Ailyth's picture
NEW-0216-050145
3c7a160
raw
history blame
26.6 kB
# Copyright 2022 Xiaomi Corp. (authors: Daniel Povey)
#
# See ../LICENSE for clarification regarding multiple authors
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import contextlib
import logging
from collections import defaultdict
from typing import List
from typing import Tuple
import torch
from torch import Tensor
from torch.optim import Optimizer
class BatchedOptimizer(Optimizer):
"""
This class adds to class Optimizer the capability to optimize parameters in batches:
it will stack the parameters and their grads for you so the optimizer can work
on tensors with an extra leading dimension. This is intended for speed with GPUs,
as it reduces the number of kernels launched in the optimizer.
Args:
params:
"""
def __init__(self, params, defaults):
super(BatchedOptimizer, self).__init__(params, defaults)
@contextlib.contextmanager
def batched_params(self, param_group, group_params_names):
"""
This function returns (technically, yields) a list of
of tuples (p, state), where
p is a `fake` parameter that is stacked (over axis 0) from real parameters
that share the same shape, and its gradient is also stacked;
`state` is the state corresponding to this batch of parameters
(it will be physically located in the "state" for one of the real
parameters, the last one that has any particular shape and dtype).
This function is decorated as a context manager so that it can
write parameters back to their "real" locations.
The idea is, instead of doing:
<code>
for p in group["params"]:
state = self.state[p]
...
</code>
you can do:
<code>
with self.batched_params(group["params"]) as batches:
for p, state, p_names in batches:
...
</code>
Args:
group: a parameter group, which is a list of parameters; should be
one of self.param_groups.
group_params_names: name for each parameter in group,
which is List[str].
"""
batches = defaultdict(
list
) # `batches` maps from tuple (dtype_as_str,*shape) to list of nn.Parameter
batches_names = defaultdict(
list
) # `batches` maps from tuple (dtype_as_str,*shape) to list of str
assert len(param_group) == len(group_params_names)
for p, named_p in zip(param_group, group_params_names):
key = (str(p.dtype), *p.shape)
batches[key].append(p)
batches_names[key].append(named_p)
batches_names_keys = list(batches_names.keys())
sorted_idx = sorted(
range(len(batches_names)), key=lambda i: batches_names_keys[i])
batches_names = [
batches_names[batches_names_keys[idx]] for idx in sorted_idx
]
batches = [batches[batches_names_keys[idx]] for idx in sorted_idx]
stacked_params_dict = dict()
# turn batches into a list, in deterministic order.
# tuples will contain tuples of (stacked_param, state, stacked_params_names),
# one for each batch in `batches`.
tuples = []
for batch, batch_names in zip(batches, batches_names):
p = batch[0]
# we arbitrarily store the state in the
# state corresponding to the 1st parameter in the
# group. class Optimizer will take care of saving/loading state.
state = self.state[p]
p_stacked = torch.stack(batch)
grad = torch.stack([
torch.zeros_like(p) if p.grad is None else p.grad for p in batch
])
p_stacked.grad = grad
stacked_params_dict[key] = p_stacked
tuples.append((p_stacked, state, batch_names))
yield tuples # <-- calling code will do the actual optimization here!
for ((stacked_params, _state, _names), batch) in zip(tuples, batches):
for i, p in enumerate(batch): # batch is list of Parameter
p.copy_(stacked_params[i])
class ScaledAdam(BatchedOptimizer):
"""
Implements 'Scaled Adam', a variant of Adam where we scale each parameter's update
proportional to the norm of that parameter; and also learn the scale of the parameter,
in log space, subject to upper and lower limits (as if we had factored each parameter as
param = underlying_param * log_scale.exp())
Args:
params: The parameters or param_groups to optimize (like other Optimizer subclasses)
lr: The learning rate. We will typically use a learning rate schedule that starts
at 0.03 and decreases over time, i.e. much higher than other common
optimizers.
clipping_scale: (e.g. 2.0)
A scale for gradient-clipping: if specified, the normalized gradients
over the whole model will be clipped to have 2-norm equal to
`clipping_scale` times the median 2-norm over the most recent period
of `clipping_update_period` minibatches. By "normalized gradients",
we mean after multiplying by the rms parameter value for this tensor
[for non-scalars]; this is appropriate because our update is scaled
by this quantity.
betas: beta1,beta2 are momentum constants for regular momentum, and moving sum-sq grad.
Must satisfy 0 < beta <= beta2 < 1.
scalar_lr_scale: A scaling factor on the learning rate, that we use to update the
scale of each parameter tensor and scalar parameters of the mode..
If each parameter were decomposed
as p * p_scale.exp(), where (p**2).mean().sqrt() == 1.0, scalar_lr_scale
would be a the scaling factor on the learning rate of p_scale.
eps: A general-purpose epsilon to prevent division by zero
param_min_rms: Minimum root-mean-square value of parameter tensor, for purposes of
learning the scale on the parameters (we'll constrain the rms of each non-scalar
parameter tensor to be >= this value)
param_max_rms: Maximum root-mean-square value of parameter tensor, for purposes of
learning the scale on the parameters (we'll constrain the rms of each non-scalar
parameter tensor to be <= this value)
scalar_max: Maximum absolute value for scalar parameters (applicable if your
model has any parameters with numel() == 1).
size_update_period: The periodicity, in steps, with which we update the size (scale)
of the parameter tensor. This is provided to save a little time
in the update.
clipping_update_period: if clipping_scale is specified, this is the period
"""
def __init__(
self,
params,
lr=3e-02,
clipping_scale=None,
betas=(0.9, 0.98),
scalar_lr_scale=0.1,
eps=1.0e-08,
param_min_rms=1.0e-05,
param_max_rms=3.0,
scalar_max=10.0,
size_update_period=4,
clipping_update_period=100,
parameters_names=None,
show_dominant_parameters=True, ):
assert parameters_names is not None, (
"Please prepare parameters_names,"
"which is a List[List[str]]. Each List[str] is for a group"
"and each str is for a parameter")
defaults = dict(
lr=lr,
clipping_scale=clipping_scale,
betas=betas,
scalar_lr_scale=scalar_lr_scale,
eps=eps,
param_min_rms=param_min_rms,
param_max_rms=param_max_rms,
scalar_max=scalar_max,
size_update_period=size_update_period,
clipping_update_period=clipping_update_period, )
super(ScaledAdam, self).__init__(params, defaults)
assert len(self.param_groups) == len(parameters_names)
self.parameters_names = parameters_names
self.show_dominant_parameters = show_dominant_parameters
def __setstate__(self, state):
super(ScaledAdam, self).__setstate__(state)
@torch.no_grad()
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:
with torch.enable_grad():
loss = closure()
batch = True
for group, group_params_names in zip(self.param_groups,
self.parameters_names):
with self.batched_params(group["params"],
group_params_names) as batches:
# batches is list of pairs (stacked_param, state). stacked_param is like
# a regular parameter, and will have a .grad, but the 1st dim corresponds to
# a stacking dim, it is not a real dim.
if (len(batches[0][1]) ==
0): # if len(first state) == 0: not yet initialized
clipping_scale = 1
else:
clipping_scale = self._get_clipping_scale(group, batches)
for p, state, _ in batches:
# Perform optimization step.
# grad is not going to be None, we handled that when creating the batches.
grad = p.grad
if grad.is_sparse:
raise RuntimeError(
"ScaledAdam optimizer does not support sparse gradients"
)
# State initialization
if len(state) == 0:
self._init_state(group, p, state)
self._step_one_batch(group, p, state, clipping_scale)
return loss
def _init_state(self, group: dict, p: Tensor, state: dict):
"""
Initializes state dict for parameter 'p'. Assumes that dim 0 of tensor p
is actually the batch dimension, corresponding to batched-together
parameters of a given shape.
Args:
group: Dict to look up configuration values.
p: The parameter that we are initializing the state for
state: Dict from string to whatever state we are initializing
"""
size_update_period = group["size_update_period"]
state["step"] = 0
kwargs = {"device": p.device, "dtype": p.dtype}
# 'delta' implements conventional momentum. There are
# several different kinds of update going on, so rather than
# compute "exp_avg" like in Adam, we store and decay a
# parameter-change "delta", which combines all forms of
# update. this is equivalent to how it's done in Adam,
# except for the first few steps.
state["delta"] = torch.zeros_like(
p, memory_format=torch.preserve_format)
batch_size = p.shape[0]
numel = p.numel() // batch_size
numel = p.numel()
if numel > 1:
# "param_rms" just periodically records the scalar root-mean-square value of
# the parameter tensor.
# it has a shape like (batch_size, 1, 1, 1, 1)
param_rms = (
(p**2).mean(dim=list(range(1, p.ndim)), keepdim=True).sqrt())
state["param_rms"] = param_rms
state["scale_exp_avg_sq"] = torch.zeros_like(param_rms)
state["scale_grads"] = torch.zeros(size_update_period,
*param_rms.shape, **kwargs)
# exp_avg_sq is the weighted sum of scaled gradients. as in Adam.
state["exp_avg_sq"] = torch.zeros_like(
p, memory_format=torch.preserve_format)
def _get_clipping_scale(self,
group: dict,
tuples: List[Tuple[Tensor, dict, List[str]]]
) -> float:
"""
Returns a scalar factor <= 1.0 that dictates gradient clipping, i.e. we will scale the gradients
by this amount before applying the rest of the update.
Args:
group: the parameter group, an item in self.param_groups
tuples: a list of tuples of (param, state, param_names)
where param is a batched set of parameters,
with a .grad (1st dim is batch dim)
and state is the state-dict where optimization parameters are kept.
param_names is a List[str] while each str is name for a parameter
in batched set of parameters "param".
"""
assert len(tuples) >= 1
clipping_scale = group["clipping_scale"]
(first_p, first_state, _) = tuples[0]
step = first_state["step"]
if clipping_scale is None or step == 0:
# no clipping. return early on step == 0 because the other
# parameters' state won't have been initialized yet.
return 1.0
clipping_update_period = group["clipping_update_period"]
tot_sumsq = torch.tensor(0.0, device=first_p.device)
for (p, state, param_names) in tuples:
grad = p.grad
if grad.is_sparse:
raise RuntimeError(
"ScaledAdam optimizer does not support sparse gradients")
if p.numel() == p.shape[0]: # a batch of scalars
tot_sumsq += (grad**2).sum() # sum() to change shape [1] to []
else:
tot_sumsq += ((grad * state["param_rms"])**2).sum()
tot_norm = tot_sumsq.sqrt()
if "model_norms" not in first_state:
first_state["model_norms"] = torch.zeros(
clipping_update_period, device=p.device)
first_state["model_norms"][step % clipping_update_period] = tot_norm
if step % clipping_update_period == 0:
# Print some stats.
# We don't reach here if step == 0 because we would have returned
# above.
sorted_norms = first_state["model_norms"].sort()[0].to("cpu")
quartiles = []
for n in range(0, 5):
index = min(
clipping_update_period - 1,
(clipping_update_period // 4) * n, )
quartiles.append(sorted_norms[index].item())
median = quartiles[2]
threshold = clipping_scale * median
first_state["model_norm_threshold"] = threshold
percent_clipped = (first_state["num_clipped"] * 100.0 /
clipping_update_period
if "num_clipped" in first_state else 0.0)
first_state["num_clipped"] = 0
quartiles = " ".join(["%.3e" % x for x in quartiles])
logging.info(
f"Clipping_scale={clipping_scale}, grad-norm quartiles {quartiles}, "
f"threshold={threshold:.3e}, percent-clipped={percent_clipped:.1f}"
)
if step < clipping_update_period:
return 1.0 # We have not yet estimated a norm to clip to.
else:
try:
model_norm_threshold = first_state["model_norm_threshold"]
except KeyError:
logging.info(
"Warning: model_norm_threshold not in state: possibly "
"you changed config when restarting, adding clipping_scale option?"
)
return 1.0
ans = min(1.0, (model_norm_threshold / (tot_norm + 1.0e-20)).item())
if ans < 1.0:
first_state["num_clipped"] += 1
if ans < 0.1:
logging.warn(
f"Scaling gradients by {ans}, model_norm_threshold={model_norm_threshold}"
)
if self.show_dominant_parameters:
assert p.shape[0] == len(param_names)
self._show_gradient_dominating_parameter(tuples, tot_sumsq)
return ans
def _show_gradient_dominating_parameter(
self, tuples: List[Tuple[Tensor, dict, List[str]]],
tot_sumsq: Tensor):
"""
Show information of parameter wihch dominanting tot_sumsq.
Args:
tuples: a list of tuples of (param, state, param_names)
where param is a batched set of parameters,
with a .grad (1st dim is batch dim)
and state is the state-dict where optimization parameters are kept.
param_names is a List[str] while each str is name for a parameter
in batched set of parameters "param".
tot_sumsq: sumsq of all parameters. Though it's could be calculated
from tuples, we still pass it to save some time.
"""
all_sumsq_orig = {}
for (p, state, batch_param_names) in tuples:
# p is a stacked batch parameters.
batch_grad = p.grad
if p.numel() == p.shape[0]: # a batch of scalars
batch_sumsq_orig = batch_grad**2
# Dummpy values used by following `zip` statement.
batch_rms_orig = torch.ones(p.shape[0])
else:
batch_rms_orig = state["param_rms"]
batch_sumsq_orig = ((batch_grad * batch_rms_orig)**2).sum(
dim=list(range(1, batch_grad.ndim)))
for name, sumsq_orig, rms, grad in zip(batch_param_names,
batch_sumsq_orig,
batch_rms_orig, batch_grad):
proportion_orig = sumsq_orig / tot_sumsq
all_sumsq_orig[name] = (proportion_orig, sumsq_orig, rms, grad)
assert torch.isclose(
sum([value[0] for value in all_sumsq_orig.values()]).cpu(),
torch.tensor(1.0), )
sorted_by_proportion = {
k: v
for k, v in sorted(
all_sumsq_orig.items(),
key=lambda item: item[1][0],
reverse=True, )
}
dominant_param_name = next(iter(sorted_by_proportion))
(dominant_proportion, dominant_sumsq, dominant_rms,
dominant_grad, ) = sorted_by_proportion[dominant_param_name]
logging.info(f"Parameter Dominanting tot_sumsq {dominant_param_name}"
f" with proportion {dominant_proportion:.2f},"
f" where dominant_sumsq=(grad_sumsq*orig_rms_sq)"
f"={dominant_sumsq:.3e},"
f" grad_sumsq = {(dominant_grad**2).sum():.3e},"
f" orig_rms_sq={(dominant_rms**2).item():.3e}")
def _step_one_batch(self,
group: dict,
p: Tensor,
state: dict,
clipping_scale: float):
"""
Do the step for one parameter, which is actually going to be a batch of
`real` parameters, with dim 0 as the batch dim.
Args:
group: dict to look up configuration values
p: parameter to update (actually multiple parameters stacked together
as a batch)
state: state-dict for p, to look up the optimizer state
"""
lr = group["lr"]
size_update_period = group["size_update_period"]
beta1 = group["betas"][0]
grad = p.grad
if clipping_scale != 1.0:
grad = grad * clipping_scale
step = state["step"]
delta = state["delta"]
delta.mul_(beta1)
batch_size = p.shape[0]
numel = p.numel() // batch_size
if numel > 1:
# Update the size/scale of p, and set param_rms
scale_grads = state["scale_grads"]
scale_grads[step % size_update_period] = (p * grad).sum(
dim=list(range(1, p.ndim)), keepdim=True)
if step % size_update_period == size_update_period - 1:
param_rms = state["param_rms"] # shape: (batch_size, 1, 1, ..)
param_rms.copy_((p**2)
.mean(dim=list(range(1, p.ndim)), keepdim=True)
.sqrt())
if step > 0:
# self._size_update() learns the overall scale on the
# parameter, by shrinking or expanding it.
self._size_update(group, scale_grads, p, state)
if numel == 1:
# For parameters with 1 element we just use regular Adam.
# Updates delta.
self._step_scalar(group, p, state)
else:
self._step(group, p, state)
state["step"] = step + 1
def _size_update(self,
group: dict,
scale_grads: Tensor,
p: Tensor,
state: dict) -> None:
"""
Called only where p.numel() > 1, this updates the scale of the parameter.
If we imagine: p = underlying_param * scale.exp(), and we are doing
gradient descent on underlying param and on scale, this function does the update
on `scale`.
Args:
group: dict to look up configuration values
scale_grads: a tensor of shape (size_update_period, batch_size, 1, 1,...) containing
grads w.r.t. the scales.
p: The parameter to update
state: The state-dict of p
"""
param_rms = state["param_rms"]
beta1, beta2 = group["betas"]
size_lr = group["lr"] * group["scalar_lr_scale"]
param_min_rms = group["param_min_rms"]
param_max_rms = group["param_max_rms"]
eps = group["eps"]
step = state["step"]
batch_size = p.shape[0]
size_update_period = scale_grads.shape[0]
# correct beta2 for the size update period: we will have
# faster decay at this level.
beta2_corr = beta2**size_update_period
scale_exp_avg_sq = state[
"scale_exp_avg_sq"] # shape: (batch_size, 1, 1, ..)
scale_exp_avg_sq.mul_(beta2_corr).add_(
(scale_grads**2).mean(dim=0), # mean over dim `size_update_period`
alpha=1 - beta2_corr, ) # shape is (batch_size, 1, 1, ...)
# The 1st time we reach here is when size_step == 1.
size_step = (step + 1) // size_update_period
bias_correction2 = 1 - beta2_corr**size_step
# we don't bother with bias_correction1; this will help prevent divergence
# at the start of training.
denom = scale_exp_avg_sq.sqrt() + eps
scale_step = (-size_lr * (bias_correction2**0.5) *
scale_grads.sum(dim=0) / denom)
is_too_small = param_rms < param_min_rms
is_too_large = param_rms > param_max_rms
# when the param gets too small, just don't shrink it any further.
scale_step.masked_fill_(is_too_small, 0.0)
# when it gets too large, stop it from getting any larger.
scale_step.masked_fill_(is_too_large, -size_lr * size_update_period)
delta = state["delta"]
# the factor of (1-beta1) relates to momentum.
delta.add_(p * scale_step, alpha=(1 - beta1))
def _step(self, group: dict, p: Tensor, state: dict):
"""
This function does the core update of self.step(), in the case where the members of
the batch have more than 1 element.
Args:
group: A dict which will be used to look up configuration values
p: The parameter to be updated
grad: The grad of p
state: The state-dict corresponding to parameter p
This function modifies p.
"""
grad = p.grad
lr = group["lr"]
beta1, beta2 = group["betas"]
eps = group["eps"]
param_min_rms = group["param_min_rms"]
step = state["step"]
exp_avg_sq = state["exp_avg_sq"]
exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=(1 - beta2))
this_step = state["step"] - (state["zero_step"]
if "zero_step" in state else 0)
bias_correction2 = 1 - beta2**(this_step + 1)
if bias_correction2 < 0.99:
# note: not in-place.
exp_avg_sq = exp_avg_sq * (1.0 / bias_correction2)
denom = exp_avg_sq.sqrt()
denom += eps
grad = grad / denom
alpha = -lr * (1 - beta1) * state["param_rms"].clamp(min=param_min_rms)
delta = state["delta"]
delta.add_(grad * alpha)
p.add_(delta)
def _step_scalar(self, group: dict, p: Tensor, state: dict):
"""
A simplified form of the core update for scalar tensors, where we cannot get a good
estimate of the parameter rms.
"""
beta1, beta2 = group["betas"]
scalar_max = group["scalar_max"]
eps = group["eps"]
lr = group["lr"] * group["scalar_lr_scale"]
grad = p.grad
exp_avg_sq = state["exp_avg_sq"] # shape: (batch_size,)
exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
# bias_correction2 is like in Adam. Don't bother with bias_correction1;
# slower update at the start will help stability anyway.
bias_correction2 = 1 - beta2**(state["step"] + 1)
denom = (exp_avg_sq / bias_correction2).sqrt() + eps
delta = state["delta"]
delta.add_(grad / denom, alpha=-lr * (1 - beta1))
p.clamp_(min=-scalar_max, max=scalar_max)
p.add_(delta)