This project serves as a basic template for any Pytorch project.

The idea is to standardize code structures as much as possible for various projects

• Clone/Fork the repository and use the structure below as a reference to add/update/delete code.
• Use the requirements.txt file to add all the dependencies in your own virtual environment.

pytorch-template-project
├── models
│   └── alexnet.py
│ 
├── preprocess
│   └── transforms.py 
│ 
├── test
│   └── test_helpers.py
│ 
├── utils
│   ├── helpers.py
│   └── setup_configs.py
│ 
├── .coverage
│ 
├── .gitignore
│ 
├── configs.json
│ 
├── coverage.svg
│ 
├── main.py
│ 
├── README.md
│ 
└── requirements.txt

This is just a base config file that I have provided. Please feel free to add additional configs attributes as needed. Although, I highly recommend not to change the names of the attributes unless necessary. Certain attributes like optimizer, criterion, data, model_name are being used to load their specific Object types. If you must change the attribute names, then update the setup_configs.py and helpers.py to use the updated attribute names.

• experiment_name : The name of the experiment
• model_name : The architecture to use. Architectures included in the code are
  • Options : AlexNet
  • Default : alexnet
• optimizer : The optimizer to use for the experiment
  • Options : adam, sgd
  • Default : sgd
• criterion : The Loss function to use for the experiment
  • Default : cross-entropy
• num_epochs : Number of epochs to train the model for
  • Default : 10
• num_classes : The number of classes for classification
  • Default : 10
• input_channels : The input channels for the model
  • Default : 3
• learning_rate : The learning rate to use for the experiments. We have not used Adaptive learning rate for the simplicity in interpreting the trends
  • Default : 0.001
• momentum : Used for the optimizer
  • Default : 0.9
• weight_decay : Used for the optimizer
  • Default : 1e-5
• batch_size : The number of images per batch in the training dataset
  • Default : 32
• data : (dict) The path for the dataset
  • root : Path for the root directory
  • train: Relative Path for the train folder from the root
  • validation : Relative Path for the validation folder from the root
  • test : Relative Path for the test folder from the root
• seed : This initializes different weights for each experiment. We average over the results to determine a general learning trend
  • Options : Length of list defines the number of different experiments for a given arch and dataset
  • Default : [0, 123, 420, 69]
• model_save_path : Path where the model will be saved after training
  • Default : ./trained_models/

• If you have any issues, then create the issue with the fix you have one and create a pull request.
• If you have new features, submit a pull request, and I will review it.