Introduction

[TO COME]

Project Architecture

Data

We obviously didn't push our dataset to Github for performance and confidentiality issues. However, there are a few requirements when it comes to the dataset folders organization. Our data loaders expect a specific architecture:

.data/
    train/
        input/
            1.mat
            2.mat
            ...
        output/
            1.mat
            2.mat
            ...
    test/
        input/
            1.mat
            2.mat
            ...
        output/
            1.mat
            2.mat
            ...
    valid/
        input/
            1.mat
            2.mat
            ...
        output/
            1.mat
            2.mat
            ...

where each input and output folder couples should have the same number of files. Furthermore, for a given input file, the corresponding output file should have the same name (i.e if the input file is 13.mat, so is the ouput file in the output folder).

Finally, you should change a few lines of code in loader.DataLoaders.py depending on the lengths of your datasets, right here:

    def __len__(self):
        if self.set_name_ == "train":
            return 21  # The size of the train dataset
        elif self.set_name_ == "test":
            return 10  # The size of the test dataset
        else:
            return 10  # The size of the validationd dataset

Feel free to modify the loader.DataLoaders.py even further, to better suit your needs and your input data.

To do

• Design the model's architecture

• Choose a loss function specific to the problem we're trying to solve

• Implement a data loading class to plug at the very beginning of any pytorch pipeline.

• Create the actual pytorch model, according to the design we agreed upon.

• Create the trainers

• Create the monitors to keep track of the performances throughout training and after: For instance, a utility that prints reconstructed images throughout training.