Building core Deep Learning algorithms in Rust.

It's kinda like the middle child of karpathy/micrograd and geohot/tinygrad.

---

Contributing

Any type of contribution is welcome as long as it adds value! i.e

• Bug fixes followed with tests to ensure the bug never resurfaces
• Increasing code readability, or run-time/memory efficiency
• Completing a To-Do task

---

To-Do

• Weight Initializers

  • Glorot
  • He
  • Lecun

• Create optimizer module for optimization algorithms

  • SGD
  • Adam
  • RMSProp

• Convolutional Neural Networks

  • Convolutional Layer
  • Pooling Layer
    • Max Pooling
    • Average Pooling
  • Padding Support
    • Padding Type
  • Dilation Support

  Loading a model from Pytorch

We need the convert_state_dict() function to convert PyTorch tensors to lists because micrograd_rs can't unpickle PyTorch tensors. This conversion will allow micrograd_rs to load pytorch models without any issues.

# we need this import to serialize the model in a compatible format 
import pickle

# changes PyTorch generated state dict to micrograd state dict
def convert_state_dict(state_dict):
    new_state_dict = {}
    for name, tensor in state_dict.items():
        new_state_dict[name] = tensor.float().flatten().tolist()
    return new_state_dict

new_state_dict = convert_state_dict(model.state_dict())

# stores new state dict
with open(path, "wb") as f:
    pickle.dump(new_state_dict, f)

// To load model in rust
model.load_state_dict(path);