A PyTorch implementation of Convolutional autoencoder (CAE) and CNN on cifar-10.

The default configuration of this repository jointly trains CAE and CNN at the same time. The training scheme is presented below.

1. Call autoencoder_main.py
2. Log files and tensorboard files recording acc and loss are produced under the directory of ./files/output

There are 2 locations to modify to change the behavior of the model and training.

• Json files under ./files/input/models/configs/*.json

  • "use_cnn": true if training CNN
  • "use_cae": true if training CAE
  • "only_train_cae_until": Only CAE is trained while specified number of epoch here (int). Afterward, CNN is jointed with CAE and the two models are trained.
  • "train_data_regulation": This parameter is map such that {cls_nm: data_num, ...}. The number of training data per class is regulated according to this parameter.
  • "train_data_expansion": This parameter is map such that {cls_nm: data_num, ...}. The number of training data per class is oversampled according to this parameter.
  • "cnn"
    • "block_sizes": This parameter is a list of lists such that [[1st_conv_depth_from_in(int), to(int), use_max_pooling(boolean)], [2nd,...,...]],
    • "use_mixed_input": true if encoded feature from CAE is inserted into the first feature maps, otherwise, encoded features are used as input of CNN.
  • "cae":
    • "enc_block_sizes": This parameter is a list of lists such that [[1st_conv_depth_from_in(int), to(int)], [2nd,...,...]]
    • "dec_block_sizes": Same with the enc_block_sizes.

• Arguments specified in autoencoder_main.py

  • An important argument is model_config_key, corresponding to the file name of Json file explained above.
  • Other arguments can be used as default. Specification is provided in autoencoder_main.py.

• dataset: Abstract and sub-class of Cifar-10
• files:
  • input/dataset: Location where Cifar-10 data file is downloaded
  • input/models/configs: There are configuration files controlling model structure and etc
  • output/board: Tensorboard files are produced
  • output/logs: Log files are outputted
  • output/images: Input and output images of CAE are produced
• models: Model definition of CAE and CNN as well as a whole network (classifier.py) of CAE and CNN
• trainer: Abstract class and sub-class of model trainer