Ternary quantization

Training models with ternary quantized weights. PyTorch implementation of https://arxiv.org/abs/1612.01064

Work in progress

• Train MNIST model in original format (float32)
• Train MNIST model with quantized weights
• Add training logs
• Analyze quantized weights
• Quantize weights keeping w_p and w_n fixed

Repo Guide

• A simple model (model_full) defined in model.py was trained on MNIST data using full precision weights. The trained weight is stored as weights/original.ckpt.
  • Code for training can be found under main_original.py.
• A copy of the above model (loaded with trained weights) was created (model_to_quantify) and was trained using quantization. The trained weight is stored as weights/quantized.ckpt.
  • Code for training can be found under main_ternary.py. The logs can be found inside the file logs/quantized_wp_wn_trainable.txt.
• I also tried updating the weights by an equal amount in the direction of their gradients. In other words, I took the sign of every parameter’s gradient and updated the parameter by a small value (0.001) like so:
  param.grad.data = torch.sign(param.grad.data) * 0.001
  • I got decent results but didn’t dig deeper into it. The weights for this model are weights/autoquantize.ckpt.

Notes:

• Full precision model gives an accuracy of 98.8%
• Quantized model gives an accuracy of as high as 98.52%
  • I slightly changed the way gradients are calculated. Using mean instead of sum in lines 15 an 16, quantification.py gave better results:

    w_p_grad = (a * grad_data).mean() # not (a * grad_data).sum()
    w_n_grad = (b * grad_data).mean() # not (b * grad_data).sum()