NovoGrad optimizer in PyTorch

novograd.py

# ported from https://github.com/NVIDIA/OpenSeq2Seq/blob/master/open_seq2seq/optimizers/novograd.py
# paper: https://arxiv.org/abs/1905.11286
# a recent NVidia's implementation in PyTorch: https://github.com/NVIDIA/DeepLearningExamples/blob/master/PyTorch/SpeechRecognition/Jasper/optimizers.py

import torch

class NovoGrad(torch.optim.Optimizer):
    def __init__(self, params, lr=1.0, betas = (0.95, 0.98), eps=1e-8, weight_decay=0.0, dampening=False):
        defaults = dict(lr=lr, betas=betas, eps=eps, weight_decay=weight_decay, dampening=dampening)
        super(NovoGrad, self).__init__(params, defaults)

    def step(self):
        for group in self.param_groups:
            for p in filter(lambda p: p.grad is not None, group['params']):
                state = self.state[p]
                g_2 = (p.grad.data ** 2).sum()
                state['_grads_ema'] = g_2 if '_grads_ema' not in state else state['_grads_ema'] * group['betas'][1] + g_2 * (1. - group['betas'][1])

                grad = p.grad.data / (state['_grads_ema'] + group['eps']).sqrt()
                if group['weight_decay'] > 0:
                    grad.add_(group['weight_decay'], p.data)
                if group['dampening']:
                    grad *= 1 - group['betas'][0]

                state['momentum_buffer'] = state['momentum_buffer'].mul_(group['betas'][0]).add_(grad) if 'momentum_buffer' in state else grad
                p.data.add_(-group['lr'], state['momentum_buffer'])