Fixed Main.c

[Advaitgaur004]

Description

This PR implements proper fixes and enhancements to the neural network library, with a focus on correcting the main.c training loop and gradient computation.

Changes

Core Operations

• Added Tensor_transpose function for matrix transposition
• Fixed GradFn_matmul to properly handle gradient computation using transposition
• Added Tensor_sub operation with appropriate gradient functions
• Enhanced broadcasting to properly handle scalar broadcasting

Neural Network Components

• Added nn_random_init function implementing Xavier/Glorot initialization
• Fixed ReLU gradient function to properly use 0.0f/1.0f values
• Corrected softmax gradient computation to handle batched inputs properly
• Added GradNode naming for better debugging and gradient tracking

Loss Functions

• Fixed nn_crossentropy to avoid numerical issues by adding epsilon
• Added nn_softmax_crossentropy to stabilize training
• Implemented proper gradients for both loss functions

Training Improvements

• Added data normalization to preprocess inputs
• Implemented dataset shuffling for better training
• Enhanced batch handling with proper actual_batch_size calculations
• Improved learning rate from 0.001 to 0.01
• Added epoch average loss tracking for monitoring

Memory Management

• Fixed memory allocation and deallocation
• Better variable initialization

[Advaitgaur004]

Works !!