config.yaml

############################ Project Settings ############################

# :Params cnn: cnn layers use 'CNN5', 'ResNet50', 'DenseNet121', 'DenseNet169', 'DenseNet201';
# :Params rnn: rnn layer use 'BiLSTM' or 'BiGRU';
# :Params use_gpu:  whether use gpu. If True, model will use CudnnGRU or CudnnLSTM to accelerate computing;
# :Params image_width: the images's width;
# :Params image_height: the images's height;
# :Params channel: the images's color channels.If channel=1, it means that program will convert the image to gray mode.
# :Params resize: tuple/list of 2 integers. Reshape the image size when make dataset;
# :Params replace_transparent: If True, Program will convert image mode 'RGBA' into 'RGB';
# :Params standardization: If True, the image standardization will be done and make binaryzation/smoo Invalid argument:
#         sequence_length(0) <= 7th ops useless;
# :Params binaryzation: If -1, not use image binaryzation else choose a value between (0,255) to set binary threshold;
# :Params smooth: If -1, not use image smooth else choose a odd value greater than 1 to set MedianBlur's kernel size;
# :Params blur: If -1, not use image blur else choose a odd value greater than 1 to set GaussianBlur's kernel size;
# :Params dataset: the dataset's name, not Null;
# :Params train_path: train data path, which contains all train images. These images will be converted to tfrecords;
# :Params test_path:  test data path, which contains all test images. These images will be converted to tfrecords;
# :Params val_path: val data path, which contains all val images. These images will be converted to tfrecords;
# :Params n_class: character numbers the corpus's labels contains;
# :Params max_seq_len: max character numbers the label sequence contains;
# :Params character: all characters the corpus's labels contains, must contain '' in the list;
# :Params epochs: train epochs;
# :Params batch_size: train batch size;
# :Params end_acc: if acc is higher than 'end_acc', the training process will stop;
# :Params end_cost: if train cost is smaller than 'end_cost', the training process will stop;
# :Params learning_rate: use it in Adam
# :Params decay: use it in Adam
# :Params regularizer_l2: use in cnn layers
# :Params rnn_units: the unit number of rnn layers, default '64';
# :Params ctc_greedy: perform much faster best-path search if `true`, default 'True';
# :Params beam_width: if `greedy` is `false`: a beam search decoder will be used with a beam of this width
# :Params top_paths: that contains the log probability of each decoded sequence.
# :Params time_major: If time_major == False, this will be a `Tensor` shaped: `[batch_size, max_time, num_classes]`,
#         else `[batch_size,max_time, num_classes]`, default 'True';
# :Params ctc_merge_repeated: If `ctc_merge_repeated` is set False, then deep within the CTC calculation, repeated
#         non-blank labels will not be merged and are interpreted as individual labels, default 'True';
# :Params preprocess_collapse_repeated: If `preprocess_collapse_repeated` is True, then a preprocessing step runs before
#         loss calculation, wherein repeated labels passed to the loss are merged into single labels, default 'False';
# :Params decode_merge_repeated: If `merge_repeated` is `True`, merge repeated classes in output, default 'false'.
model:
  cnn: ResNet50
  rnn: BiLSTM
  use_gpu: False
  image_width: 150
  image_height: 50
  channel: 3
  resize: [150, 50]
  replace_transparent: True
  standardization: False
  binaryzation: -1
  smooth: -1
  blur: -1
  dataset: tencent
  train_path: cyclegan/datasets/tencent/trainA
  test_path: cyclegan/datasets/tencent/testA
  val_path: cyclegan/datasets/tencent/valA
  n_class: 38
  max_seq_len: 4
  epochs: 100000
  batch_size: 64
  end_acc: 0.9
  end_cost: 1.0
  learning_rate: 0.0001
  decay: 0.1
  regularizer_l2: 0.01
  rnn_units: 64
  ctc_greedy: True
  beam_width: 5
  top_paths: 1
  time_major: True
  ctc_merge_repeated: True
  preprocess_collapse_repeated: False
  ctc_decode_merge_repeated: False
  characters: ['a', 'c', 'b', 'd', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l',
               'm', 'n', 'o', 'p', 'q', 'r', 's', 't', 'u', 'v', 'w', 'x',
               'y', 'z', '1', '2', '3', '4', '5', '6', '7', '8', '9', '0', '']