Конфигурация.

configs/rec_custom.yml

@@ -1,8 +1,42 @@
 Global:
+  model_name: PP-OCRv5_server_rec # To use static model for inference.
+  debug: false
   use_gpu: false
-  character_dict_path: train_data/dict.txt
-  save_model_dir: ./output/
-  pretrained_model: pretrained_models/PP-OCRv5_server_rec_pretrained.pdparams
+  device: cpu
+  epoch_num: 75
+  log_smooth_window: 20
+  print_batch_step: 10
+  save_model_dir: ./output/PP-OCRv5_server_rec
+  save_epoch_step: 1
+  eval_batch_step: [0, 2000]
+  cal_metric_during_train: true
+  calc_epoch_interval: 1
+  pretrained_model: 
+  checkpoints:
+  save_inference_dir:
+  use_visualdl: false
+  infer_img: ./train_data/images/1C4HJXEN3MW645094.jpg
+  character_dict_path: ./train_data/dict.txt
+  max_text_length: &max_text_length 19
+  infer_mode: false
+  use_space_char: true
+  distributed: true
+  save_res_path: ./output/rec/predicts_ppocrv5.txt
+  d2s_train_image_shape: [3, 48, 320]
+
+
+Optimizer:
+  name: Adam
+  beta1: 0.9
+  beta2: 0.999
+  lr:
+    name: Cosine
+    learning_rate: 0.0005
+    warmup_epoch: 1
+  regularizer:
+    name: L2
+    factor: 3.0e-05
+
 
 Architecture:
   model_type: rec
@@ -26,7 +60,13 @@ Architecture:
             fc_decay: 0.00001
       - NRTRHead:
           nrtr_dim: 384
-          max_text_length: 17
+          max_text_length: *max_text_length
+
+Loss:
+  name: MultiLoss
+  loss_config_list:
+    - CTCLoss:
+    - NRTRLoss:
 
 PostProcess:  
   name: CTCLabelDecode
@@ -37,27 +77,50 @@ Metric:
 
 Train:
   dataset:
-    name: SimpleDataSet
-    data_dir: train_data/
-    label_file_list: 
-      - train_data/train.txt
+    name: MultiScaleDataSet
+    ds_width: false
+    data_dir: ./train_data/images
+    ext_op_transform_idx: 1
+    label_file_list:
+    - ./train_data/train.txt
+    transforms:
+    - DecodeImage:
+        img_mode: BGR
+        channel_first: false
+    - RecAug:
+    - MultiLabelEncode:
+        gtc_encode: NRTRLabelEncode
+    - KeepKeys:
+        keep_keys:
+        - image
+        - label_ctc
+        - label_gtc
+        - length
+        - valid_ratio
+  sampler:
+    name: MultiScaleSampler
+    scales: [[320, 32], [320, 48], [320, 64]]
+    first_bs: &bs 128
+    fix_bs: false
+    divided_factor: [8, 16] # w, h
+    is_training: True
   loader:
-    batch_size_per_card: 16  # Уменьшите для экономии памяти
-    num_workers: 2
-
+    shuffle: true
+    batch_size_per_card: *bs
+    drop_last: true
+    num_workers: 16
 Eval:
   dataset:
     name: SimpleDataSet
-    data_dir: train_data/
+    data_dir: ./train_data/images
     label_file_list:
-      - train_data/val.txt
+    - ./train_data/val.txt
     transforms:
     - DecodeImage:
         img_mode: BGR
         channel_first: false
     - MultiLabelEncode:
         gtc_encode: NRTRLabelEncode
-        max_text_length: 17
     - RecResizeImg:
         image_shape: [3, 48, 320]
     - KeepKeys:

scripts/dataset_validation.py

@@ -1,5 +1,7 @@
-import os
 import cv2
+import os
+import random
+from PIL import Image
 
 def rename_files(directory_path: str):
     extensions = set()
@@ -37,8 +39,51 @@ def check_symbols(dir_path: str):
             if any([char not in dict_chars for char in label]):
                 print(filename)
 
+def max_height(dir_path: str):
+    max_height = 0
+    max_filename = ''
+    for filename in os.listdir(dir_path):
+        im = Image.open(os.path.join(dir_path, filename))
+        if im.height > max_height:
+            max_height = im.height
+            max_filename = filename
+
+    print(max_filename, max_height)
+
+def resize_to_height(dir_path: str, target_height=48):
+    for filename in os.listdir(dir_path):
+        with Image.open(os.path.join(dir_path, filename)) as img:
+            width_percent = target_height / float(img.height)
+            new_width = int(float(img.width) * width_percent)
+            resized_img = img.resize((new_width, target_height), Image.LANZOS)
+            resized_img.save(os.path.join(dir_path, filename))
+
+def split_dataset(dir_path: str):
+    images_dir = os.path.join(dir_path, "images")
+    filenames = os.listdir(images_dir)
+    dataset_length = len(filenames)
+    random.shuffle(filenames)
+    train_ratio = 0.8
+    val_ratio = 0.2
+    train_files_len = round(dataset_length * train_ratio)
+    val_files_len = round(dataset_length * val_ratio)
+    train_files = filenames[:train_files_len]
+    val_files = filenames[train_files_len:train_files_len+val_files_len]
+    train_file_path = os.path.join(dir_path, "train.txt")
+    val_file_path = os.path.join(dir_path, "val.txt")
+    with open(train_file_path, "w") as train_file:
+        for filename in train_files:
+            label, _ = filename.split(".")
+            train_file.write(f"{filename}\t{label}\n")
+
+    with open(val_file_path, "w") as val_file:
+        for filename in val_files:
+            label, _ = filename.split(".")
+            val_file.write(f"{filename}\t{label}\n")
+
 # rename_files("train_data/images/")
 # check_images("train_data/images/")
 # check_labels("train_data/images/")
-check_symbols("train_data/")
-
+# check_symbols("train_data/")
+# max_height("train_data/images")
+split_dataset("train_data/")