Neograd

- Supports Python 3.11
- Uses dill to save params as well and moved away from h5py, thus having one package to manage all saving and loading of objects
- Restructured layers.py into multiple modules for ease of access

ಕನ್ನಡವೇ ಸತ್ಯ ಕನ್ನಡವೇ ನಿತ್ಯ

- get_params renamed as parameters in all Model, Layer, Container
- Freezing of params is now accomplished by requires_grad and not a separate frozen attribute which allows for dynamic freezing and unfreezing of Params
- Backward can now be called on intermediate tensors in the graph
- Refactor unbroadcasting in backward pass to be more efficient
- Fixed buggy Dropout and LeakyReLU layers
- Tests now run in lesser time(13s) as opposed to 24s earlier because the size of inputs has now been decreased

- Gradient checking added to check the correctness of gradients calculated by autograd
- Optimization algorithms like Momentum, RMSProp, Adam added
- Convolutional Neural Networks added with MaxPooling
- Save models, weights to disk and load them whenever required
- Add checkpoints while training to prevent loss of trained weights in case of catastrophes like running out of battery
- Tests added for layers, loss functions, activations
- Documentation available at neograd.readthedocs.io/

Initial release of neograd, a deep learning framework created from scratch using Python and NumPy with a PyTorch like API. Autograd is the automatic differentiation engine for the framework, loss functions such as MSE, BCE are provided, basic gradient descent optimiser takes care of optimization. Will be adding several new features in the coming days! Keep an eye out! Cheers!