Medix-AI/coqui-tts
Medix-AI
/
coqui-tts
mirror of https://github.com/coqui-ai/TTS.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

radam pytorch 1.5 update

This commit is contained in:
erogol 2020-05-12 13:49:23 +02:00
parent 574968b249
commit 3b2d726e2d
1 changed files with 41 additions and 98 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

135
utils/radam.py
View File

@ -1,17 +1,31 @@
# from https://github.com/LiyuanLucasLiu/RAdam
import math import math
import torch import torch
from torch.optim.optimizer import Optimizer from torch.optim.optimizer import Optimizer, required
# adapted from https://github.com/LiyuanLucasLiu/RAdam
class RAdam(Optimizer): class RAdam(Optimizer):
def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8, weight_decay=0): def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8, weight_decay=0, degenerated_to_sgd=True):
defaults = dict(lr=lr, betas=betas, eps=eps, weight_decay=weight_decay) if not 0.0 <= lr:
self.buffer = [[None, None, None] for ind in range(10)] raise ValueError("Invalid learning rate: {}".format(lr))
if not 0.0 <= eps:
raise ValueError("Invalid epsilon value: {}".format(eps))
if not 0.0 <= betas[0] < 1.0:
raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
if not 0.0 <= betas[1] < 1.0:
raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
self.degenerated_to_sgd = degenerated_to_sgd
if isinstance(params, (list, tuple)) and len(params) > 0 and isinstance(params[0], dict):
for param in params:
if 'betas' in param and (param['betas'][0] != betas[0] or param['betas'][1] != betas[1]):
param['buffer'] = [[None, None, None] for _ in range(10)]
defaults = dict(lr=lr, betas=betas, eps=eps, weight_decay=weight_decay, buffer=[[None, None, None] for _ in range(10)])
super(RAdam, self).__init__(params, defaults) super(RAdam, self).__init__(params, defaults)
def __setstate__(self, state): # pylint: disable= useless-super-delegation def __setstate__(self, state):
super(RAdam, self).__setstate__(state) super(RAdam, self).__setstate__(state)
def step(self, closure=None): def step(self, closure=None):
@ -27,128 +41,57 @@ class RAdam(Optimizer):
continue continue
grad = p.grad.data.float() grad = p.grad.data.float()
if grad.is_sparse: if grad.is_sparse:
raise RuntimeError( raise RuntimeError('RAdam does not support sparse gradients')
'RAdam does not support sparse gradients')
p_data_fp32 = p.data.float() p_data_fp32 = p.data.float()
state = self.state[p] state = self.state[p]
if not state: if len(state) == 0:
state['step'] = 0 state['step'] = 0
state['exp_avg'] = torch.zeros_like(p_data_fp32) state['exp_avg'] = torch.zeros_like(p_data_fp32)
state['exp_avg_sq'] = torch.zeros_like(p_data_fp32) state['exp_avg_sq'] = torch.zeros_like(p_data_fp32)
else: else:
state['exp_avg'] = state['exp_avg'].type_as(p_data_fp32) state['exp_avg'] = state['exp_avg'].type_as(p_data_fp32)
state['exp_avg_sq'] = state['exp_avg_sq'].type_as( state['exp_avg_sq'] = state['exp_avg_sq'].type_as(p_data_fp32)
p_data_fp32)
exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq'] exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
beta1, beta2 = group['betas'] beta1, beta2 = group['betas']
exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad) exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
exp_avg.mul_(beta1).add_(1 - beta1, grad) exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1)
state['step'] += 1 state['step'] += 1
buffered = self.buffer[int(state['step'] % 10)] buffered = group['buffer'][int(state['step'] % 10)]
if state['step'] == buffered[0]: if state['step'] == buffered[0]:
N_sma, step_size = buffered[1], buffered[2] N_sma, step_size = buffered[1], buffered[2]
else: else:
buffered[0] = state['step'] buffered[0] = state['step']
beta2_t = beta2 ** state['step'] beta2_t = beta2 ** state['step']
N_sma_max = 2 / (1 - beta2) - 1 N_sma_max = 2 / (1 - beta2) - 1
N_sma = N_sma_max - 2 * \ N_sma = N_sma_max - 2 * state['step'] * beta2_t / (1 - beta2_t)
state['step'] * beta2_t / (1 - beta2_t)
buffered[1] = N_sma buffered[1] = N_sma
# more conservative since it's an approximated value # more conservative since it's an approximated value
if N_sma >= 5: if N_sma >= 5:
step_size = group['lr'] * math.sqrt((1 - beta2_t) * (N_sma - 4) / (N_sma_max - 4) * ( step_size = math.sqrt((1 - beta2_t) * (N_sma - 4) / (N_sma_max - 4) * (N_sma - 2) / N_sma * N_sma_max / (N_sma_max - 2)) / (1 - beta1 ** state['step'])
N_sma - 2) / N_sma * N_sma_max / (N_sma_max - 2)) / (1 - beta1 ** state['step']) elif self.degenerated_to_sgd:
step_size = 1.0 / (1 - beta1 ** state['step'])
else: else:
step_size = group['lr'] / (1 - beta1 ** state['step']) step_size = -1
buffered[2] = step_size buffered[2] = step_size
if group['weight_decay'] != 0:
p_data_fp32.add_(-group['weight_decay']
* group['lr'], p_data_fp32)
# more conservative since it's an approximated value # more conservative since it's an approximated value
if N_sma >= 5: if N_sma >= 5:
if group['weight_decay'] != 0:
p_data_fp32.add_(p_data_fp32, alpha=-group['weight_decay'] * group['lr'])
denom = exp_avg_sq.sqrt().add_(group['eps']) denom = exp_avg_sq.sqrt().add_(group['eps'])
p_data_fp32.addcdiv_(-step_size, exp_avg, denom) p_data_fp32.addcdiv_(exp_avg, denom, value=-step_size * group['lr'])
else: p.data.copy_(p_data_fp32)
p_data_fp32.add_(-step_size, exp_avg) elif step_size > 0:
if group['weight_decay'] != 0:
p.data.copy_(p_data_fp32) p_data_fp32.add_(p_data_fp32, alpha=-group['weight_decay'] * group['lr'])
p_data_fp32.add_(exp_avg, alpha=-step_size * group['lr'])
return loss
class PlainRAdam(Optimizer):
def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8, weight_decay=0):
defaults = dict(lr=lr, betas=betas, eps=eps, weight_decay=weight_decay)
super(PlainRAdam, self).__init__(params, defaults)
def __setstate__(self, state): # pylint: disable= useless-super-delegation
super(PlainRAdam, self).__setstate__(state)
def step(self, closure=None):
loss = None
if closure is not None:
loss = closure()
for group in self.param_groups:
for p in group['params']:
if p.grad is None:
continue
grad = p.grad.data.float()
if grad.is_sparse:
raise RuntimeError(
'RAdam does not support sparse gradients')
p_data_fp32 = p.data.float()
state = self.state[p]
if not state:
state['step'] = 0
state['exp_avg'] = torch.zeros_like(p_data_fp32)
state['exp_avg_sq'] = torch.zeros_like(p_data_fp32)
else:
state['exp_avg'] = state['exp_avg'].type_as(p_data_fp32)
state['exp_avg_sq'] = state['exp_avg_sq'].type_as(
p_data_fp32)
exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
beta1, beta2 = group['betas']
exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)
exp_avg.mul_(beta1).add_(1 - beta1, grad)
state['step'] += 1
beta2_t = beta2 ** state['step']
N_sma_max = 2 / (1 - beta2) - 1
N_sma = N_sma_max - 2 * state['step'] * beta2_t / (1 - beta2_t)
if group['weight_decay'] != 0:
p_data_fp32.add_(-group['weight_decay']
* group['lr'], p_data_fp32)
# more conservative since it's an approximated value
if N_sma >= 5:
step_size = group['lr'] * math.sqrt((1 - beta2_t) * (N_sma - 4) / (N_sma_max - 4) * (
N_sma - 2) / N_sma * N_sma_max / (N_sma_max - 2)) / (1 - beta1 ** state['step'])
denom = exp_avg_sq.sqrt().add_(group['eps'])
p_data_fp32.addcdiv_(-step_size, exp_avg, denom)
else:
step_size = group['lr'] / (1 - beta1 ** state['step'])
p_data_fp32.add_(-step_size, exp_avg)
p.data.copy_(p_data_fp32) p.data.copy_(p_data_fp32)
return loss return loss