mirror of
https://github.com/storytold/storyteller-ml.git
synced 2026-10-29 09:18:21 +00:00
add UniAnimate (see notes)
@ d7814fa44a0a1154524b92fce0e3133a2604d333 https://github.com/ali-vilab/UniAnimate
This commit is contained in:
Brandon Thomas
@@ -0,0 +1,2 @@
|
||||
from .lr_scheduler import *
|
||||
from .adafactor import *
|
||||
@@ -0,0 +1,230 @@
|
||||
import math
|
||||
import torch
|
||||
from torch.optim import Optimizer
|
||||
from torch.optim.lr_scheduler import LambdaLR
|
||||
|
||||
__all__ = ['Adafactor']
|
||||
|
||||
class Adafactor(Optimizer):
|
||||
"""
|
||||
AdaFactor pytorch implementation can be used as a drop in replacement for Adam original fairseq code:
|
||||
https://github.com/pytorch/fairseq/blob/master/fairseq/optim/adafactor.py
|
||||
Paper: *Adafactor: Adaptive Learning Rates with Sublinear Memory Cost* https://arxiv.org/abs/1804.04235 Note that
|
||||
this optimizer internally adjusts the learning rate depending on the `scale_parameter`, `relative_step` and
|
||||
`warmup_init` options. To use a manual (external) learning rate schedule you should set `scale_parameter=False` and
|
||||
`relative_step=False`.
|
||||
Arguments:
|
||||
params (`Iterable[nn.parameter.Parameter]`):
|
||||
Iterable of parameters to optimize or dictionaries defining parameter groups.
|
||||
lr (`float`, *optional*):
|
||||
The external learning rate.
|
||||
eps (`Tuple[float, float]`, *optional*, defaults to (1e-30, 1e-3)):
|
||||
Regularization constants for square gradient and parameter scale respectively
|
||||
clip_threshold (`float`, *optional*, defaults 1.0):
|
||||
Threshold of root mean square of final gradient update
|
||||
decay_rate (`float`, *optional*, defaults to -0.8):
|
||||
Coefficient used to compute running averages of square
|
||||
beta1 (`float`, *optional*):
|
||||
Coefficient used for computing running averages of gradient
|
||||
weight_decay (`float`, *optional*, defaults to 0):
|
||||
Weight decay (L2 penalty)
|
||||
scale_parameter (`bool`, *optional*, defaults to `True`):
|
||||
If True, learning rate is scaled by root mean square
|
||||
relative_step (`bool`, *optional*, defaults to `True`):
|
||||
If True, time-dependent learning rate is computed instead of external learning rate
|
||||
warmup_init (`bool`, *optional*, defaults to `False`):
|
||||
Time-dependent learning rate computation depends on whether warm-up initialization is being used
|
||||
This implementation handles low-precision (FP16, bfloat) values, but we have not thoroughly tested.
|
||||
Recommended T5 finetuning settings (https://discuss.huggingface.co/t/t5-finetuning-tips/684/3):
|
||||
- Training without LR warmup or clip_threshold is not recommended.
|
||||
- use scheduled LR warm-up to fixed LR
|
||||
- use clip_threshold=1.0 (https://arxiv.org/abs/1804.04235)
|
||||
- Disable relative updates
|
||||
- Use scale_parameter=False
|
||||
- Additional optimizer operations like gradient clipping should not be used alongside Adafactor
|
||||
Example:
|
||||
```python
|
||||
Adafactor(model.parameters(), scale_parameter=False, relative_step=False, warmup_init=False, lr=1e-3)
|
||||
```
|
||||
Others reported the following combination to work well:
|
||||
```python
|
||||
Adafactor(model.parameters(), scale_parameter=True, relative_step=True, warmup_init=True, lr=None)
|
||||
```
|
||||
When using `lr=None` with [`Trainer`] you will most likely need to use [`~optimization.AdafactorSchedule`]
|
||||
scheduler as following:
|
||||
```python
|
||||
from transformers.optimization import Adafactor, AdafactorSchedule
|
||||
optimizer = Adafactor(model.parameters(), scale_parameter=True, relative_step=True, warmup_init=True, lr=None)
|
||||
lr_scheduler = AdafactorSchedule(optimizer)
|
||||
trainer = Trainer(..., optimizers=(optimizer, lr_scheduler))
|
||||
```
|
||||
Usage:
|
||||
```python
|
||||
# replace AdamW with Adafactor
|
||||
optimizer = Adafactor(
|
||||
model.parameters(),
|
||||
lr=1e-3,
|
||||
eps=(1e-30, 1e-3),
|
||||
clip_threshold=1.0,
|
||||
decay_rate=-0.8,
|
||||
beta1=None,
|
||||
weight_decay=0.0,
|
||||
relative_step=False,
|
||||
scale_parameter=False,
|
||||
warmup_init=False,
|
||||
)
|
||||
```"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
params,
|
||||
lr=None,
|
||||
eps=(1e-30, 1e-3),
|
||||
clip_threshold=1.0,
|
||||
decay_rate=-0.8,
|
||||
beta1=None,
|
||||
weight_decay=0.0,
|
||||
scale_parameter=True,
|
||||
relative_step=True,
|
||||
warmup_init=False,
|
||||
):
|
||||
r"""require_version("torch>=1.5.0") # add_ with alpha
|
||||
"""
|
||||
if lr is not None and relative_step:
|
||||
raise ValueError("Cannot combine manual `lr` and `relative_step=True` options")
|
||||
if warmup_init and not relative_step:
|
||||
raise ValueError("`warmup_init=True` requires `relative_step=True`")
|
||||
|
||||
defaults = dict(
|
||||
lr=lr,
|
||||
eps=eps,
|
||||
clip_threshold=clip_threshold,
|
||||
decay_rate=decay_rate,
|
||||
beta1=beta1,
|
||||
weight_decay=weight_decay,
|
||||
scale_parameter=scale_parameter,
|
||||
relative_step=relative_step,
|
||||
warmup_init=warmup_init,
|
||||
)
|
||||
super().__init__(params, defaults)
|
||||
|
||||
@staticmethod
|
||||
def _get_lr(param_group, param_state):
|
||||
rel_step_sz = param_group["lr"]
|
||||
if param_group["relative_step"]:
|
||||
min_step = 1e-6 * param_state["step"] if param_group["warmup_init"] else 1e-2
|
||||
rel_step_sz = min(min_step, 1.0 / math.sqrt(param_state["step"]))
|
||||
param_scale = 1.0
|
||||
if param_group["scale_parameter"]:
|
||||
param_scale = max(param_group["eps"][1], param_state["RMS"])
|
||||
return param_scale * rel_step_sz
|
||||
|
||||
@staticmethod
|
||||
def _get_options(param_group, param_shape):
|
||||
factored = len(param_shape) >= 2
|
||||
use_first_moment = param_group["beta1"] is not None
|
||||
return factored, use_first_moment
|
||||
|
||||
@staticmethod
|
||||
def _rms(tensor):
|
||||
return tensor.norm(2) / (tensor.numel() ** 0.5)
|
||||
|
||||
@staticmethod
|
||||
def _approx_sq_grad(exp_avg_sq_row, exp_avg_sq_col):
|
||||
# copy from fairseq's adafactor implementation:
|
||||
# https://github.com/huggingface/transformers/blob/8395f14de6068012787d83989c3627c3df6a252b/src/transformers/optimization.py#L505
|
||||
r_factor = (exp_avg_sq_row / exp_avg_sq_row.mean(dim=-1, keepdim=True)).rsqrt_().unsqueeze(-1)
|
||||
c_factor = exp_avg_sq_col.unsqueeze(-2).rsqrt()
|
||||
return torch.mul(r_factor, c_factor)
|
||||
|
||||
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()
|
||||
|
||||
for group in self.param_groups:
|
||||
for p in group["params"]:
|
||||
if p.grad is None:
|
||||
continue
|
||||
grad = p.grad.data
|
||||
if grad.dtype in {torch.float16, torch.bfloat16}:
|
||||
grad = grad.float()
|
||||
if grad.is_sparse:
|
||||
raise RuntimeError("Adafactor does not support sparse gradients.")
|
||||
|
||||
state = self.state[p]
|
||||
grad_shape = grad.shape
|
||||
|
||||
factored, use_first_moment = self._get_options(group, grad_shape)
|
||||
# State Initialization
|
||||
if len(state) == 0:
|
||||
state["step"] = 0
|
||||
|
||||
if use_first_moment:
|
||||
# Exponential moving average of gradient values
|
||||
state["exp_avg"] = torch.zeros_like(grad)
|
||||
if factored:
|
||||
state["exp_avg_sq_row"] = torch.zeros(grad_shape[:-1]).to(grad)
|
||||
state["exp_avg_sq_col"] = torch.zeros(grad_shape[:-2] + grad_shape[-1:]).to(grad)
|
||||
else:
|
||||
state["exp_avg_sq"] = torch.zeros_like(grad)
|
||||
|
||||
state["RMS"] = 0
|
||||
else:
|
||||
if use_first_moment:
|
||||
state["exp_avg"] = state["exp_avg"].to(grad)
|
||||
if factored:
|
||||
state["exp_avg_sq_row"] = state["exp_avg_sq_row"].to(grad)
|
||||
state["exp_avg_sq_col"] = state["exp_avg_sq_col"].to(grad)
|
||||
else:
|
||||
state["exp_avg_sq"] = state["exp_avg_sq"].to(grad)
|
||||
|
||||
p_data_fp32 = p.data
|
||||
if p.data.dtype in {torch.float16, torch.bfloat16}:
|
||||
p_data_fp32 = p_data_fp32.float()
|
||||
|
||||
state["step"] += 1
|
||||
state["RMS"] = self._rms(p_data_fp32)
|
||||
lr = self._get_lr(group, state)
|
||||
|
||||
beta2t = 1.0 - math.pow(state["step"], group["decay_rate"])
|
||||
update = (grad**2) + group["eps"][0]
|
||||
if factored:
|
||||
exp_avg_sq_row = state["exp_avg_sq_row"]
|
||||
exp_avg_sq_col = state["exp_avg_sq_col"]
|
||||
|
||||
exp_avg_sq_row.mul_(beta2t).add_(update.mean(dim=-1), alpha=(1.0 - beta2t))
|
||||
exp_avg_sq_col.mul_(beta2t).add_(update.mean(dim=-2), alpha=(1.0 - beta2t))
|
||||
|
||||
# Approximation of exponential moving average of square of gradient
|
||||
update = self._approx_sq_grad(exp_avg_sq_row, exp_avg_sq_col)
|
||||
update.mul_(grad)
|
||||
else:
|
||||
exp_avg_sq = state["exp_avg_sq"]
|
||||
|
||||
exp_avg_sq.mul_(beta2t).add_(update, alpha=(1.0 - beta2t))
|
||||
update = exp_avg_sq.rsqrt().mul_(grad)
|
||||
|
||||
update.div_((self._rms(update) / group["clip_threshold"]).clamp_(min=1.0))
|
||||
update.mul_(lr)
|
||||
|
||||
if use_first_moment:
|
||||
exp_avg = state["exp_avg"]
|
||||
exp_avg.mul_(group["beta1"]).add_(update, alpha=(1 - group["beta1"]))
|
||||
update = exp_avg
|
||||
|
||||
if group["weight_decay"] != 0:
|
||||
p_data_fp32.add_(p_data_fp32, alpha=(-group["weight_decay"] * lr))
|
||||
|
||||
p_data_fp32.add_(-update)
|
||||
|
||||
if p.data.dtype in {torch.float16, torch.bfloat16}:
|
||||
p.data.copy_(p_data_fp32)
|
||||
|
||||
return loss
|
||||
@@ -0,0 +1,58 @@
|
||||
import math
|
||||
from torch.optim.lr_scheduler import _LRScheduler
|
||||
|
||||
__all__ = ['AnnealingLR']
|
||||
|
||||
class AnnealingLR(_LRScheduler):
|
||||
|
||||
def __init__(self, optimizer, base_lr, warmup_steps, total_steps, decay_mode='cosine', min_lr=0.0, last_step=-1):
|
||||
assert decay_mode in ['linear', 'cosine', 'none']
|
||||
self.optimizer = optimizer
|
||||
self.base_lr = base_lr
|
||||
self.warmup_steps = warmup_steps
|
||||
self.total_steps = total_steps
|
||||
self.decay_mode = decay_mode
|
||||
self.min_lr = min_lr
|
||||
self.current_step = last_step + 1
|
||||
self.step(self.current_step)
|
||||
|
||||
def get_lr(self):
|
||||
if self.warmup_steps > 0 and self.current_step <= self.warmup_steps:
|
||||
return self.base_lr * self.current_step / self.warmup_steps
|
||||
else:
|
||||
ratio = (self.current_step - self.warmup_steps) / (self.total_steps - self.warmup_steps)
|
||||
ratio = min(1.0, max(0.0, ratio))
|
||||
if self.decay_mode == 'linear':
|
||||
return self.base_lr * (1 - ratio)
|
||||
elif self.decay_mode == 'cosine':
|
||||
return self.base_lr * (math.cos(math.pi * ratio) + 1.0) / 2.0
|
||||
else:
|
||||
return self.base_lr
|
||||
|
||||
def step(self, current_step=None):
|
||||
if current_step is None:
|
||||
current_step = self.current_step + 1
|
||||
self.current_step = current_step
|
||||
new_lr = max(self.min_lr, self.get_lr())
|
||||
if isinstance(self.optimizer, list):
|
||||
for o in self.optimizer:
|
||||
for group in o.param_groups:
|
||||
group['lr'] = new_lr
|
||||
else:
|
||||
for group in self.optimizer.param_groups:
|
||||
group['lr'] = new_lr
|
||||
|
||||
def state_dict(self):
|
||||
return {
|
||||
'base_lr': self.base_lr,
|
||||
'warmup_steps': self.warmup_steps,
|
||||
'total_steps': self.total_steps,
|
||||
'decay_mode': self.decay_mode,
|
||||
'current_step': self.current_step}
|
||||
|
||||
def load_state_dict(self, state_dict):
|
||||
self.base_lr = state_dict['base_lr']
|
||||
self.warmup_steps = state_dict['warmup_steps']
|
||||
self.total_steps = state_dict['total_steps']
|
||||
self.decay_mode = state_dict['decay_mode']
|
||||
self.current_step = state_dict['current_step']
|
||||
Reference in New Issue
Block a user