Merge branch 'dygraph' of https://github.com/PaddlePaddle/PaddleOCR into fix_doc

benchmark/run_benchmark_det.sh

+#!/usr/bin/env bash
+set -xe
+# 运行示例：CUDA_VISIBLE_DEVICES=0 bash run_benchmark.sh ${run_mode} ${bs_item} ${fp_item} 500 ${model_mode}
+# 参数说明
+function _set_params(){
+    run_mode=${1:-"sp"}          # 单卡sp|多卡mp
+    batch_size=${2:-"64"}
+    fp_item=${3:-"fp32"}        # fp32|fp16
+    max_iter=${4:-"500"}       # 可选，如果需要修改代码提前中断
+    model_name=${5:-"model_name"}
+    run_log_path=${TRAIN_LOG_DIR:-$(pwd)}  # TRAIN_LOG_DIR 后续QA设置该参数
+
+#   以下不用修改   
+    device=${CUDA_VISIBLE_DEVICES//,/ }
+    arr=(${device})
+    num_gpu_devices=${#arr[*]}
+    log_file=${run_log_path}/${model_name}_${run_mode}_bs${batch_size}_${fp_item}_${num_gpu_devices}
+}
+function _train(){
+    echo "Train on ${num_gpu_devices} GPUs"
+    echo "current CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES, gpus=$num_gpu_devices, batch_size=$batch_size"
+
+    train_cmd="-c configs/det/${model_name}.yml
+               -o Train.loader.batch_size_per_card=${batch_size}
+               -o Global.epoch_num=${max_iter} "   
+    case ${run_mode} in
+      sp) 
+        train_cmd="python3.7 tools/train.py "${train_cmd}""
+        ;;
+      mp)
+        train_cmd="python3.7 -m paddle.distributed.launch --log_dir=./mylog --gpus=$CUDA_VISIBLE_DEVICES tools/train.py ${train_cmd}"
+        ;;
+      *) echo "choose run_mode(sp or mp)"; exit 1;
+    esac
+# 以下不用修改
+    timeout 15m ${train_cmd} > ${log_file} 2>&1
+    if [ $? -ne 0 ];then
+            echo -e "${model_name}, FAIL"
+        export job_fail_flag=1
+    else
+        echo -e "${model_name}, SUCCESS"
+        export job_fail_flag=0
+    fi
+    kill -9 `ps -ef|grep 'python3.7'|awk '{print $2}'`
+
+    if [ $run_mode = "mp" -a -d mylog ]; then
+        rm ${log_file}
+        cp mylog/workerlog.0 ${log_file}
+    fi
+}
+
+_set_params $@
+_train
+

benchmark/run_det.sh

+# 提供可稳定复现性能的脚本，默认在标准docker环境内py37执行： paddlepaddle/paddle:latest-gpu-cuda10.1-cudnn7  paddle=2.1.2  py=37
+# 执行目录：需说明
+cd PaddleOCR
+# 1 安装该模型需要的依赖 (如需开启优化策略请注明)
+python3.7 -m pip install -r requirements.txt
+# 2 拷贝该模型需要数据、预训练模型
+wget -p ./tain_data/  https://paddleocr.bj.bcebos.com/dygraph_v2.0/test/icdar2015.tar && cd train_data  && tar xf icdar2015.tar && cd ../
+wget -p ./pretrain_models/ https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ResNet50_vd_pretrained.pdparams
+# 3 批量运行（如不方便批量，1，2需放到单个模型中）
+
+model_mode_list=(det_mv3_db det_r50_vd_east)
+fp_item_list=(fp32)
+bs_list=(256 128)
+for model_mode in ${model_mode_list[@]}; do
+      for fp_item in ${fp_item_list[@]}; do
+          for bs_item in ${bs_list[@]}; do
+            echo "index is speed, 1gpus, begin, ${model_name}"
+            run_mode=sp
+            CUDA_VISIBLE_DEVICES=0 bash benchmark/run_benchmark.sh ${run_mode} ${bs_item} ${fp_item} 10 ${model_mode}     #  (5min)
+            sleep 60
+            echo "index is speed, 8gpus, run_mode is multi_process, begin, ${model_name}"
+            run_mode=mp
+            CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash benchmark/run_benchmark.sh ${run_mode} ${bs_item} ${fp_item} 10 ${model_mode} 
+            sleep 60
+            done
+      done
+done
+
+

configs/det/det_r50_vd_east.yml

@@ -8,7 +8,7 @@ Global:
   # evaluation is run every 5000 iterations after the 4000th iteration
   eval_batch_step: [4000, 5000]
   cal_metric_during_train: False
-  pretrained_model: ./pretrain_models/ResNet50_vd_pretrained/
+  pretrained_model: ./pretrain_models/ResNet50_vd_pretrained
   checkpoints: 
   save_inference_dir:
   use_visualdl: False

configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec_enhanced_ctc_loss.yml

+Global:
+  debug: false
+  use_gpu: true
+  epoch_num: 800
+  log_smooth_window: 20
+  print_batch_step: 10
+  save_model_dir: ./output/rec_mobile_pp-OCRv2_enhanced_ctc_loss
+  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_mobile_pp-OCRv2_enhanced_ctc_loss.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: rec
+  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
+    return_feats: true
+
+Loss:
+  name: CombinedLoss
+  loss_config_list:
+  - CTCLoss:
+      use_focal_loss: false
+      weight: 1.0
+  - CenterLoss:
+      weight: 0.05
+      num_classes: 6625
+      feat_dim: 96
+      init_center: false
+      center_file_path: "./train_center.pkl"
+  # you can also try to add ace loss on your own dataset
+  # - ACELoss:
+  #     weight: 0.1
+
+PostProcess:
+  name: CTCLabelDecode
+
+Metric:
+  name: RecMetric
+  main_indicator: acc
+
+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
+        - label_ace
+  loader:
+    shuffle: true
+    batch_size_per_card: 128
+    drop_last: true
+    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

configs/rec/rec_resnet_stn_bilstm_att.yml

+Global:
+  use_gpu: True
+  epoch_num: 400
+  log_smooth_window: 20
+  print_batch_step: 10
+  save_model_dir: ./output/rec/seed
+  save_epoch_step: 3
+  # evaluation is run every 5000 iterations after the 4000th iteration
+  eval_batch_step: [0, 2000]
+  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_symbol
+  max_text_length: 100
+  infer_mode: False
+  use_space_char: False
+  save_res_path: ./output/rec/predicts_seed.txt
+
+
+Optimizer:
+  name: Adadelta
+  weight_deacy: 0.0
+  momentum: 0.9
+  lr:
+    name: Piecewise
+    decay_epochs: [4,5,8]
+    values: [1.0, 0.1, 0.01]
+  regularizer:
+    name: 'L2'
+    factor: 2.0e-05
+
+
+Architecture:
+  model_type: rec
+  algorithm: SEED
+  Transform:
+    name: STN_ON
+    tps_inputsize: [32, 64]
+    tps_outputsize: [32, 100]
+    num_control_points: 20
+    tps_margins: [0.05,0.05]
+    stn_activation: none
+  Backbone:
+    name: ResNet_ASTER
+  Head:
+    name: AsterHead  # AttentionHead
+    sDim: 512
+    attDim: 512
+    max_len_labels: 100
+
+Loss:
+  name: AsterLoss
+
+PostProcess:
+  name: SEEDLabelDecode
+
+Metric:
+  name: RecMetric
+  main_indicator: acc
+  is_filter: True
+
+Train:
+  dataset:
+    name: LMDBDataSet
+    data_dir: ./train_data/data_lmdb_release/training/
+    transforms:
+      - Fasttext:
+          path: "./cc.en.300.bin"
+      - DecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - SEEDLabelEncode: # Class handling label
+      - RecResizeImg:
+          character_type: en
+          image_shape: [3, 64, 256]
+          padding: False
+      - KeepKeys:
+          keep_keys: ['image', 'label', 'length', 'fast_label'] # dataloader will return list in this order
+  loader:
+    shuffle: True
+    batch_size_per_card: 256
+    drop_last: True
+    num_workers: 6
+
+Eval:
+  dataset:
+    name: LMDBDataSet
+    data_dir: ./train_data/data_lmdb_release/evaluation/
+    transforms:
+      - DecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - SEEDLabelEncode: # Class handling label
+      - RecResizeImg:
+          character_type: en
+          image_shape: [3, 64, 256]
+          padding: False
+      - KeepKeys:
+          keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order
+  loader:
+    shuffle: False
+    drop_last: True
+    batch_size_per_card: 256
+    num_workers: 4

deploy/slim/prune/sensitivity_anal.py

@@ -110,25 +110,42 @@ def main(config, device, logger, vdl_writer):
         logger.info("metric['hmean']: {}".format(metric['hmean']))
         return metric['hmean']
 
-    params_sensitive = pruner.sensitive(
-        eval_func=eval_fn,
-        sen_file="./sen.pickle",
-        skip_vars=[
-            "conv2d_57.w_0", "conv2d_transpose_2.w_0", "conv2d_transpose_3.w_0"
-        ])
-
-    logger.info(
-        "The sensitivity analysis results of model parameters saved in sen.pickle"
-    )
-    # calculate pruned params's ratio
-    params_sensitive = pruner._get_ratios_by_loss(params_sensitive, loss=0.02)
-    for key in params_sensitive.keys():
-        logger.info("{}, {}".format(key, params_sensitive[key]))
-
-    #params_sensitive = {}
-    #for param in model.parameters():
-    #    if 'transpose' not in param.name and 'linear' not in param.name:
-    #        params_sensitive[param.name] = 0.1  
+    run_sensitive_analysis = False
+    """
+    run_sensitive_analysis=True: 
+        Automatically compute the sensitivities of convolutions in a model. 
+        The sensitivity of a convolution is the losses of accuracy on test dataset in 
+        differenct pruned ratios. The sensitivities can be used to get a group of best 
+        ratios with some condition.
+    
+    run_sensitive_analysis=False: 
+        Set prune trim ratio to a fixed value, such as 10%. The larger the value, 
+        the more convolution weights will be cropped.
+
+    """
+
+    if run_sensitive_analysis:
+        params_sensitive = pruner.sensitive(
+            eval_func=eval_fn,
+            sen_file="./deploy/slim/prune/sen.pickle",
+            skip_vars=[
+                "conv2d_57.w_0", "conv2d_transpose_2.w_0",
+                "conv2d_transpose_3.w_0"
+            ])
+        logger.info(
+            "The sensitivity analysis results of model parameters saved in sen.pickle"
+        )
+        # calculate pruned params's ratio
+        params_sensitive = pruner._get_ratios_by_loss(
+            params_sensitive, loss=0.02)
+        for key in params_sensitive.keys():
+            logger.info("{}, {}".format(key, params_sensitive[key]))
+    else:
+        params_sensitive = {}
+        for param in model.parameters():
+            if 'transpose' not in param.name and 'linear' not in param.name:
+                # set prune ratio as 10%. The larger the value, the more convolution weights will be cropped
+                params_sensitive[param.name] = 0.1
 
     plan = pruner.prune_vars(params_sensitive, [0])
 

doc/doc_ch/algorithm_overview.md

@@ -50,6 +50,7 @@ PaddleOCR基于动态图开源的文本识别算法列表：
 - [x]  SRN([paper](https://arxiv.org/abs/2003.12294))
 - [x]  NRTR([paper](https://arxiv.org/abs/1806.00926v2))
 - [x]  SAR([paper](https://arxiv.org/abs/1811.00751v2))
+- [x]  SEED([paper](https://arxiv.org/pdf/2005.10977.pdf))
 
 参考[DTRB](https://arxiv.org/abs/1904.01906) 文字识别训练和评估流程，使用MJSynth和SynthText两个文字识别数据集训练，在IIIT, SVT, IC03, IC13, IC15, SVTP, CUTE数据集上进行评估，算法效果如下：
 
@@ -66,5 +67,5 @@ PaddleOCR基于动态图开源的文本识别算法列表：
 |SRN|Resnet50_vd_fpn| 88.52% | rec_r50fpn_vd_none_srn | [下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_r50_vd_srn_train.tar) |
 |NRTR|NRTR_MTB| 84.3% | rec_mtb_nrtr | [下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_mtb_nrtr_train.tar) |
 |SAR|Resnet31| 87.2% | rec_r31_sar | [下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.1/rec/rec_r31_sar_train.tar) |
-
+|SEED| Aster_Resnet | 85.2% | rec_resnet_stn_bilstm_att | [下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.1/rec/rec_resnet_stn_bilstm_att.tar)|
 PaddleOCR文本识别算法的训练和使用请参考文档教程中[模型训练/评估中的文本识别部分](./recognition.md)。

doc/doc_ch/recognition.md

@@ -234,6 +234,9 @@ PaddleOCR支持训练和评估交替进行, 可以在 `configs/rec/rec_icdar15_t
 | rec_r50fpn_vd_none_srn.yml    | SRN | Resnet50_fpn_vd    | None    | rnn | srn |
 | rec_mtb_nrtr.yml    | NRTR | nrtr_mtb    | None    | transformer encoder | transformer decoder |
 | rec_r31_sar.yml               | SAR | ResNet31 | None | LSTM encoder | LSTM decoder |
+| rec_resnet_stn_bilstm_att.yml | SEED | Aster_Resnet | STN | BiLSTM | att |
+
+*其中SEED模型需要额外加载FastText训练好的[语言模型](https://dl.fbaipublicfiles.com/fasttext/vectors-crawl/cc.en.300.bin.gz)
 
 训练中文数据，推荐使用[rec_chinese_lite_train_v2.0.yml](../../configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml)，如您希望尝试其他算法在中文数据集上的效果，请参考下列说明修改配置文件：
 
@@ -460,5 +463,3 @@ python3 tools/export_model.py -c configs/rec/ch_ppocr_v2.0/rec_chinese_lite_trai
   ```
   python3 tools/infer/predict_rec.py --image_dir="./doc/imgs_words_en/word_336.png" --rec_model_dir="./your inference model" --rec_image_shape="3, 32, 100" --rec_char_type="ch" --rec_char_dict_path="your text dict path"
   ```
-
-  

ppocr/data/imaug/label_ops.py

@@ -215,6 +215,11 @@ class CTCLabelEncode(BaseRecLabelEncode):
         data['length'] = np.array(len(text))
         text = text + [0] * (self.max_text_len - len(text))
         data['label'] = np.array(text)
+
+        label = [0] * len(self.character)
+        for x in text:
+            label[x] += 1
+        data['label_ace'] = np.array(label)
         return data
 
     def add_special_char(self, dict_character):
@@ -342,6 +347,38 @@ class AttnLabelEncode(BaseRecLabelEncode):
         return idx
 
 
+class SEEDLabelEncode(BaseRecLabelEncode):
+    """ Convert between text-label and text-index """
+
+    def __init__(self,
+                 max_text_length,
+                 character_dict_path=None,
+                 character_type='ch',
+                 use_space_char=False,
+                 **kwargs):
+        super(SEEDLabelEncode,
+              self).__init__(max_text_length, character_dict_path,
+                             character_type, use_space_char)
+
+    def add_special_char(self, dict_character):
+        self.end_str = "eos"
+        dict_character = dict_character + [self.end_str]
+        return dict_character
+
+    def __call__(self, data):
+        text = data['label']
+        text = self.encode(text)
+        if text is None:
+            return None
+        if len(text) >= self.max_text_len:
+            return None
+        data['length'] = np.array(len(text)) + 1  # conclude eos
+        text = text + [len(self.character) - 1] * (self.max_text_len - len(text)
+                                                   )
+        data['label'] = np.array(text)
+        return data
+
+
 class SRNLabelEncode(BaseRecLabelEncode):
     """ Convert between text-label and text-index """
 
@@ -421,7 +458,6 @@ class TableLabelEncode(object):
             substr = lines[0].decode('utf-8').strip("\r\n").split("\t")
             character_num = int(substr[0])
             elem_num = int(substr[1])
-
             for cno in range(1, 1 + character_num):
                 character = lines[cno].decode('utf-8').strip("\r\n")
                 list_character.append(character)

ppocr/data/imaug/operators.py

@@ -23,6 +23,7 @@ import sys
 import six
 import cv2
 import numpy as np
+import fasttext
 
 
 class DecodeImage(object):
@@ -83,12 +84,13 @@ class NRTRDecodeImage(object):
         elif self.img_mode == 'RGB':
             assert img.shape[2] == 3, 'invalid shape of image[%s]' % (img.shape)
             img = img[:, :, ::-1]
-        img = cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)
+        img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
         if self.channel_first:
             img = img.transpose((2, 0, 1))
         data['image'] = img
         return data
 
+
 class NormalizeImage(object):
     """ normalize image such as substract mean, divide std
     """
@@ -133,6 +135,17 @@ class ToCHWImage(object):
         return data
 
 
+class Fasttext(object):
+    def __init__(self, path="None", **kwargs):
+        self.fast_model = fasttext.load_model(path)
+
+    def __call__(self, data):
+        label = data['label']
+        fast_label = self.fast_model[label]
+        data['fast_label'] = fast_label
+        return data
+
+
 class KeepKeys(object):
     def __init__(self, keep_keys, **kwargs):
         self.keep_keys = keep_keys

ppocr/data/imaug/rec_img_aug.py

@@ -88,17 +88,19 @@ class RecResizeImg(object):
                  image_shape,
                  infer_mode=False,
                  character_type='ch',
+                 padding=True,
                  **kwargs):
         self.image_shape = image_shape
         self.infer_mode = infer_mode
         self.character_type = character_type
+        self.padding = padding
 
     def __call__(self, data):
         img = data['image']
         if self.infer_mode and self.character_type == "ch":
             norm_img = resize_norm_img_chinese(img, self.image_shape)
         else:
-            norm_img = resize_norm_img(img, self.image_shape)
+            norm_img = resize_norm_img(img, self.image_shape, self.padding)
         data['image'] = norm_img
         return data
 
@@ -174,16 +176,21 @@ def resize_norm_img_sar(img, image_shape, width_downsample_ratio=0.25):
     return padding_im, resize_shape, pad_shape, valid_ratio
 
 
-def resize_norm_img(img, image_shape):
+def resize_norm_img(img, image_shape, padding=True):
     imgC, imgH, imgW = image_shape
     h = img.shape[0]
     w = img.shape[1]
-    ratio = w / float(h)
-    if math.ceil(imgH * ratio) > imgW:
+    if not padding:
+        resized_image = cv2.resize(
+            img, (imgW, imgH), interpolation=cv2.INTER_LINEAR)
         resized_w = imgW
     else:
-        resized_w = int(math.ceil(imgH * ratio))
-    resized_image = cv2.resize(img, (resized_w, imgH))
+        ratio = w / float(h)
+        if math.ceil(imgH * ratio) > imgW:
+            resized_w = imgW
+        else:
+            resized_w = int(math.ceil(imgH * ratio))
+        resized_image = cv2.resize(img, (resized_w, imgH))
     resized_image = resized_image.astype('float32')
     if image_shape[0] == 1:
         resized_image = resized_image / 255

ppocr/losses/__init__.py

@@ -28,6 +28,8 @@ from .rec_att_loss import AttentionLoss
 from .rec_srn_loss import SRNLoss
 from .rec_nrtr_loss import NRTRLoss
 from .rec_sar_loss import SARLoss
+from .rec_aster_loss import AsterLoss
+
 # cls loss
 from .cls_loss import ClsLoss
 
@@ -48,9 +50,8 @@ def build_loss(config):
     support_dict = [
         'DBLoss', 'PSELoss', 'EASTLoss', 'SASTLoss', 'CTCLoss', 'ClsLoss',
         'AttentionLoss', 'SRNLoss', 'PGLoss', 'CombinedLoss', 'NRTRLoss',
-        'TableAttentionLoss', 'SARLoss'
+        'TableAttentionLoss', 'SARLoss', 'AsterLoss'
     ]
-
     config = copy.deepcopy(config)
     module_name = config.pop('name')
     assert module_name in support_dict, Exception('loss only support {}'.format(

ppocr/losses/ace_loss.py

+# copyright (c) 2021 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
+import paddle.nn as nn
+
+
+class ACELoss(nn.Layer):
+    def __init__(self, **kwargs):
+        super().__init__()
+        self.loss_func = nn.CrossEntropyLoss(
+            weight=None,
+            ignore_index=0,
+            reduction='none',
+            soft_label=True,
+            axis=-1)
+
+    def __call__(self, predicts, batch):
+        if isinstance(predicts, (list, tuple)):
+            predicts = predicts[-1]
+        B, N = predicts.shape[:2]
+        div = paddle.to_tensor([N]).astype('float32')
+
+        predicts = nn.functional.softmax(predicts, axis=-1)
+        aggregation_preds = paddle.sum(predicts, axis=1)
+        aggregation_preds = paddle.divide(aggregation_preds, div)
+
+        length = batch[2].astype("float32")
+        batch = batch[3].astype("float32")
+        batch[:, 0] = paddle.subtract(div, length)
+
+        batch = paddle.divide(batch, div)
+
+        loss = self.loss_func(aggregation_preds, batch)
+
+        return {"loss_ace": loss}

ppocr/losses/center_loss.py

+#copyright (c) 2021 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 os
+import pickle
+
+import paddle
+import paddle.nn as nn
+import paddle.nn.functional as F
+
+
+class CenterLoss(nn.Layer):
+    """
+    Reference: Wen et al. A Discriminative Feature Learning Approach for Deep Face Recognition. ECCV 2016.
+    """
+
+    def __init__(self,
+                 num_classes=6625,
+                 feat_dim=96,
+                 init_center=False,
+                 center_file_path=None):
+        super().__init__()
+        self.num_classes = num_classes
+        self.feat_dim = feat_dim
+        self.centers = paddle.randn(
+            shape=[self.num_classes, self.feat_dim]).astype(
+                "float64")  #random center
+
+        if init_center:
+            assert os.path.exists(
+                center_file_path
+            ), f"center path({center_file_path}) must exist when init_center is set as True."
+            with open(center_file_path, 'rb') as f:
+                char_dict = pickle.load(f)
+                for key in char_dict.keys():
+                    self.centers[key] = paddle.to_tensor(char_dict[key])
+
+    def __call__(self, predicts, batch):
+        assert isinstance(predicts, (list, tuple))
+        features, predicts = predicts
+
+        feats_reshape = paddle.reshape(
+            features, [-1, features.shape[-1]]).astype("float64")
+        label = paddle.argmax(predicts, axis=2)
+        label = paddle.reshape(label, [label.shape[0] * label.shape[1]])
+
+        batch_size = feats_reshape.shape[0]
+
+        #calc feat * feat   
+        dist1 = paddle.sum(paddle.square(feats_reshape), axis=1, keepdim=True)
+        dist1 = paddle.expand(dist1, [batch_size, self.num_classes])
+
+        #dist2 of centers
+        dist2 = paddle.sum(paddle.square(self.centers), axis=1,
+                           keepdim=True)  #num_classes
+        dist2 = paddle.expand(dist2,
+                              [self.num_classes, batch_size]).astype("float64")
+        dist2 = paddle.transpose(dist2, [1, 0])
+
+        #first x * x + y * y
+        distmat = paddle.add(dist1, dist2)
+        tmp = paddle.matmul(feats_reshape,
+                            paddle.transpose(self.centers, [1, 0]))
+        distmat = distmat - 2.0 * tmp
+
+        #generate the mask
+        classes = paddle.arange(self.num_classes).astype("int64")
+        label = paddle.expand(
+            paddle.unsqueeze(label, 1), (batch_size, self.num_classes))
+        mask = paddle.equal(
+            paddle.expand(classes, [batch_size, self.num_classes]),
+            label).astype("float64")  #get mask
+        dist = paddle.multiply(distmat, mask)
+        loss = paddle.sum(paddle.clip(dist, min=1e-12, max=1e+12)) / batch_size
+        return {'loss_center': loss}

ppocr/losses/combined_loss.py

@@ -15,6 +15,10 @@
 import paddle
 import paddle.nn as nn
 
+from .rec_ctc_loss import CTCLoss
+from .center_loss import CenterLoss
+from .ace_loss import ACELoss
+
 from .distillation_loss import DistillationCTCLoss
 from .distillation_loss import DistillationDMLLoss
 from .distillation_loss import DistillationDistanceLoss, DistillationDBLoss, DistillationDilaDBLoss

ppocr/losses/distillation_loss.py

@@ -112,7 +112,7 @@ class DistillationDMLLoss(DMLLoss):
                     if isinstance(loss, dict):
                         for key in loss:
                             loss_dict["{}_{}_{}_{}_{}".format(key, pair[
-                                0], pair[1], map_name, idx)] = loss[key]
+                                0], pair[1], self.maps_name, idx)] = loss[key]
                     else:
                         loss_dict["{}_{}_{}".format(self.name, self.maps_name[
                             _c], idx)] = loss

ppocr/losses/rec_aster_loss.py

+# copyright (c) 2021 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
+
+
+class CosineEmbeddingLoss(nn.Layer):
+    def __init__(self, margin=0.):
+        super(CosineEmbeddingLoss, self).__init__()
+        self.margin = margin
+        self.epsilon = 1e-12
+
+    def forward(self, x1, x2, target):
+        similarity = paddle.fluid.layers.reduce_sum(
+            x1 * x2, dim=-1) / (paddle.norm(
+                x1, axis=-1) * paddle.norm(
+                    x2, axis=-1) + self.epsilon)
+        one_list = paddle.full_like(target, fill_value=1)
+        out = paddle.fluid.layers.reduce_mean(
+            paddle.where(
+                paddle.equal(target, one_list), 1. - similarity,
+                paddle.maximum(
+                    paddle.zeros_like(similarity), similarity - self.margin)))
+
+        return out
+
+
+class AsterLoss(nn.Layer):
+    def __init__(self,
+                 weight=None,
+                 size_average=True,
+                 ignore_index=-100,
+                 sequence_normalize=False,
+                 sample_normalize=True,
+                 **kwargs):
+        super(AsterLoss, self).__init__()
+        self.weight = weight
+        self.size_average = size_average
+        self.ignore_index = ignore_index
+        self.sequence_normalize = sequence_normalize
+        self.sample_normalize = sample_normalize
+        self.loss_sem = CosineEmbeddingLoss()
+        self.is_cosin_loss = True
+        self.loss_func_rec = nn.CrossEntropyLoss(weight=None, reduction='none')
+
+    def forward(self, predicts, batch):
+        targets = batch[1].astype("int64")
+        label_lengths = batch[2].astype('int64')
+        sem_target = batch[3].astype('float32')
+        embedding_vectors = predicts['embedding_vectors']
+        rec_pred = predicts['rec_pred']
+
+        if not self.is_cosin_loss:
+            sem_loss = paddle.sum(self.loss_sem(embedding_vectors, sem_target))
+        else:
+            label_target = paddle.ones([embedding_vectors.shape[0]])
+            sem_loss = paddle.sum(
+                self.loss_sem(embedding_vectors, sem_target, label_target))
+
+        # rec loss
+        batch_size, def_max_length = targets.shape[0], targets.shape[1]
+
+        mask = paddle.zeros([batch_size, def_max_length])
+        for i in range(batch_size):
+            mask[i, :label_lengths[i]] = 1
+        mask = paddle.cast(mask, "float32")
+        max_length = max(label_lengths)
+        assert max_length == rec_pred.shape[1]
+        targets = targets[:, :max_length]
+        mask = mask[:, :max_length]
+        rec_pred = paddle.reshape(rec_pred, [-1, rec_pred.shape[2]])
+        input = nn.functional.log_softmax(rec_pred, axis=1)
+        targets = paddle.reshape(targets, [-1, 1])
+        mask = paddle.reshape(mask, [-1, 1])
+        output = -paddle.index_sample(input, index=targets) * mask
+        output = paddle.sum(output)
+        if self.sequence_normalize:
+            output = output / paddle.sum(mask)
+        if self.sample_normalize:
+            output = output / batch_size
+
+        loss = output + sem_loss * 0.1
+        return {'loss': loss}

ppocr/losses/rec_ctc_loss.py

@@ -21,16 +21,24 @@ from paddle import nn
 
 
 class CTCLoss(nn.Layer):
-    def __init__(self, **kwargs):
+    def __init__(self, use_focal_loss=False, **kwargs):
         super(CTCLoss, self).__init__()
         self.loss_func = nn.CTCLoss(blank=0, reduction='none')
+        self.use_focal_loss = use_focal_loss
 
     def forward(self, predicts, batch):
+        if isinstance(predicts, (list, tuple)):
+            predicts = predicts[-1]
         predicts = predicts.transpose((1, 0, 2))
         N, B, _ = predicts.shape
         preds_lengths = paddle.to_tensor([N] * B, dtype='int64')
         labels = batch[1].astype("int32")
         label_lengths = batch[2].astype('int64')
         loss = self.loss_func(predicts, labels, preds_lengths, label_lengths)
+        if self.use_focal_loss:
+            weight = paddle.exp(-loss)
+            weight = paddle.subtract(paddle.to_tensor([1.0]), weight)
+            weight = paddle.square(weight) * self.focal_loss_alpha
+            loss = paddle.multiply(loss, weight)
         loss = loss.mean()  # sum
         return {'loss': loss}

ppocr/metrics/rec_metric.py

@@ -13,13 +13,20 @@
 # limitations under the License.
 
 import Levenshtein
+import string
 
 
 class RecMetric(object):
-    def __init__(self, main_indicator='acc', **kwargs):
+    def __init__(self, main_indicator='acc', is_filter=False, **kwargs):
         self.main_indicator = main_indicator
+        self.is_filter = is_filter
         self.reset()
 
+    def _normalize_text(self, text):
+        text = ''.join(
+            filter(lambda x: x in (string.digits + string.ascii_letters), text))
+        return text.lower()
+
     def __call__(self, pred_label, *args, **kwargs):
         preds, labels = pred_label
         correct_num = 0
@@ -28,6 +35,9 @@ class RecMetric(object):
         for (pred, pred_conf), (target, _) in zip(preds, labels):
             pred = pred.replace(" ", "")
             target = target.replace(" ", "")
+            if self.is_filter:
+                pred = self._normalize_text(pred)
+                target = self._normalize_text(target)
             norm_edit_dis += Levenshtein.distance(pred, target) / max(
                 len(pred), len(target), 1)
             if pred == target:
@@ -57,4 +67,3 @@ class RecMetric(object):
         self.correct_num = 0
         self.all_num = 0
         self.norm_edit_dis = 0
-