/MyTorch

A deep learning library implemented from scratch.

Primary LanguagePythonOtherNOASSERTION

⚠️ IMPORTANT DISCLAIMER: This repository contains code from an academic assignment completed at Carnegie Mellon University (CMU). It is shared here solely to showcase my work. Please do not use, reproduce, distribute, or reference this code in any form. The intention is purely to display the project and not to provide a solution or reference for current or future students. If you're a student at CMU or any other institution, please adhere to your university's academic integrity guidelines. Misuse of this code is a violation of the repository's license.

MyTorch®

This repository is an implementation of a deep learning library written in Python. Inspired by PyTorch, this library, MyTorch® is used to create everything from multilayer perceptrons (MLP), convolutional neural networks (CNN), to recurrent neural networks with gated recurrent units (GRU) and long-short term memory (LSTM) structures.

MyTorch® is structurally similar to to PyTorch and TensorFlow and its modules can be reused in other applications.

Setup

  1. Clone the repository and run poetry install.
  2. Run make help for further instructions.

Reference

  • MLP:
    • Linear Layer [mytorch.nn.Linear]
    • Activation Functions:
      • Sigmoid [mytorch.nn.Sigmoid]
      • ReLU [mytorch.nn.ReLU]
      • Tanh [mytorch.nn.Tanh]
    • Neural Network Models:
      • MLP0 (Hidden Layers = 0) [mytorch.models.MLP0]
      • MLP1 (Hidden Layers = 1) [mytorch.models.MLP1]
      • MLP4 (Hidden Layers = 4) [mytorch.models.MLP4]
    • Criterion: Loss Functions
      • MSE Loss [mytorch.nn.MSELoss]
      • Cross-Entropy Loss [mytorch.nn.CrossEntropyLoss]
    • Optimizers:
      • Stochastic Gradient Descent (SGD) [mytorch.optim.SGD]
    • Regularization:
      • Batch Normalization [mytorch.nn.BatchNorm1d]
  • CNN:
    • Resampling:
      • Upsample1d [mytorch.nn.Upsample1d]
      • Downsampling1d [mytorch.nn.Downsampling1d]
      • Upsampling2d [mytorch.nn.Upsampling2d]
      • Downsampling2d [mytorch.nn.Downsampling2d]
    • Convolutional Layer:
      • Conv1d_stride1 [mytorch.nn.Conv1d_stride1]
      • Conv1d [mytorch.nn.Conv1d]
      • Conv2d_stride1 [mytorch.nn.Conv2d_stride1]
      • Conv2d [mytorch.nn.Conv2d]
    • Transposed Convolution:
      • ConvTranspose1d [mytorch.nn.ConvTranspose1d]
      • ConvTranspose2d [mytorch.nn.ConvTranspose2d]
    • Pooling:
      • MaxPool2d_stride1 or MeanPool2d_stride1 [mytorch.nn.MaxPool2d_stride1] or [mytorch.nn.MeanPool2d_stride1]
      • MaxPool2d or MeanPool2d [mytorch.nn.MaxPool2d] or [mytorch.nn.MeanPool2d]
  • RNN:
    • RNN Cell [mytorch.nn.rnn_cell]
    • RNN Phoneme Classifier [models.rnn_classifier]
    • GRU Cell [mytorch.nn.gru_cell]
    • CTC [mytorch.nn.ctc]
      • CTC Loss
    • CTC Decoding: Greedy Search and Beam Search [mytorch.nn.CTCDecoding]

Made with ☕ as part of Carnegie Mellon University's 11-785 Introduction to Deep Learning.