Pytorch Multichannel Image Classification

In this repository, you can find PyTorch based image classification code which can be used for data with any number of channels. This is in general important as some applications such as satellite imagery or biomedical imaging can contain more than 3 channels of information.

Folders

configs: includes the config files
data_loaders: includes the dataset and dataloders
sample_data: example of MNIST as well as Reconstructing cell cycle progression for running the code
sample_output: example of output files for a LeNet architecture
eval: code for predicting over a new dataset
losses: code for loss functions
models: code for different models, characteristics of the models and their feature extration
optimizers: code for optimizeritation functions
losses: code for loss functions
tensorboard_writer: includes the methods for tensorboard as well as saving results
train: includes the methods for training the model, calculating the metircs as well as the learning rate scheduler

Input parameters

For using this code, you need to use the config file following information:

data_dir : directory of the images in their corresponding folders (each folder is a class) (format: str)
test_data_dir : directory of the test set images in their corresponding folders (each folder is a class) (defualt: None, format: str)
batch_size: Batch size
validation_split: percentage of train-validation split
data_map: list of the maps which can be used for mapping data from one range to another. In case [] is passed, no transformation will be done on the data. At the moment, possible options are: normalize, map_zero_one and map_minus_one_to_one
augmentation: list which can be used for data augemtnation, at the moment the options are ["random_crop", "random_rotation", "random_flip"]. In case there is no need for augmentation, you can pass []
test_split: percentage of test vs train-validation split (It will be discarded if the test_data_dir is None)
tensorboard_path: path to save tensorboard as well as other outputs
file_extension: file extension which exists in the data, for example .png
checkpoint: the path, for transfer learning. In case it is not passed it will not be considered. The checkpoint should include: 'epoch', model.state_dict(), optimizer.state_dict() and the loss.
model_name: the used architecture. The name should be exactly the same from the file models.py:
num_epochs: Number of epochs
device: It can be cpu or cuda. I have intentionally kept this option and the code does not look if cuda exists
optimization_parameters: a dictionary including the options which are needed for the optimizers.py
optimization_method: name of the optimizer in the file optimizers.py
loss_function: name of the loss function in the file losses.py
metrics_of_interest: list of metrics which should be trace. They should be same as the metrics in the files metrics.py
call_back: dictionary of parameters which are needed for early stopping as well as saving models. It includes saving_period which indicates how many epochs are the models are saved. patience which the model can tolerate not getting improved as well as the criteria which we look at for best model selection. The criteria must be one of the metrics_of_interest.

Running the Code

For running this code, it is enough to just pass the config file and run the main.py. For testing whether the code works or not, you can use the folder sample_data which includes MNIST dataset with two channels (the second channel is artificially created)

python main.py --config ./configs/sample_config.json

Libraries

This code is based Python 3.7.6 (Anaconda) using these libraries:

torch=='1.4.0'
tensorboard=='2.1.0'
pandas=='1.0.0'
sklearn=='0.22.1'
skimage=='0.16.2'
pillow=='6.2.0'

Naming Convention

Each run is named using multiple conditions to be able to distinguish their results. The naming is based on:

DATETIME_NOW + _ + MODEL + OPTIMIZATION_METHOD + _ + LR