fix rec distillation (#3995)

* fix rec distillation

* add dist cfg

* fix cfg

configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec.yml

@@ -4,7 +4,7 @@ Global:
   epoch_num: 800
   log_smooth_window: 20
   print_batch_step: 10
-  save_model_dir: ./output/rec_chinese_lite_distillation_v2.1
+  save_model_dir: ./output/rec_mobile_pp-OCRv2
   save_epoch_step: 3
   eval_batch_step: [0, 2000]
   cal_metric_during_train: true
@@ -19,7 +19,7 @@ Global:
   infer_mode: false
   use_space_char: true
   distributed: true
-  save_res_path: ./output/rec/predicts_chinese_lite_distillation_v2.1.txt
+  save_res_path: ./output/rec/predicts_mobile_pp-OCRv2.txt
 
 
 Optimizer:
@@ -35,16 +35,9 @@ Optimizer:
     name: L2
     factor: 2.0e-05
 
+
 Architecture:
-  model_type: &model_type "rec"
-  name: DistillationModel
-  algorithm: Distillation
-  Models:
-    Teacher:
-      pretrained:
-      freeze_params: false
-      return_all_feats: true
-      model_type: *model_type
+  model_type: rec
   algorithm: CRNN
   Transform:
   Backbone:
@@ -58,56 +51,16 @@ Architecture:
     name: CTCHead
     mid_channels: 96
     fc_decay: 0.00002
-    Student:
-      pretrained:
-      freeze_params: false
-      return_all_feats: true
-      model_type: *model_type
-      algorithm: CRNN
-      Transform:
-      Backbone:
-        name: MobileNetV1Enhance
-        scale: 0.5
-      Neck:
-        name: SequenceEncoder
-        encoder_type: rnn
-        hidden_size: 64
-      Head:
-        name: CTCHead
-        mid_channels: 96
-        fc_decay: 0.00002
-  
 
 Loss:
-  name: CombinedLoss
-  loss_config_list:
-  - DistillationCTCLoss:
-      weight: 1.0
-      model_name_list: ["Student", "Teacher"]
-      key: head_out
-  - DistillationDMLLoss:
-      weight: 1.0
-      act: "softmax"
-      model_name_pairs:
-      - ["Student", "Teacher"]
-      key: head_out
-  - DistillationDistanceLoss:
-      weight: 1.0
-      mode: "l2"
-      model_name_pairs:
-      - ["Student", "Teacher"]
-      key: backbone_out
+  name: CTCLoss
 
 PostProcess:
-  name: DistillationCTCLabelDecode
-  model_name: ["Student", "Teacher"]
-  key: head_out
+  name: CTCLabelDecode
 
 Metric:
-  name: DistillationMetric
-  base_metric_name: RecMetric
+  name: RecMetric
   main_indicator: acc
-  key: "Student"
 
 Train:
   dataset:
@@ -132,7 +85,6 @@ Train:
     shuffle: true
     batch_size_per_card: 128
     drop_last: true
-    num_sections: 1
     num_workers: 8
 Eval:
   dataset:

configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec_distillation.yml

+Global:
+  debug: false
+  use_gpu: true
+  epoch_num: 800
+  log_smooth_window: 20
+  print_batch_step: 10
+  save_model_dir: ./output/rec_pp-OCRv2_distillation
+  save_epoch_step: 3
+  eval_batch_step: [0, 2000]
+  cal_metric_during_train: true
+  pretrained_model:
+  checkpoints:
+  save_inference_dir:
+  use_visualdl: false
+  infer_img: doc/imgs_words/ch/word_1.jpg
+  character_dict_path: ppocr/utils/ppocr_keys_v1.txt
+  character_type: ch
+  max_text_length: 25
+  infer_mode: false
+  use_space_char: true
+  distributed: true
+  save_res_path: ./output/rec/predicts_pp-OCRv2_distillation.txt
+
+
+Optimizer:
+  name: Adam
+  beta1: 0.9
+  beta2: 0.999
+  lr:
+    name: Piecewise
+    decay_epochs : [700, 800]
+    values : [0.001, 0.0001]
+    warmup_epoch: 5
+  regularizer:
+    name: L2
+    factor: 2.0e-05
+
+Architecture:
+  model_type: &model_type "rec"
+  name: DistillationModel
+  algorithm: Distillation
+  Models:
+    Teacher:
+      pretrained:
+      freeze_params: false
+      return_all_feats: true
+      model_type: *model_type
+      algorithm: CRNN
+      Transform:
+      Backbone:
+        name: MobileNetV1Enhance
+        scale: 0.5
+      Neck:
+        name: SequenceEncoder
+        encoder_type: rnn
+        hidden_size: 64
+      Head:
+        name: CTCHead
+        mid_channels: 96
+        fc_decay: 0.00002
+    Student:
+      pretrained:
+      freeze_params: false
+      return_all_feats: true
+      model_type: *model_type
+      algorithm: CRNN
+      Transform:
+      Backbone:
+        name: MobileNetV1Enhance
+        scale: 0.5
+      Neck:
+        name: SequenceEncoder
+        encoder_type: rnn
+        hidden_size: 64
+      Head:
+        name: CTCHead
+        mid_channels: 96
+        fc_decay: 0.00002
+  
+
+Loss:
+  name: CombinedLoss
+  loss_config_list:
+  - DistillationCTCLoss:
+      weight: 1.0
+      model_name_list: ["Student", "Teacher"]
+      key: head_out
+  - DistillationDMLLoss:
+      weight: 1.0
+      act: "softmax"
+      use_log: true
+      model_name_pairs:
+      - ["Student", "Teacher"]
+      key: head_out
+  - DistillationDistanceLoss:
+      weight: 1.0
+      mode: "l2"
+      model_name_pairs:
+      - ["Student", "Teacher"]
+      key: backbone_out
+
+PostProcess:
+  name: DistillationCTCLabelDecode
+  model_name: ["Student", "Teacher"]
+  key: head_out
+
+Metric:
+  name: DistillationMetric
+  base_metric_name: RecMetric
+  main_indicator: acc
+  key: "Student"
+
+Train:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data/
+    label_file_list:
+    - ./train_data/train_list.txt
+    transforms:
+    - DecodeImage:
+        img_mode: BGR
+        channel_first: false
+    - RecAug:
+    - CTCLabelEncode:
+    - RecResizeImg:
+        image_shape: [3, 32, 320]
+    - KeepKeys:
+        keep_keys:
+        - image
+        - label
+        - length
+  loader:
+    shuffle: true
+    batch_size_per_card: 128
+    drop_last: true
+    num_sections: 1
+    num_workers: 8
+Eval:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data
+    label_file_list:
+    - ./train_data/val_list.txt
+    transforms:
+    - DecodeImage:
+        img_mode: BGR
+        channel_first: false
+    - CTCLabelEncode:
+    - RecResizeImg:
+        image_shape: [3, 32, 320]
+    - KeepKeys:
+        keep_keys:
+        - image
+        - label
+        - length
+  loader:
+    shuffle: false
+    drop_last: false
+    batch_size_per_card: 128
+    num_workers: 8

doc/doc_ch/knowledge_distillation.md

@@ -39,7 +39,7 @@ PaddleOCR中集成了知识蒸馏的算法，具体地，有以下几个主要
 
 ### 2.1 识别配置文件解析
 
-配置文件在[ch_PP-OCRv2_rec.yml](../../configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec.yml)。
+配置文件在[ch_PP-OCRv2_rec_distillation.yml](../../configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec_distillation.yml)。
 
 #### 2.1.1 模型结构
 
@@ -246,6 +246,39 @@ Metric:
 关于`DistillationMetric`更加具体的实现可以参考: [distillation_metric.py](../../ppocr/metrics/distillation_metric.py#L24)。
 
 
+#### 2.1.5 蒸馏模型微调
+
+对蒸馏得到的识别蒸馏进行微调有2种方式。
+
+（1）基于知识蒸馏的微调：这种情况比较简单，下载预训练模型，在[ch_PP-OCRv2_rec_distillation.yml](../../configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec_distillation.yml)中配置好预训练模型路径以及自己的数据路径，即可进行模型微调训练。
+
+（2）微调时不使用知识蒸馏：这种情况，需要首先将预训练模型中的学生模型参数提取出来，具体步骤如下。
+
+* 首先下载预训练模型并解压。
+```shell
+# 下面预训练模型并解压
+wget https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_train.tar
+tar -xf ch_PP-OCRv2_rec_train.tar
+```
+
+* 然后使用python，对其中的学生模型参数进行提取
+
+```python
+import paddle
+# 加载预训练模型
+all_params = paddle.load("ch_PP-OCRv2_rec_train/best_accuracy.pdparams")
+# 查看权重参数的keys
+print(all_params.keys())
+# 学生模型的权重提取
+s_params = {key[len("Student."):]: all_params[key] for key in all_params if "Student." in key}
+# 查看学生模型权重参数的keys
+print(s_params.keys())
+# 保存
+paddle.save(s_params, "ch_PP-OCRv2_rec_train/student.pdparams")
+```
+
+转化完成之后，使用[ch_PP-OCRv2_rec.yml](../../configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec.yml)，修改预训练模型的路径（为导出的`student.pdparams`模型路径）以及自己的数据路径，即可进行模型微调。
+
 ### 2.2 检测配置文件解析
 
 * coming soon!

ppocr/losses/basic_loss.py

@@ -56,31 +56,34 @@ class CELoss(nn.Layer):
 
 class KLJSLoss(object):
     def __init__(self, mode='kl'):
-        assert mode in ['kl', 'js', 'KL', 'JS'], "mode can only be one of ['kl', 'js', 'KL', 'JS']"
+        assert mode in ['kl', 'js', 'KL', 'JS'
+                        ], "mode can only be one of ['kl', 'js', 'KL', 'JS']"
         self.mode = mode
 
     def __call__(self, p1, p2, reduction="mean"):
 
-        loss = paddle.multiply(p2, paddle.log( (p2+1e-5)/(p1+1e-5) + 1e-5))
+        loss = paddle.multiply(p2, paddle.log((p2 + 1e-5) / (p1 + 1e-5) + 1e-5))
 
         if self.mode.lower() == "js":
-            loss += paddle.multiply(p1, paddle.log((p1+1e-5)/(p2+1e-5) + 1e-5))
+            loss += paddle.multiply(
+                p1, paddle.log((p1 + 1e-5) / (p2 + 1e-5) + 1e-5))
             loss *= 0.5
         if reduction == "mean":
-            loss = paddle.mean(loss, axis=[1,2])
-        elif reduction=="none" or reduction is None:
+            loss = paddle.mean(loss, axis=[1, 2])
+        elif reduction == "none" or reduction is None:
             return loss
         else:
-            loss = paddle.sum(loss, axis=[1,2])
+            loss = paddle.sum(loss, axis=[1, 2])
 
         return loss
 
+
 class DMLLoss(nn.Layer):
     """
     DMLLoss
     """
 
-    def __init__(self, act=None):
+    def __init__(self, act=None, use_log=False):
         super().__init__()
         if act is not None:
             assert act in ["softmax", "sigmoid"]
@@ -91,19 +94,23 @@ class DMLLoss(nn.Layer):
         else:
             self.act = None
 
+        self.use_log = use_log
+
         self.jskl_loss = KLJSLoss(mode="js")
 
     def forward(self, out1, out2):
         if self.act is not None:
             out1 = self.act(out1)
             out2 = self.act(out2)
-        if len(out1.shape) < 2:
+        if self.use_log:
+            # for recognition distillation, log is needed for feature map
             log_out1 = paddle.log(out1)
             log_out2 = paddle.log(out2)
             loss = (F.kl_div(
                 log_out1, out2, reduction='batchmean') + F.kl_div(
                     log_out2, out1, reduction='batchmean')) / 2.0
         else:
+            # for detection distillation log is not needed
             loss = self.jskl_loss(out1, out2)
         return loss
 

ppocr/losses/combined_loss.py

@@ -49,11 +49,15 @@ class CombinedLoss(nn.Layer):
             loss = loss_func(input, batch, **kargs)
             if isinstance(loss, paddle.Tensor):
                 loss = {"loss_{}_{}".format(str(loss), idx): loss}
+
             weight = self.loss_weight[idx]
-            for key in loss.keys():
-                if key == "loss":
-                    loss_all += loss[key] * weight
+
+            loss = {key: loss[key] * weight for key in loss}
+
+            if "loss" in loss:
+                loss_all += loss["loss"]
             else:
-                    loss_dict["{}_{}".format(key, idx)] = loss[key]
+                loss_all += paddle.add_n(list(loss.values()))
+            loss_dict.update(loss)
         loss_dict["loss"] = loss_all
         return loss_dict

ppocr/losses/distillation_loss.py

@@ -44,10 +44,11 @@ class DistillationDMLLoss(DMLLoss):
     def __init__(self,
                  model_name_pairs=[],
                  act=None,
+                 use_log=False,
                  key=None,
                  maps_name=None,
                  name="dml"):
-        super().__init__(act=act)
+        super().__init__(act=act, use_log=use_log)
         assert isinstance(model_name_pairs, list)
         self.key = key
         self.model_name_pairs = self._check_model_name_pairs(model_name_pairs)
@@ -57,7 +58,8 @@ class DistillationDMLLoss(DMLLoss):
     def _check_model_name_pairs(self, model_name_pairs):
         if not isinstance(model_name_pairs, list):
             return []
-        elif isinstance(model_name_pairs[0], list) and isinstance(model_name_pairs[0][0], str):
+        elif isinstance(model_name_pairs[0], list) and isinstance(
+                model_name_pairs[0][0], str):
             return model_name_pairs
         else:
             return [model_name_pairs]
@@ -112,8 +114,8 @@ class DistillationDMLLoss(DMLLoss):
                             loss_dict["{}_{}_{}_{}_{}".format(key, pair[
                                 0], pair[1], map_name, idx)] = loss[key]
                     else:
-                        loss_dict["{}_{}_{}".format(self.name, self.maps_name[_c],
-                                                    idx)] = loss
+                        loss_dict["{}_{}_{}".format(self.name, self.maps_name[
+                            _c], idx)] = loss
 
         loss_dict = _sum_loss(loss_dict)
 

ppocr/utils/save_load.py

@@ -116,6 +116,7 @@ def load_dygraph_params(config, model, logger, optimizer):
         logger.info(f"loaded pretrained_model successful from {pm}")
         return {}
 
+
 def load_pretrained_params(model, path):
     if path is None:
         return False
@@ -138,6 +139,7 @@ def load_pretrained_params(model, path):
     print(f"load pretrain successful from {path}")
     return model
 
+
 def save_model(model,
                optimizer,
                model_path,