calvinfeng/deep-learning-playground

Run experiments and try different models on GCP GPU instance

Deep Learning Playground

TODO

• Parallel batch encoding for all models
• Data augmentation logic
• Tensorboard Visualizer
• Implement NMS for decoding
• SSD
  • Anchor Generator
  • Encoder
  • Decoder
  • Loss
  • Training Loop
• CenterNet
  • Keypoint Encoder
  • Keypoint Decoder
  • Loss
  • Training Loop
• Swin Transformer
  • Model Architecture
  • Detection Head
• ResNeXT

Parallel Encoding Example

import torch.multiprocessing as mp

def encode_sample(sample):
    # This function encodes a single sample.
    # Add your own encoding logic here.
    return encoded_sample

def encode_batch(batch):
    with mp.Pool() as pool:
        encoded_batch = pool.map(encode_sample, batch)
    return torch.stack(encoded_batch)

PyTorch Details

• Use view (in PyTorch) or reshape (in numpy) to change the shape of the data without changing the order of the axes. view requires that Tensor is contiguous.
• Use permute (in PyTorch) or transpose/swapaxes (in numpy) to rearrange the axes of your data.