deeptensor 0.5.0

A minimal PyTorch-like deep learning library focused on custom autograd and efficient tensor operations.

DeepTensor 🔥

• DeepTensor: A minimal PyTorch-like deep learning library focused on custom autograd and efficient tensor operations.

---

Features at a Glance 🚀

• Automatic gradient computation with a custom autograd engine.
• Weight initialization schemes:
  • Xavier/Glorot and He initialization in both uniform and normal variants.
• Activation functions:
  • ReLU, GeLU, Sigmoid, Tanh, SoftMax, LeakyReLU, and more.
• Built-in loss functions:
  • Mean Squared Error (MSE), Cross Entropy, and Binary Cross Entropy.
• Optimizers:
  • SGD, Momentum, AdaGrad, RMSprop, and Adam.

---

Why DeepTensor?

DeepTensor offers a hands-on implementation of deep learning fundamentals with a focus on customizability and learning the internals of deep learning frameworks like PyTorch.

---

Installation

pip install deeptensor

---

Setup the project for development

git clone --recurse-submodules -j8 [e-mail]:deependujha/DeepTensor.git
cd DeepTensor

# run ctests
make ctest

# install python package in editable mode
pip install -e .

# run pytest
make test

---

Checkout Demo

• play with latest demo

---

Check Docs

• visit docs

---

Basic Usage

from deeptensor import (
    # model
    Model,

    # Layers
    Conv2D,
    MaxPooling2D,
    Flatten,
    LinearLayer,

    # activation layers
    GeLu,
    LeakyReLu,
    ReLu,
    Sigmoid,
    SoftMax,
    Tanh,

    # core objects
    Tensor,
    Value,

    # optimizers
    SGD,
    Momentum,
    AdaGrad,
    RMSprop,
    Adam,

    # losses
    mean_squared_error,
    cross_entropy,
    binary_cross_entropy,
)

model = Model(
    [
        LinearLayer(2, 16),
        ReLu(),
        LinearLayer(16, 16),
        LeakyReLu(0.1),
        LinearLayer(16, 1),
        Sigmoid(),
    ],
    False,  # using_cuda
)

opt = Adam(model, 0.01) # learning rate

print(model)

tensor_input = Tensor([2])
tensor_input.set(0, Value(2.4))
tensor_input.set(1, Value(5.2))

out = model(tensor_input)

loss = mean_squared_error(out, YOUR_EXPECTED_OUTPUT)

# backprop
loss.backward()
opt.step()
opt.zero_grad()

---

Features expected to be added

• Save & Load model
• Train a character-level transformer model
• Add support for DDP
• Add support for CUDA execution ⭐️

---

Open to Opportunities 🎅🏻🎁

I am actively seeking new opportunities to contribute to impactful projects in the deep learning and AI space.

If you are interested in collaborating or have a position that aligns with my expertise, feel free to reach out!

You can connect with me on GitHub, X (formerly twitter), or email me: [e-mail].