yaml

@@ -39,7 +39,7 @@ Optimizer:
     values: [0.1, 0.01, 0.001, 0.0001]
   regularizer:
     name: 'L2'
-    coeff: 0.0003
+    coeff: 0.00003
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/MobileNetV1/MobileNetV1_x0_5.yaml

@@ -39,7 +39,7 @@ Optimizer:
     values: [0.1, 0.01, 0.001, 0.0001]
   regularizer:
     name: 'L2'
-    coeff: 0.0003
+    coeff: 0.00003
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/MobileNetV1/MobileNetV1_x0_75.yaml

@@ -39,7 +39,7 @@ Optimizer:
     values: [0.1, 0.01, 0.001, 0.0001]
   regularizer:
     name: 'L2'
-    coeff: 0.0003
+    coeff: 0.00003
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/MobileNetV2/MobileNetV2.yaml

@@ -39,7 +39,7 @@ Optimizer:
     learning_rate: 0.045
   regularizer:
     name: 'L2'
-    coeff: 0.0004
+    coeff: 0.00004
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/MobileNetV2/MobileNetV2_x0_25.yaml

@@ -37,7 +37,7 @@ Optimizer:
     learning_rate: 0.045
   regularizer:
     name: 'L2'
-    coeff: 0.0003
+    coeff: 0.00003
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/MobileNetV2/MobileNetV2_x0_5.yaml

@@ -37,7 +37,7 @@ Optimizer:
     learning_rate: 0.045
   regularizer:
     name: 'L2'
-    coeff: 0.0003
+    coeff: 0.00003
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/MobileNetV2/MobileNetV2_x0_75.yaml

@@ -37,7 +37,7 @@ Optimizer:
     learning_rate: 0.045
   regularizer:
     name: 'L2'
-    coeff: 0.0004
+    coeff: 0.00004
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/MobileNetV2/MobileNetV2_x1_5.yaml

@@ -37,7 +37,7 @@ Optimizer:
     learning_rate: 0.045
   regularizer:
     name: 'L2'
-    coeff: 0.0004
+    coeff: 0.00004
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/MobileNetV2/MobileNetV2_x2_0.yaml

@@ -37,7 +37,7 @@ Optimizer:
     learning_rate: 0.045
   regularizer:
     name: 'L2'
-    coeff: 0.0004
+    coeff: 0.00004
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/ShuffleNet/ShuffleNetV2_swish.yaml

+# global configs
+Global:
+  checkpoints: null
+  pretrained_model: null
+  output_dir: ./output/
+  device: gpu
+  save_interval: 1
+  eval_during_train: True
+  eval_interval: 1
+  epochs: 240
+  print_batch_step: 10
+  use_visualdl: False
+  # used for static mode and model export
+  image_shape: [3, 224, 224]
+  save_inference_dir: ./inference
+
+# model architecture
+Arch:
+  name: ShuffleNetV2_swish
+  class_num: 1000
+ 
+# loss function config for traing/eval process
+Loss:
+  Train:
+    - CELoss:
+        weight: 1.0
+  Eval:
+    - CELoss:
+        weight: 1.0
+
+
+Optimizer:
+  name: Momentum
+  momentum: 0.9
+  lr:
+    name: Cosine
+    learning_rate: 0.5
+    warmup_epoch: 5
+  regularizer:
+    name: 'L2'
+    coeff: 0.00004
+
+
+# data loader for train and eval
+DataLoader:
+  Train:
+    dataset:
+      name: ImageNetDataset
+      image_root: ./dataset/ILSVRC2012/
+      cls_label_path: ./dataset/ILSVRC2012/train_list.txt
+      transform_ops:
+        - DecodeImage:
+            to_rgb: True
+            channel_first: False
+        - RandCropImage:
+            size: 224
+        - RandFlipImage:
+            flip_code: 1
+        - NormalizeImage:
+            scale: 1.0/255.0
+            mean: [0.485, 0.456, 0.406]
+            std: [0.229, 0.224, 0.225]
+            order: ''
+
+    sampler:
+      name: DistributedBatchSampler
+      batch_size: 256
+      drop_last: False
+      shuffle: True
+    loader:
+      num_workers: 4
+      use_shared_memory: True
+
+  Eval:
+    dataset: 
+      name: ImageNetDataset
+      image_root: ./dataset/ILSVRC2012/
+      cls_label_path: ./dataset/ILSVRC2012/val_list.txt
+      transform_ops:
+        - DecodeImage:
+            to_rgb: True
+            channel_first: False
+        - ResizeImage:
+            resize_short: 256
+        - CropImage:
+            size: 224
+        - NormalizeImage:
+            scale: 1.0/255.0
+            mean: [0.485, 0.456, 0.406]
+            std: [0.229, 0.224, 0.225]
+            order: ''
+    sampler:
+      name: DistributedBatchSampler
+      batch_size: 64
+      drop_last: False
+      shuffle: False
+    loader:
+      num_workers: 4
+      use_shared_memory: True
+
+Infer:
+  infer_imgs: docs/images/whl/demo.jpg
+  batch_size: 10
+  transforms:
+    - DecodeImage:
+        to_rgb: True
+        channel_first: False
+    - ResizeImage:
+        resize_short: 256
+    - CropImage:
+        size: 224
+    - NormalizeImage:
+        scale: 1.0/255.0
+        mean: [0.485, 0.456, 0.406]
+        std: [0.229, 0.224, 0.225]
+        order: ''
+    - ToCHWImage:
+  PostProcess:
+    name: Topk
+    topk: 5
+    class_id_map_file: ppcls/utils/imagenet1k_label_list.txt
+
+Metric:
+  Train:
+    - TopkAcc:
+        topk: [1, 5]
+  Eval:
+    - TopkAcc:
+        topk: [1, 5]

ppcls/configs/ImageNet/ShuffleNet/ShuffleNetV2_x0_25.yaml

@@ -38,7 +38,7 @@ Optimizer:
     warmup_epoch: 5
   regularizer:
     name: 'L2'
-    coeff: 0.0003
+    coeff: 0.00003
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/ShuffleNet/ShuffleNetV2_x0_33.yaml

@@ -38,7 +38,7 @@ Optimizer:
     warmup_epoch: 5
   regularizer:
     name: 'L2'
-    coeff: 0.0003
+    coeff: 0.00003
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/ShuffleNet/ShuffleNetV2_x0_5.yaml

@@ -38,7 +38,7 @@ Optimizer:
     warmup_epoch: 5
   regularizer:
     name: 'L2'
-    coeff: 0.0003
+    coeff: 0.00003
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/ShuffleNet/ShuffleNetV2_x1_0.yaml

@@ -38,7 +38,7 @@ Optimizer:
     warmup_epoch: 5
   regularizer:
     name: 'L2'
-    coeff: 0.0004
+    coeff: 0.00004
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/ShuffleNet/ShuffleNetV2_x1_5.yaml

@@ -38,7 +38,7 @@ Optimizer:
     warmup_epoch: 5
   regularizer:
     name: 'L2'
-    coeff: 0.0004
+    coeff: 0.00004
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/ShuffleNet/ShuffleNetV2_x2_0.yaml

@@ -38,7 +38,7 @@ Optimizer:
     warmup_epoch: 5
   regularizer:
     name: 'L2'
-    coeff: 0.0004
+    coeff: 0.00004
 
 
 # data loader for train and eval

ppcls/engine/trainer.py

@@ -588,7 +588,7 @@ class Trainer(object):
             if len(batch) == 3:
                 has_unique_id = True
                 batch[2] = batch[2].reshape([-1, 1]).astype("int64")
-            out = self.model(batch[0], batch[1])
+            out = self.forward(batch)
             batch_feas = out["features"]
 
             # do norm
@@ -653,7 +653,7 @@ class Trainer(object):
             image_file_list.append(image_file)
             if len(batch_data) >= batch_size or idx == len(image_list) - 1:
                 batch_tensor = paddle.to_tensor(batch_data)
-                out = self.model(batch_tensor)
+                out = self.forward([batch_tensor])
                 if isinstance(out, list):
                     out = out[0]
                 result = postprocess_func(out, image_file_list)

tests/test.sh

@@ -185,7 +185,7 @@ function func_inference(){
         elif [ ${use_gpu} = "True" ] || [ ${use_gpu} = "gpu" ]; then
             for use_trt in ${use_trt_list[*]}; do
                 for precision in ${precision_list[*]}; do
-                    if [ ${precision} = "False" ] && [ ${use_trt} = "False" ]; then
+                    if [ ${precision} = "True" ] && [ ${use_trt} = "False" ]; then
                         continue
                     fi
                     if [[ ${use_trt} = "False" || ${precision} =~ "int8" ]] && [ ${_flag_quant} = "True" ]; then