puretorch 0.1.2

A PyTorch-like library built from scratch

---

Table of contents

Title	Desc
Version	Current version and features
Purpose	What and why of this repo
Structure	File structure of various modules
Upcoming features	Features that im working on
Setup	How to setup locally

---

Version info

Current version - 0.1.2
New features - new Flatten layer added inside PureTorch.layers

Runs on CPU only. Might add GPU support later.

---

Purpose

Raw implementation of PyTorch using NumPy. The structure and essence of torch remains the same, but its fully implemented using NumPy.

---

Structure

Current file structure:

• PureTorch
  • activations
    • ReLU
    • Softmax
    • Tanh
  • layers
    • Sequential
    • Linear
    • Flatten
  • loss
    • CrossEntropyLoss
    • SparseCrossEntropyLoss

Will be adding other layers, activations, losses, optimizers.
As of now, the default optimizer is SGD.
To check the "development code" check the "dev" branch

---

Upcoming features

These are the features that im working on, and will soon be a part of PureTorch.

• Convolutional layer(s)
  • 2D first, then 1D and multi-dim

---

Setup

Install git before running this command in your env, then run:

pip install "git+https://github.com/Dristro/PureTorch"

Run: (to verify the installation in python)

import PureTorch
print(PureTorch.__version__)

or: (to verify the installation on the terminal)

python3 -c "import PureTorch; print(PureTorch.__version__)"

---