add DBNet and DBNet++ support

configs/det/db++_r50_icdar15.yaml

@@ -0,0 +1,152 @@
+system:
+  mode: 1 # 0 for graph mode, 1 for pynative mode in MindSpore
+  distribute: False
+  amp_level: 'O0'
+  seed: 42
+  val_while_train: True
+  ckpt_save_dir: './tmp_det'
+
+model:
+  type: det
+  transform: null
+  backbone:
+    name: det_resnet50
+    pretrained: True
+  neck:
+    name: DBFPN
+    out_channels: 256
+    bias: False
+    use_asf: True
+  head:
+    name: DBHead
+    k: 50
+    bias: False
+    adaptive: True
+
+postprocess:
+  name: DBPostprocess
+  region_type: 'quad'
+  thresh: 0.3
+  box_thresh: 0.55 # TODO: this value is 0.55 in modelzoo and but 0.7 in paddle
+  max_candidates: 1000
+  unclip_ratio: 1.5
+
+metric:
+  name: DetMetric
+  main_indicator: f-score
+
+loss:
+  name: L1BalancedCELoss
+  eps: 1.0e-6
+  l1_scale: 10
+  bce_scale: 5
+  bce_replace: bceloss
+
+scheduler:
+  scheduler: polynomial_decay
+  min_lr: 0.
+  lr: 0.007
+  num_epochs: 1200
+  warmup_epochs: 3
+
+optimizer:
+  opt: SGD
+  filter_bias_and_bn: false
+  momentum: 0.9
+  weight_decay: 1.0e-4
+  loss_scale: 1.0
+
+train:
+  dataset_sink_mode: False
+  dataset:
+    type: DetDataset
+    data_dir: /path_to_image_folder/
+    label_files: /path_to_labels.txt
+    sample_ratios: [ 1.0 ]
+    shuffle: True
+    transform_pipeline:
+      - DecodeImage:
+          img_mode: BGR
+          to_float32: False
+      - DetLabelEncode:
+      - MZResizeByGrid:
+          divisor: 32
+          transform_polys: True # originally in modelzoo, it doesn't transform polys
+      - MZRandomScaleByShortSide:
+          short_side: 736
+      - IaaAugment:
+          augmenter_args:
+            - { 'type': 'Affine', 'args': { 'rotate': [ -10, 10 ] } }
+            - { 'type': 'Fliplr', 'args': { 'p': 0.5 } }
+      - MZRandomCropData:
+          max_tries: 100
+          min_crop_side_ratio: 0.1
+          crop_size: [ 640, 640 ]
+      #- MZResizeByGrid:  # TODO: necessary? 640 is divisable by 32
+      #        denominator: 32
+      #        divisor: 32
+      #    transform_polys: True
+      #- MakeShrinkMap: 
+      - MZMakeSegDetectionData:
+          min_text_size: 8
+          shrink_ratio: 0.4
+      #- 'MakeBorderMap': 
+      - MZMakeBorderMap:
+          shrink_ratio: 0.4
+          thresh_min: 0.3
+          thresh_max: 0.7
+      - MZRandomColorAdjust:
+          brightness: 0.1255 #32.0 / 255
+          saturation: 0.5
+          to_numpy: True
+      #{'MZIrregularNormToCHW': None},
+      - NormalizeImage:
+          bgr_to_rgb: True
+          is_hwc: True
+          mean: imagenet
+          std: imagenet
+      - ToCHWImage:
+    #  the order of the dataloader list, matching the network input and the input labels for the loss function, and optional data for debug/visaulize 
+    output_keys: [ 'image', 'shrink_map', 'shrink_mask', 'threshold_map', 'threshold_mask' ] #'img_path']
+    num_keys_to_net: 1 # num inputs for network forward func in output_keys
+#    keys_for_loss: 4 # num labels for loss func
+
+  loader:
+    shuffle: True # TODO: tbc
+    batch_size: 20
+    drop_remainder: False
+    max_rowsize: 20
+    num_workers: 10
+
+eval:
+  dataset_sink_mode: False
+  dataset:
+    type: DetDataset
+    data_dir: /path_to_image_folder/
+    label_files: /path_to_labels.txt
+    sample_ratios: [ 1.0 ]
+    shuffle: False
+    transform_pipeline:
+      - DecodeImage:
+          img_mode: BGR
+          to_float32: False
+      - DetLabelEncode:
+      - MZScalePad:
+          eval_size: [ 736, 1280 ] # h, w
+      - NormalizeImage:
+          bgr_to_rgb: True
+          is_hwc: True
+          mean: imagenet
+          std: imagenet
+      - ToCHWImage:
+    #  the order of the dataloader list, matching the network input and the labels for evalution 
+    output_keys: [ 'image', 'polys', 'ignore_tags' ] #'shape'] #'img_path']
+    num_keys_to_net: 1 # num inputs for network forward func
+    num_keys_of_labels: 2 # num labels 
+
+  loader:
+    shuffle: False
+    batch_size: 1 # TODO: due to dynamic shape of polygons (num of boxes varies), BS has to be 1
+    drop_remainder: False
+    max_rowsize: 20
+    num_workers: 1

configs/det/db_r50_icdar15.yaml

@@ -1,6 +1,6 @@
 system:
   mode: 1 # 0 for graph mode, 1 for pynative mode in MindSpore
-  distribute: False 
+  distribute: False
   amp_level: 'O0'
   seed: 42
   val_while_train: True
@@ -11,152 +11,147 @@ model:
   transform: null
   backbone:
     name: det_resnet50
-    pretrained: True 
+    pretrained: True
   neck:
     name: DBFPN
     out_channels: 256
     bias: False
     use_asf: False # enable it for DB++ 
   head:
-    name: DBHead 
+    name: DBHead
     k: 50
     bias: False
     adaptive: True
-    serial: False
-
+
 postprocess:
   name: DBPostprocess
   region_type: 'quad'
   thresh: 0.3
   box_thresh: 0.55 # TODO: this value is 0.55 in modelzoo and but 0.7 in paddle
   max_candidates: 1000
   unclip_ratio: 1.5
- 
+
 metric:
   name: DetMetric
-  main_indicator: hmean
+  main_indicator: f-score
 
 loss:
-  name: L1BalanceCELoss  
-  eps: 0.000001
+  name: L1BalancedCELoss
+  eps: 1.0e-6
   l1_scale: 10
-  bce_scale: 5 
+  bce_scale: 5
   bce_replace: bceloss
 
-scheduler: 
-  scheduler: "cosine_decay"
-  min_lr: 0.00001
-  lr: 0.001
+scheduler:
+  scheduler: polynomial_decay
+  min_lr: 0.
+  lr: 0.007
   num_epochs: 1200
-  warmup_epochs: 20
-  decay_epochs: 1180
+  warmup_epochs: 3
 
 optimizer:
-  opt: "adamw"
-  filter_bias_and_bn: True
+  opt: SGD
+  filter_bias_and_bn: false
   momentum: 0.9
-  weight_decay: 0.001
+  weight_decay: 1.0e-4
   loss_scale: 1.0
-  #use_nesterov: False
 
 train:
   dataset_sink_mode: True 
   dataset:
     type: DetDataset
-    #data_dir: /data/ocr_datasets/ic15/text_localization/train/ch4_training_images
-    #label_files: /data/ocr_datasets/ic15/text_localization/train/det_gt.txt
-    data_dir: /Users/Samit/Data/datasets/ic15/det/train/ch4_training_images
-    label_files: /Users/Samit/Data/datasets/ic15/det/train/det_gt.txt
-    sample_ratios: [1.0]
+    data_dir: /data/ocr_datasets/ic15/text_localization/train
+    label_files: /data/ocr_datasets/ic15/text_localization/train/train_icdar15_label.txt
+    #data_dir: /Users/Samit/Data/datasets/ic15/det/train
+    #label_files: /Users/Samit/Data/datasets/ic15/det/train/train_icdar2015_label.txt
+    sample_ratios: [ 1.0 ]
     shuffle: True
     transform_pipeline:
-      - DecodeImage: 
+      - DecodeImage:
           img_mode: BGR
           to_float32: False
-      - DetLabelEncode:  
+      - DetLabelEncode:
       - MZResizeByGrid:
           divisor: 32
-          transform_polys: True
-      - MZRandomScaleByShortSide: 
+          transform_polys: True # originally in modelzoo, it doesn't transform polys
+      - MZRandomScaleByShortSide:
           short_side: 736
-      - IaaAugment: 
+      - IaaAugment:
           augmenter_args:
-            - {'type': 'Affine', 'args': {'rotate': [-10, 10]}}
-            - {'type': 'Fliplr', 'args': {'p': 0.5}}
-      - MZRandomCropData: 
-          max_tries: 100 
+            - { 'type': 'Affine', 'args': { 'rotate': [ -10, 10 ] } }
+            - { 'type': 'Fliplr', 'args': { 'p': 0.5 } }
+      - MZRandomCropData:
+          max_tries: 100
           min_crop_side_ratio: 0.1
-          crop_size: [640, 640]
-      - MZMakeSegDetectionData: 
+          crop_size: [ 640, 640 ]
+      - MZMakeSegDetectionData:
           min_text_size: 8
           shrink_ratio: 0.4
       - MZMakeBorderMap:
           shrink_ratio: 0.4
           thresh_min: 0.3
           thresh_max: 0.7
-      - MZRandomColorAdjust: 
+      - MZRandomColorAdjust:
           brightness: 0.1255 #32.0 / 255
           saturation: 0.5
           to_numpy: True
-      - NormalizeImage: 
+      - NormalizeImage:
           bgr_to_rgb: True
           is_hwc: True
-          mean : [123.675, 116.28, 103.53]
-          std : [58.395, 57.12, 57.375]
-      - ToCHWImage: 
+          mean: imagenet
+          std: imagenet
+      - ToCHWImage:
     #  the order of the dataloader list, matching the network input and the input labels for the loss function, and optional data for debug/visaulize 
     output_keys: ['image', 'shrink_map', 'shrink_mask', 'threshold_map', 'threshold_mask'] #'img_path'] 
     #output_keys: ['image'] # for debug op performance
     num_keys_to_net: 1 # num inputs for network forward func in output_keys
-    #keys_for_loss: 4 # num labels for loss func
-     
+#    keys_for_loss: 4 # num labels for loss func
+
   loader:
-      shuffle: True
-      batch_size: 16
-      drop_remainder: True
-      max_rowsize: 20
-      num_workers: 1
+    shuffle: True # TODO: tbc
+    batch_size: 20
+    drop_remainder: False
+    max_rowsize: 20
+    num_workers: 10 # TODO: may lead to OOM
 
 eval:
   dataset_sink_mode: False
   dataset:
     type: DetDataset
-    #data_dir: /data/ocr_datasets/ic15/text_localization/test/ch4_test_images
-    #label_files: /data/ocr_datasets/ic15/text_localization/test/det_gt.txt
-    data_dir: /Users/Samit/Data/datasets/ic15/det/test/ch4_test_images
-    label_files: /Users/Samit/Data/datasets/ic15/det/test/det_gt.txt
-    sample_ratios: [1.0]
+    data_dir: /data/ocr_datasets/ic15/text_localization/test
+    label_files: /data/ocr_datasets/ic15/text_localization/test/test_icdar2015_label.txt
+    #data_dir: /Users/Samit/Data/datasets/ic15/det/test
+    #label_files: /Users/Samit/Data/datasets/ic15/det/test/test_icdar2015_label.txt
+    sample_ratios: [ 1.0 ]
     shuffle: False
     transform_pipeline:
-      - DecodeImage: 
+      - DecodeImage:
           img_mode: BGR
           to_float32: False
-      - DetLabelEncode:  
-      - MZScalePad: 
-          eval_size: [736, 1280] # h, w
-      - NormalizeImage: 
+      - DetLabelEncode:
+      - MZScalePad:
+          eval_size: [ 736, 1280 ] # h, w
+      - NormalizeImage:
           bgr_to_rgb: True
           is_hwc: True
-          mean : [123.675, 116.28, 103.53]
-          std : [58.395, 57.12, 57.375]
-      - ToCHWImage: 
+          mean: imagenet
+          std: imagenet
+      - ToCHWImage:
     #  the order of the dataloader list, matching the network input and the labels for evalution 
-    output_keys: ['image', 'polys', 'ignore_tags'] #'shape'] #'img_path'] 
+    output_keys: [ 'image', 'polys', 'ignore_tags' ] #'shape'] #'img_path']
     num_keys_to_net: 1 # num inputs for network forward func
     num_keys_of_labels: 2 # num labels 
-
-
+
   loader:
-      shuffle: False
-      batch_size: 1 # TODO: due to dynamic shape of polygons (num of boxes varies), BS has to be 1
-      drop_remainder: False 
-      max_rowsize: 20
-      num_workers: 1
+    shuffle: False
+    batch_size: 1 # TODO: due to dynamic shape of polygons (num of boxes varies), BS has to be 1
+    drop_remainder: False 
+    max_rowsize: 20
+    num_workers: 1
 
-modelarts: # for running on modelarts or openi
+modelarts: # TODO: for running on modelarts or openi. Not making effect currently.
    enable_modelarts: False
    data_url: /cache/data/  # path to dataset
    multi_data_url: /cache/data/ # path to multi dataset 
    ckpt_url: /cache/output/  # pretrained model path
    train_url: /cache/output/ # model save folder
-

mindocr/data/base_dataset.py

@@ -1,6 +1,3 @@
-#from __future__ import absolute_import
-from __future__ import division
-
 from typing import Union, List
 import random
 import os

mindocr/data/det_dataset.py

@@ -1,6 +1,3 @@
-from __future__ import absolute_import
-from __future__ import division
-
 from .base_dataset import BaseDataset
 
 __all__ = ['DetDataset']

mindocr/data/rec_dataset.py

@@ -1,6 +1,3 @@
-from __future__ import absolute_import
-from __future__ import division
-
 from .base_dataset import BaseDataset
 
 __all__ = ['RecDataset']