File size: 4,974 Bytes
f32c034 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 | import math
import torch
from torch.optim import Optimizer
# From: https://github.com/uds-lsv/bert-stable-fine-tuning
# Paper: On the Stability of Fine-tuning BERT: Misconceptions, Explanations, and Strong Baselines
# Marius Mosbach, Maksym Andriushchenko, Dietrich Klakow
# https://arxiv.org/abs/2006.04884
class mAdamW(Optimizer):
""" Implements Adam algorithm with weight decay fix.
Parameters:
lr (float): learning rate. Default 1e-3.
betas (tuple of 2 floats): Adams beta parameters (b1, b2). Default: (0.9, 0.999)
eps (float): Adams epsilon. Default: 1e-6
weight_decay (float): Weight decay. Default: 0.0
correct_bias (bool): can be set to False to avoid correcting bias in Adam (e.g. like in Bert TF repository). Default True.
local_normalization (bool):
max_grad_norm (bool):
"""
def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-6, weight_decay=0.0, correct_bias=True,
local_normalization=False, max_grad_norm=-1):
if lr < 0.0:
raise ValueError("Invalid learning rate: {} - should be >= 0.0".format(lr))
if not 0.0 <= betas[0] < 1.0:
raise ValueError("Invalid beta parameter: {} - should be in [0.0, 1.0[".format(betas[0]))
if not 0.0 <= betas[1] < 1.0:
raise ValueError("Invalid beta parameter: {} - should be in [0.0, 1.0[".format(betas[1]))
if not 0.0 <= eps:
raise ValueError("Invalid epsilon value: {} - should be >= 0.0".format(eps))
defaults = dict(lr=lr, betas=betas, eps=eps, weight_decay=weight_decay, correct_bias=correct_bias,
local_normalization=local_normalization, max_grad_norm=max_grad_norm)
super().__init__(params, defaults)
def step(self, closure=None):
"""Performs a single optimization step.
Arguments:
closure (callable, optional): A closure that reevaluates the model
and returns the loss.
"""
loss = None
if closure is not None:
loss = closure()
updates = []
for group in self.param_groups:
group_updates = []
for p in group["params"]:
if p.grad is None:
continue
# Normalize gradients locally (layer-wise)
if group["local_normalization"]:
torch.nn.utils.clip_grad_norm_(p, group["max_grad_norm"])
grad = p.grad.data
if grad.is_sparse:
raise RuntimeError("Adam does not support sparse gradients, please consider SparseAdam instead")
state = self.state[p]
# State initialization
if len(state) == 0:
state["step"] = 0
# Exponential moving average of gradient values
state["exp_avg"] = torch.zeros_like(p.data)
# Exponential moving average of squared gradient values
state["exp_avg_sq"] = torch.zeros_like(p.data)
exp_avg, exp_avg_sq = state["exp_avg"], state["exp_avg_sq"]
beta1, beta2 = group["betas"]
state["step"] += 1
# Decay the first and second moment running average coefficient
# In-place operations to update the averages at the same time
exp_avg.mul_(beta1).add_(grad, alpha=1.0 - beta1)
exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1.0 - beta2)
denom = exp_avg_sq.sqrt().add_(group["eps"])
step_size = group["lr"]
if group["correct_bias"]: # No bias correction for Bert
bias_correction1 = 1.0 - beta1 ** state["step"]
bias_correction2 = 1.0 - beta2 ** state["step"]
step_size = step_size * math.sqrt(bias_correction2) / bias_correction1
p.data.addcdiv_(-step_size, exp_avg, denom)
# Just adding the square of the weights to the loss function is *not*
# the correct way of using L2 regularization/weight decay with Adam,
# since that will interact with the m and v parameters in strange ways.
#
# Instead we want to decay the weights in a manner that doesn't interact
# with the m/v parameters. This is equivalent to adding the square
# of the weights to the loss with plain (non-momentum) SGD.
# Add weight decay at the end (fixed version)
if group["weight_decay"] > 0.0:
p.data.add_(p.data, alpha=-group["lr"] * group["weight_decay"])
group_updates.append((exp_avg, denom))
updates.append(group_updates)
return loss, updates |