Merge pull request #952 from WenmuZhou/dygraph

使用PaddleClass的resnet_vd

configs/det/det_mv3_db.yml

@@ -3,7 +3,7 @@ Global:
   epoch_num: 1200
   log_smooth_window: 20
   print_batch_step: 2
-  save_model_dir: ./output/20201010/
+  save_model_dir: ./output/db_mv3/
   save_epoch_step: 1200
   # evaluation is run every 5000 iterations after the 4000th iteration
   eval_batch_step: 8
@@ -66,9 +66,9 @@ Metric:
 TRAIN:
   dataset:
     name: SimpleDataSet
-    data_dir: /home/zhoujun20/detection/
+    data_dir: ./detection/
     file_list:
-      - /home/zhoujun20/detection/train_icdar2015_label.txt # dataset1
+      - ./detection/train_icdar2015_label.txt # dataset1
     ratio_list: [1.0]
     transforms:
       - DecodeImage: # load image
@@ -103,14 +103,14 @@ TRAIN:
     shuffle: True
     drop_last: False
     batch_size: 16
-    num_workers: 6
+    num_workers: 8
 
 EVAL:
   dataset:
     name: SimpleDataSet
-    data_dir: /home/zhoujun20/detection/
+    data_dir: ./detection/
     file_list:
-      - /home/zhoujun20/detection/test_icdar2015_label.txt
+      - ./detection/test_icdar2015_label.txt
     transforms:
       - DecodeImage: # load image
           img_mode: BGR
@@ -130,4 +130,4 @@ EVAL:
     shuffle: False
     drop_last: False
     batch_size: 1 # must be 1
-    num_workers: 6
+    num_workers: 8

configs/det/det_r50_vd_db.yml

@@ -3,14 +3,14 @@ Global:
   epoch_num: 1200
   log_smooth_window: 20
   print_batch_step: 2
-  save_model_dir: ./output/20201010/
+  save_model_dir: ./output/20201015_r50/
   save_epoch_step: 1200
   # evaluation is run every 5000 iterations after the 4000th iteration
   eval_batch_step: 8
   # if pretrained_model is saved in static mode, load_static_weights must set to True
   load_static_weights: True
   cal_metric_during_train: False
-  pretrained_model: /home/zhoujun20/pretrain_models/MobileNetV3_large_x0_5_pretrained
+  pretrained_model: /home/zhoujun20/pretrain_models/ResNet50_vd_ssld_pretrained/
   checkpoints: #./output/det_db_0.001_DiceLoss_256_pp_config_2.0b_4gpu/best_accuracy
   save_inference_dir:
   use_visualdl: True
@@ -102,7 +102,7 @@ TRAIN:
     shuffle: True
     drop_last: False
     batch_size: 16
-    num_workers: 6
+    num_workers: 8
 
 EVAL:
   dataset:
@@ -129,4 +129,4 @@ EVAL:
     shuffle: False
     drop_last: False
     batch_size: 1 # must be 1
-    num_workers: 6
+    num_workers: 8

configs/rec/rec_mv3_none_bilstm_ctc.yml

@@ -84,7 +84,7 @@ TRAIN:
     batch_size: 256
     shuffle: True
     drop_last: True
-    num_workers: 6
+    num_workers: 8
 
 EVAL:
   dataset:
@@ -105,4 +105,4 @@ EVAL:
     shuffle: False
     drop_last: False
     batch_size: 256
-    num_workers: 6
+    num_workers: 8

configs/rec/rec_mv3_none_bilstm_ctc_lmdb.yml

@@ -83,7 +83,7 @@ TRAIN:
     batch_size: 256
     shuffle: True
     drop_last: True
-    num_workers: 6
+    num_workers: 8
 
 EVAL:
   dataset:
@@ -103,4 +103,4 @@ EVAL:
     shuffle: False
     drop_last: False
     batch_size: 256
-    num_workers: 6
+    num_workers: 8

configs/rec/rec_mv3_none_none_ctc_lmdb.yml

@@ -0,0 +1,105 @@
+Global:
+  use_gpu: false
+  epoch_num: 500
+  log_smooth_window: 20
+  print_batch_step: 1
+  save_model_dir: ./output/rec/test/
+  save_epoch_step: 500
+  # evaluation is run every 5000 iterations after the 4000th iteration
+  eval_batch_step: 1016
+  # if pretrained_model is saved in static mode, load_static_weights must set to True
+  load_static_weights: True
+  cal_metric_during_train: True
+  pretrained_model:
+  checkpoints: #output/rec/rec_crnn/best_accuracy
+  save_inference_dir:
+  use_visualdl: True
+  infer_img: doc/imgs_words/ch/word_1.jpg
+  # for data or label process
+  max_text_length: 80
+  character_dict_path: /home/zhoujun20/rec/lmdb/dict.txt
+  character_type: 'en'
+  use_space_char: True
+  infer_mode: False
+  use_tps: False
+
+
+Optimizer:
+  name: Adam
+  beta1: 0.9
+  beta2: 0.999
+  learning_rate:
+    name: Cosine
+    lr: 0.0005
+    warmup_epoch: 1
+  regularizer:
+    name: 'L2'
+    factor: 0.00001
+
+Architecture:
+  type: rec
+  algorithm: CRNN
+  Transform:
+  Backbone:
+    name: MobileNetV3
+    scale: 0.5
+    model_name: small
+    small_stride: [ 1, 2, 2, 2 ]
+  Neck:
+    name: SequenceEncoder
+    encoder_type: reshape
+  Head:
+    name: CTC
+    fc_decay: 0.00001
+
+Loss:
+  name: CTCLoss
+
+PostProcess:
+  name: CTCLabelDecode
+
+Metric:
+  name: RecMetric
+  main_indicator: acc
+
+TRAIN:
+  dataset:
+    name: LMDBDateSet
+    file_list:
+      - /Users/zhoujun20/Downloads/evaluation_new # dataset1
+    ratio_list: [ 0.4,0.6 ]
+    transforms:
+      - DecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - CTCLabelEncode: # Class handling label
+      - RecAug:
+      - RecResizeImg:
+          image_shape: [ 3,32,320 ]
+      - keepKeys:
+          keep_keys: [ 'image','label','length' ] # dataloader将按照此顺序返回list
+  loader:
+    batch_size: 256
+    shuffle: True
+    drop_last: True
+    num_workers: 8
+
+EVAL:
+  dataset:
+    name: LMDBDateSet
+    file_list:
+      - /home/zhoujun20/rec/lmdb/val
+    transforms:
+      - DecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - CTCLabelEncode: # Class handling label
+      - RecResizeImg:
+          image_shape: [ 3,32,320 ]
+      - keepKeys:
+          keep_keys: [ 'image','label','length' ] # dataloader将按照此顺序返回list
+  loader:
+    shuffle: False
+    drop_last: False
+    batch_size: 256
+    num_workers: 8

configs/rec/rec_r34_vd_none_bilstm_ctc.yml

@@ -42,7 +42,7 @@ Architecture:
   Transform:
   Backbone:
     name: ResNet
-    layers: 200
+    layers: 34
   Neck:
     name: SequenceEncoder
     encoder_type: fc
@@ -82,7 +82,7 @@ TRAIN:
     batch_size: 256
     shuffle: True
     drop_last: True
-    num_workers: 6
+    num_workers: 8
 
 EVAL:
   dataset:
@@ -103,4 +103,4 @@ EVAL:
     shuffle: False
     drop_last: False
     batch_size: 256
-    num_workers: 6
+    num_workers: 8

ppocr/modeling/architectures/model.py

@@ -94,13 +94,11 @@ def check_static():
     from ppocr.utils.logging import get_logger
     from tools import program
 
-    config = program.load_config('configs/det/det_r50_vd_db.yml')
+    config = program.load_config('configs/rec/rec_r34_vd_none_bilstm_ctc.yml')
 
-    # import cv2
-    # data = cv2.imread('doc/imgs/1.jpg')
-    # data = normalize(data)
     logger = get_logger()
-    data = np.zeros((1, 3, 640, 640), dtype=np.float32)
+    np.random.seed(0)
+    data = np.random.rand(1, 3, 32, 320).astype(np.float32)
     paddle.disable_static()
 
     config['Architecture']['in_channels'] = 3
@@ -110,17 +108,15 @@ def check_static():
     load_dygraph_pretrain(
         model,
         logger,
-        '/Users/zhoujun20/Desktop/code/PaddleOCR/db/db',
+        '/Users/zhoujun20/Desktop/code/PaddleOCR/cnn_ctc/cnn_ctc',
         load_static_weights=True)
-    x = paddle.to_variable(data)
+    x = paddle.to_tensor(data)
     y = model(x)
     for y1 in y:
         print(y1.shape)
-    #
-    # # from matplotlib import pyplot as plt
-    # # plt.imshow(y.numpy())
-    # # plt.show()
-    static_out = np.load('/Users/zhoujun20/Desktop/code/PaddleOCR/db/db.npy')
+
+    static_out = np.load(
+        '/Users/zhoujun20/Desktop/code/PaddleOCR/output/conv.npy')
     diff = y.numpy() - static_out
     print(y.shape, static_out.shape, diff.mean())
 

ppocr/modeling/necks/rnn.py

@@ -116,7 +116,7 @@ class EncoderWithFC(nn.Layer):
 
 
 class SequenceEncoder(nn.Layer):
-    def __init__(self, in_channels, encoder_type, hidden_size, **kwargs):
+    def __init__(self, in_channels, encoder_type, hidden_size=48, **kwargs):
         super(SequenceEncoder, self).__init__()
         self.encoder_reshape = EncoderWithReshape(in_channels)
         self.out_channels = self.encoder_reshape.out_channels

tools/train.py

@@ -88,20 +88,23 @@ def main(config, device, logger, vdl_writer):
                   best_model_dict, logger, vdl_writer)
 
 
-def test_reader(config, place, logger):
-    train_loader = build_dataloader(config['TRAIN'], place)
+def test_reader(config, place, logger, global_config):
+    train_loader, _ = build_dataloader(
+        config['TRAIN'], place, global_config=global_config)
     import time
     starttime = time.time()
     count = 0
     try:
-        for data in train_loader():
+        for data in train_loader:
             count += 1
             if count % 1 == 0:
                 batch_time = time.time() - starttime
                 starttime = time.time()
-                logger.info("reader: {}, {}, {}".format(count,
-                                                        len(data), batch_time))
+                logger.info("reader: {}, {}, {}".format(
+                    count, len(data[0]), batch_time))
     except Exception as e:
+        import traceback
+        traceback.print_exc()
         logger.info(e)
     logger.info("finish reader: {}, Success!".format(count))
 
@@ -130,7 +133,7 @@ def dis_main():
                                                             device))
 
     main(config, device, logger, vdl_writer)
-    # test_reader(config, place, logger)
+    # test_reader(config, device, logger, config['Global'])
 
 
 if __name__ == '__main__':