Han Zhu 48088cb807
Refactor optimizer (#1837)
* Print indexes of largest grad
2024-12-30 15:30:02 +08:00

1238 lines
49 KiB
Python

# 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
import random
from collections import defaultdict
from typing import Dict, List, Optional, Tuple, Union
import torch
from lhotse.utils import fix_random_seed
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])
def basic_step(group, p, state, grad):
# computes basic Adam update using beta2 (dividing by gradient stddev) only. no momentum yet.
lr = group["lr"]
if p.numel() == p.shape[0]:
lr = lr * group["scalar_lr_scale"]
beta2 = group["betas"][1]
eps = group["eps"]
# p shape: (batch_size,) or (batch_size, 1, [1,..])
try:
exp_avg_sq = state[
"exp_avg_sq"
] # shape: (batch_size,) or (batch_size, 1, [1,..])
except KeyError:
exp_avg_sq = torch.zeros(*p.shape, device=p.device, dtype=torch.float)
state["exp_avg_sq"] = exp_avg_sq
exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
# bias_correction2 is like in Adam.
# slower update at the start will help stability anyway.
bias_correction2 = 1 - beta2 ** (state["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().add_(eps)
return -lr * grad / denom
def scaling_step(group, p, state, grad):
delta = basic_step(group, p, state, grad)
if p.numel() == p.shape[0]:
return delta # there is no scaling for scalar parameters. (p.shape[0] is the batch of parameters.)
step = state["step"]
size_update_period = group["size_update_period"]
try:
param_rms = state["param_rms"]
scale_grads = state["scale_grads"]
scale_exp_avg_sq = state["scale_exp_avg_sq"]
except KeyError:
# we know p.ndim > 1 because we'd have returned above if not, so don't worry
# about the speial case of dim=[] that pytorch treats inconsistently.
param_rms = (p**2).mean(dim=list(range(1, p.ndim)), keepdim=True).sqrt()
param_rms = param_rms.to(torch.float)
scale_exp_avg_sq = torch.zeros_like(param_rms)
scale_grads = torch.zeros(
size_update_period, *param_rms.shape, dtype=torch.float, device=p.device
)
state["param_rms"] = param_rms
state["scale_grads"] = scale_grads
state["scale_exp_avg_sq"] = scale_exp_avg_sq
# on every step, update the gradient w.r.t. the scale of the parameter, we
# store these as a batch and periodically update the size (for speed only, to
# avoid too many operations).
scale_grads[step % size_update_period] = (p * grad).sum(
dim=list(range(1, p.ndim)), keepdim=True
)
# periodically recompute the value of param_rms.
if step % size_update_period == size_update_period - 1:
param_rms.copy_((p**2).mean(dim=list(range(1, p.ndim)), keepdim=True).sqrt())
param_min_rms = group["param_min_rms"]
# scale the step size by param_rms. This is the most important "scaling" part of
# ScaledAdam
delta *= param_rms.clamp(min=param_min_rms)
if step % size_update_period == size_update_period - 1 and step > 0:
# This block updates the size of parameter by adding a step ("delta") value in
# the direction of either shrinking or growing it.
beta2 = group["betas"][1]
size_lr = group["lr"] * group["scalar_lr_scale"]
param_max_rms = group["param_max_rms"]
eps = group["eps"]
batch_size = p.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.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
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
# when the param gets too small, just don't shrink it any further.
scale_step.masked_fill_(is_too_small, 0.0)
# The following may help prevent instability: don't allow the scale step to be too large in
# either direction.
scale_step.clamp_(min=-0.1, max=0.1)
# and ensure the parameter rms after update never exceeds param_max_rms.
# We have to look at the trained model for parameters at or around the
# param_max_rms, because sometimes they can indicate a problem with the
# topology or settings.
scale_step = torch.minimum(scale_step, (param_max_rms - param_rms) / param_rms)
delta.add_(p * scale_step)
return delta
def momentum_step(group, p, state, grad):
delta = scaling_step(group, p, state, grad)
beta1 = group["betas"][0]
try:
stored_delta = state["delta"]
except KeyError:
stored_delta = torch.zeros(*p.shape, device=p.device, dtype=torch.float)
state["delta"] = stored_delta
stored_delta.mul_(beta1)
stored_delta.add_(delta, alpha=(1 - beta1))
# we don't bother doing the "bias correction" part of Adam for beta1 because this is just
# an edge effect that affects the first 10 or so batches; and the effect of not doing it
# is just to do a slower update for the first few batches, which will help stability.
return stored_delta
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)
Unlike common optimizers, which accept model.parameters() or groups of parameters(),
this optimizer could accept model.named_parameters() or groups of named_parameters().
See comments of function _get_names_of_parameters for its 4 possible cases.
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,
):
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,
)
# If params only contains parameters or group of parameters,
# i.e when parameter names are not given,
# this flag will be set to False in funciton _get_names_of_parameters.
self.show_dominant_parameters = True
param_groups, parameters_names = self._get_names_of_parameters(params)
super(ScaledAdam, self).__init__(param_groups, defaults)
assert len(self.param_groups) == len(parameters_names)
self.parameters_names = parameters_names
def _get_names_of_parameters(
self, params_or_named_params
) -> Tuple[List[Dict], List[List[str]]]:
"""
Args:
params_or_named_params: according to the way ScaledAdam is initialized in train.py,
this argument could be one of following 4 cases,
case 1, a generator of parameter, e.g.:
optimizer = ScaledAdam(model.parameters(), lr=params.base_lr, clipping_scale=3.0)
case 2, a list of parameter groups with different config, e.g.:
model_param_groups = [
{'params': model.encoder.parameters(), 'lr': 0.05},
{'params': model.decoder.parameters(), 'lr': 0.01},
{'params': model.joiner.parameters(), 'lr': 0.03},
]
optimizer = ScaledAdam(model_param_groups, lr=params.base_lr, clipping_scale=3.0)
case 3, a generator of named_parameter, e.g.:
optimizer = ScaledAdam(model.named_parameters(), lr=params.base_lr, clipping_scale=3.0)
case 4, a list of named_parameter groups with different config, e.g.:
model_named_param_groups = [
{'named_params': model.encoder.named_parameters(), 'lr': 0.05},
{'named_params': model.decoder.named_parameters(), 'lr': 0.01},
{'named_params': model.joiner.named_parameters(), 'lr': 0.03},
]
optimizer = ScaledAdam(model_named_param_groups, lr=params.base_lr, clipping_scale=3.0)
For case 1 and case 2, input params is used to initialize the underlying torch.optimizer.
For case 3 and case 4, firstly, names and params are extracted from input named_params,
then, these extracted params are used to initialize the underlying torch.optimizer,
and these extracted names are mainly used by function
`_show_gradient_dominating_parameter`
Returns:
Returns a tuple containing 2 elements:
- `param_groups` with type List[Dict], each Dict element is a parameter group.
An example of `param_groups` could be:
[
{'params': `one iterable of Parameter`, 'lr': 0.05},
{'params': `another iterable of Parameter`, 'lr': 0.08},
{'params': `a third iterable of Parameter`, 'lr': 0.1},
]
- `param_gruops_names` with type List[List[str]],
each `List[str]` is for a group['params'] in param_groups,
and each `str` is the name of a parameter.
A dummy name "foo" is related to each parameter,
if input are params without names, i.e. case 1 or case 2.
"""
# variable naming convention in this function:
# p is short for param.
# np is short for named_param.
# p_or_np is short for param_or_named_param.
# cur is short for current.
# group is a dict, e.g. {'params': iterable of parameter, 'lr': 0.05, other fields}.
# groups is a List[group]
iterable_or_groups = list(params_or_named_params)
if len(iterable_or_groups) == 0:
raise ValueError("optimizer got an empty parameter list")
# The first value of returned tuple. A list of dicts containing at
# least 'params' as a key.
param_groups = []
# The second value of returned tuple,
# a List[List[str]], each sub-List is for a group.
param_groups_names = []
if not isinstance(iterable_or_groups[0], dict):
# case 1 or case 3,
# the input is an iterable of parameter or named parameter.
param_iterable_cur_group = []
param_names_cur_group = []
for p_or_np in iterable_or_groups:
if isinstance(p_or_np, tuple):
# case 3
name, param = p_or_np
else:
# case 1
assert isinstance(p_or_np, torch.Tensor)
param = p_or_np
# Assign a dummy name as a placeholder
name = "foo"
self.show_dominant_parameters = False
param_iterable_cur_group.append(param)
param_names_cur_group.append(name)
param_groups.append({"params": param_iterable_cur_group})
param_groups_names.append(param_names_cur_group)
else:
# case 2 or case 4
# the input is groups of parameter or named parameter.
for cur_group in iterable_or_groups:
if "named_params" in cur_group:
name_list = [x[0] for x in cur_group["named_params"]]
p_list = [x[1] for x in cur_group["named_params"]]
del cur_group["named_params"]
cur_group["params"] = p_list
else:
assert "params" in cur_group
name_list = ["foo" for _ in cur_group["params"]]
param_groups.append(cur_group)
param_groups_names.append(name_list)
return param_groups, param_groups_names
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"
)
try:
cur_step = state["step"]
except KeyError:
state["step"] = 0
cur_step = 0
grad = (
p.grad if clipping_scale == 1.0 else p.grad.mul_(clipping_scale)
)
p += momentum_step(group, p.detach(), state, grad)
if p.numel() == p.shape[0]: # scalar parameter
scalar_max = group["scalar_max"]
p.clamp_(min=-scalar_max, max=scalar_max)
state["step"] = cur_step + 1
return loss
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"]
scalar_lr_scale = group["scalar_lr_scale"]
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() * (
scalar_lr_scale**2
) # 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
irregular_estimate_steps = [
i for i in [10, 20, 40] if i < clipping_update_period
]
if step % clipping_update_period == 0 or step in irregular_estimate_steps:
# 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")
if step in irregular_estimate_steps:
sorted_norms = sorted_norms[-step:]
num_norms = sorted_norms.numel()
quartiles = []
for n in range(0, 5):
index = min(num_norms - 1, (num_norms // 4) * n)
quartiles.append(sorted_norms[index].item())
median = quartiles[2]
if median - median != 0:
raise RuntimeError("Too many grads were not finite")
threshold = clipping_scale * median
if step in irregular_estimate_steps:
# use larger thresholds on first few steps of estimating threshold,
# as norm may be changing rapidly.
threshold = threshold * 2.0
first_state["model_norm_threshold"] = threshold
percent_clipped = (
first_state["num_clipped"] * 100.0 / num_norms
if "num_clipped" in first_state
else 0.0
)
first_state["num_clipped"] = 0
quartiles = " ".join(["%.3e" % x for x in quartiles])
logging.warning(
f"Clipping_scale={clipping_scale}, grad-norm quartiles {quartiles}, "
f"threshold={threshold:.3e}, percent-clipped={percent_clipped:.1f}"
)
try:
model_norm_threshold = first_state["model_norm_threshold"]
except KeyError:
return 1.0 # threshold has not yet been set.
ans = min(1.0, (model_norm_threshold / (tot_norm + 1.0e-20)).item())
if ans != ans: # e.g. ans is nan
ans = 0.0
if ans < 1.0:
first_state["num_clipped"] += 1
if ans < 0.5:
logging.warning(
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, group["scalar_lr_scale"]
)
self._show_param_with_unusual_grad(tuples)
if ans == 0.0:
for (p, state, param_names) in tuples:
p.grad.zero_() # get rid of infinity()
return ans
def _show_param_with_unusual_grad(
self,
tuples: List[Tuple[Tensor, dict, List[str]]],
):
"""
Print information about parameter which has the largest ratio of grad-on-this-batch
divided by normal grad size.
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".
"""
largest_ratio = 0.0
largest_name = ""
# ratios_names is a list of 3-tuples: (grad_ratio, param_name, tensor)
ratios_names = []
for (p, state, batch_param_names) in tuples:
dims = list(range(1, p.ndim))
def mean(x):
# workaround for bad interface of torch's "mean" for when dims is the empty list.
if len(dims) > 0:
return x.mean(dim=dims)
else:
return x
grad_ratio = (
(mean(p.grad**2) / state["exp_avg_sq"].mean(dim=dims))
.sqrt()
.to("cpu")
)
ratios_names += zip(
grad_ratio.tolist(), batch_param_names, p.grad.unbind(dim=0)
)
ratios_names = sorted(ratios_names, reverse=True)
ratios_names = ratios_names[:10]
ratios_names = [
(ratio, name, largest_index(tensor))
for (ratio, name, tensor) in ratios_names
]
logging.warning(
f"Parameters with most larger-than-usual grads, with ratios, are: {ratios_names}"
)
def _show_gradient_dominating_parameter(
self,
tuples: List[Tuple[Tensor, dict, List[str]]],
tot_sumsq: Tensor,
scalar_lr_scale: float,
):
"""
Show information of parameter which dominates 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
# Dummy values used by following `zip` statement.
batch_rms_orig = torch.full(
p.shape, scalar_lr_scale, device=batch_grad.device
)
else:
batch_rms_orig = state["param_rms"]
batch_sumsq_orig = (batch_grad * batch_rms_orig) ** 2
if batch_grad.ndim > 1:
# need to guard it with if-statement because sum() sums over
# all dims if dim == ().
batch_sumsq_orig = batch_sumsq_orig.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)
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.warning(
f"Parameter dominating 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 largest_index(x: Tensor):
x = x.contiguous()
argmax = x.abs().argmax().item()
return [(argmax // x.stride(i)) % x.size(i) for i in range(x.ndim)]
class LRScheduler(object):
"""
Base-class for learning rate schedulers where the learning-rate depends on both the
batch and the epoch.
"""
def __init__(self, optimizer: Optimizer, verbose: bool = False):
# Attach optimizer
if not isinstance(optimizer, Optimizer):
raise TypeError("{} is not an Optimizer".format(type(optimizer).__name__))
self.optimizer = optimizer
self.verbose = verbose
for group in optimizer.param_groups:
group.setdefault("base_lr", group["lr"])
self.base_lrs = [group["base_lr"] for group in optimizer.param_groups]
self.epoch = 0
self.batch = 0
def state_dict(self):
"""Returns the state of the scheduler as a :class:`dict`.
It contains an entry for every variable in self.__dict__ which
is not the optimizer.
"""
return {
# the user might try to override the base_lr, so don't include this in the state.
# previously they were included.
# "base_lrs": self.base_lrs,
"epoch": self.epoch,
"batch": self.batch,
}
def load_state_dict(self, state_dict):
"""Loads the schedulers state.
Args:
state_dict (dict): scheduler state. Should be an object returned
from a call to :meth:`state_dict`.
"""
# the things with base_lrs are a work-around for a previous problem
# where base_lrs were written with the state dict.
base_lrs = self.base_lrs
self.__dict__.update(state_dict)
self.base_lrs = base_lrs
def get_last_lr(self) -> List[float]:
"""Return last computed learning rate by current scheduler. Will be a list of float."""
return self._last_lr
def get_lr(self):
# Compute list of learning rates from self.epoch and self.batch and
# self.base_lrs; this must be overloaded by the user.
# e.g. return [some_formula(self.batch, self.epoch, base_lr) for base_lr in self.base_lrs ]
raise NotImplementedError
def step_batch(self, batch: Optional[int] = None) -> None:
# Step the batch index, or just set it. If `batch` is specified, it
# must be the batch index from the start of training, i.e. summed over
# all epochs.
# You can call this in any order; if you don't provide 'batch', it should
# of course be called once per batch.
if batch is not None:
self.batch = batch
else:
self.batch = self.batch + 1
self._set_lrs()
def step_epoch(self, epoch: Optional[int] = None):
# Step the epoch index, or just set it. If you provide the 'epoch' arg,
# you should call this at the start of the epoch; if you don't provide the 'epoch'
# arg, you should call it at the end of the epoch.
if epoch is not None:
self.epoch = epoch
else:
self.epoch = self.epoch + 1
self._set_lrs()
def _set_lrs(self):
values = self.get_lr()
assert len(values) == len(self.optimizer.param_groups)
for i, data in enumerate(zip(self.optimizer.param_groups, values)):
param_group, lr = data
param_group["lr"] = lr
self.print_lr(self.verbose, i, lr)
self._last_lr = [group["lr"] for group in self.optimizer.param_groups]
def print_lr(self, is_verbose, group, lr):
"""Display the current learning rate."""
if is_verbose:
logging.warning(
f"Epoch={self.epoch}, batch={self.batch}: adjusting learning rate"
f" of group {group} to {lr:.4e}."
)
class Eden(LRScheduler):
"""
Eden scheduler.
The basic formula (before warmup) is:
lr = base_lr * (((batch**2 + lr_batches**2) / lr_batches**2) ** -0.25 *
(((epoch**2 + lr_epochs**2) / lr_epochs**2) ** -0.25)) * warmup
where `warmup` increases from linearly 0.5 to 1 over `warmup_batches` batches
and then stays constant at 1.
If you don't have the concept of epochs, or one epoch takes a very long time,
you can replace the notion of 'epoch' with some measure of the amount of data
processed, e.g. hours of data or frames of data, with 'lr_epochs' being set to
some measure representing "quite a lot of data": say, one fifth or one third
of an entire training run, but it doesn't matter much. You could also use
Eden2 which has only the notion of batches.
We suggest base_lr = 0.04 (passed to optimizer) if used with ScaledAdam
Args:
optimizer: the optimizer to change the learning rates on
lr_batches: the number of batches after which we start significantly
decreasing the learning rate, suggest 5000.
lr_epochs: the number of epochs after which we start significantly
decreasing the learning rate, suggest 6 if you plan to do e.g.
20 to 40 epochs, but may need smaller number if dataset is huge
and you will do few epochs.
"""
def __init__(
self,
optimizer: Optimizer,
lr_batches: Union[int, float],
lr_epochs: Union[int, float],
warmup_batches: Union[int, float] = 500.0,
warmup_start: float = 0.5,
verbose: bool = False,
):
super(Eden, self).__init__(optimizer, verbose)
self.lr_batches = lr_batches
self.lr_epochs = lr_epochs
self.warmup_batches = warmup_batches
assert 0.0 <= warmup_start <= 1.0, warmup_start
self.warmup_start = warmup_start
def get_lr(self):
factor = (
(self.batch**2 + self.lr_batches**2) / self.lr_batches**2
) ** -0.25 * (
((self.epoch**2 + self.lr_epochs**2) / self.lr_epochs**2) ** -0.25
)
warmup_factor = (
1.0
if self.batch >= self.warmup_batches
else self.warmup_start
+ (1.0 - self.warmup_start) * (self.batch / self.warmup_batches)
# else 0.5 + 0.5 * (self.batch / self.warmup_batches)
)
return [x * factor * warmup_factor for x in self.base_lrs]
class Eden2(LRScheduler):
"""
Eden2 scheduler, simpler than Eden because it does not use the notion of epoch,
only batches.
The basic formula (before warmup) is:
lr = base_lr * ((batch**2 + lr_batches**2) / lr_batches**2) ** -0.5) * warmup
where `warmup` increases from linearly 0.5 to 1 over `warmup_batches` batches
and then stays constant at 1.
E.g. suggest base_lr = 0.04 (passed to optimizer) if used with ScaledAdam
Args:
optimizer: the optimizer to change the learning rates on
lr_batches: the number of batches after which we start significantly
decreasing the learning rate, suggest 5000.
"""
def __init__(
self,
optimizer: Optimizer,
lr_batches: Union[int, float],
warmup_batches: Union[int, float] = 500.0,
warmup_start: float = 0.5,
verbose: bool = False,
):
super().__init__(optimizer, verbose)
self.lr_batches = lr_batches
self.warmup_batches = warmup_batches
assert 0.0 <= warmup_start <= 1.0, warmup_start
self.warmup_start = warmup_start
def get_lr(self):
factor = (
(self.batch**2 + self.lr_batches**2) / self.lr_batches**2
) ** -0.5
warmup_factor = (
1.0
if self.batch >= self.warmup_batches
else self.warmup_start
+ (1.0 - self.warmup_start) * (self.batch / self.warmup_batches)
# else 0.5 + 0.5 * (self.batch / self.warmup_batches)
)
return [x * factor * warmup_factor for x in self.base_lrs]
def _test_eden():
m = torch.nn.Linear(100, 100)
optim = ScaledAdam(m.parameters(), lr=0.03)
scheduler = Eden(optim, lr_batches=100, lr_epochs=2, verbose=True)
for epoch in range(10):
scheduler.step_epoch(epoch) # sets epoch to `epoch`
for step in range(20):
x = torch.randn(200, 100).detach()
x.requires_grad = True
y = m(x)
dy = torch.randn(200, 100).detach()
f = (y * dy).sum()
f.backward()
optim.step()
scheduler.step_batch()
optim.zero_grad()
logging.info(f"last lr = {scheduler.get_last_lr()}")
logging.info(f"state dict = {scheduler.state_dict()}")
# This is included mostly as a baseline for ScaledAdam.
class Eve(Optimizer):
"""
Implements Eve algorithm. This is a modified version of AdamW with a special
way of setting the weight-decay / shrinkage-factor, which is designed to make the
rms of the parameters approach a particular target_rms (default: 0.1). This is
for use with networks with 'scaled' versions of modules (see scaling.py), which
will be close to invariant to the absolute scale on the parameter matrix.
The original Adam algorithm was proposed in `Adam: A Method for Stochastic Optimization`_.
The AdamW variant was proposed in `Decoupled Weight Decay Regularization`_.
Eve is unpublished so far.
Arguments:
params (iterable): iterable of parameters to optimize or dicts defining
parameter groups
lr (float, optional): learning rate (default: 1e-3)
betas (Tuple[float, float], optional): coefficients used for computing
running averages of gradient and its square (default: (0.9, 0.999))
eps (float, optional): term added to the denominator to improve
numerical stability (default: 1e-8)
weight_decay (float, optional): weight decay coefficient (default: 3e-4;
this value means that the weight would decay significantly after
about 3k minibatches. Is not multiplied by learning rate, but
is conditional on RMS-value of parameter being > target_rms.
target_rms (float, optional): target root-mean-square value of
parameters, if they fall below this we will stop applying weight decay.
.. _Adam: A Method for Stochastic Optimization:
https://arxiv.org/abs/1412.6980
.. _Decoupled Weight Decay Regularization:
https://arxiv.org/abs/1711.05101
.. _On the Convergence of Adam and Beyond:
https://openreview.net/forum?id=ryQu7f-RZ
"""
def __init__(
self,
params,
lr=1e-3,
betas=(0.9, 0.98),
eps=1e-8,
weight_decay=1e-3,
target_rms=0.1,
):
if not 0.0 <= lr:
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]))
if not 0 <= weight_decay <= 0.1:
raise ValueError("Invalid weight_decay value: {}".format(weight_decay))
if not 0 < target_rms <= 10.0:
raise ValueError("Invalid target_rms value: {}".format(target_rms))
defaults = dict(
lr=lr,
betas=betas,
eps=eps,
weight_decay=weight_decay,
target_rms=target_rms,
)
super(Eve, self).__init__(params, defaults)
def __setstate__(self, state):
super(Eve, 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()
for group in self.param_groups:
for p in group["params"]:
if p.grad is None:
continue
# Perform optimization step
grad = p.grad
if grad.is_sparse:
raise RuntimeError("AdamW does not support sparse gradients")
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, memory_format=torch.preserve_format
)
# Exponential moving average of squared gradient values
state["exp_avg_sq"] = torch.zeros_like(
p, memory_format=torch.preserve_format
)
exp_avg, exp_avg_sq = state["exp_avg"], state["exp_avg_sq"]
beta1, beta2 = group["betas"]
state["step"] += 1
bias_correction1 = 1 - beta1 ** state["step"]
bias_correction2 = 1 - beta2 ** state["step"]
# Decay the first and second moment running average coefficient
exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1)
exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
denom = (exp_avg_sq.sqrt() * (bias_correction2**-0.5)).add_(
group["eps"]
)
step_size = group["lr"] / bias_correction1
target_rms = group["target_rms"]
weight_decay = group["weight_decay"]
if p.numel() > 1:
# avoid applying this weight-decay on "scaling factors"
# (which are scalar).
is_above_target_rms = p.norm() > (target_rms * (p.numel() ** 0.5))
p.mul_(1 - (weight_decay * is_above_target_rms))
p.addcdiv_(exp_avg, denom, value=-step_size)
if random.random() < 0.0005:
step = (exp_avg / denom) * step_size
logging.info(
f"Delta rms = {(step**2).mean().item()}, shape = {step.shape}"
)
return loss
def _test_scaled_adam(hidden_dim: int):
import timeit
from scaling import ScaledLinear
E = 100
B = 4
T = 2
logging.info("in test_eve_cain")
# device = torch.device('cuda')
device = torch.device("cpu")
dtype = torch.float32
fix_random_seed(42)
# these input_magnitudes and output_magnitudes are to test that
# Abel is working as we expect and is able to adjust scales of
# different dims differently.
input_magnitudes = (1.0 * torch.randn(E, dtype=dtype, device=device)).exp()
output_magnitudes = (1.0 * torch.randn(E, dtype=dtype, device=device)).exp()
for iter in [1, 0]:
fix_random_seed(42)
Linear = torch.nn.Linear if iter == 0 else ScaledLinear
m = torch.nn.Sequential(
Linear(E, hidden_dim),
torch.nn.PReLU(),
Linear(hidden_dim, hidden_dim),
torch.nn.PReLU(),
Linear(hidden_dim, E),
).to(device)
train_pairs = [
(
100.0
* torch.randn(B, T, E, device=device, dtype=dtype)
* input_magnitudes,
torch.randn(B, T, E, device=device, dtype=dtype) * output_magnitudes,
)
for _ in range(20)
]
if iter == 0:
optim = Eve(m.parameters(), lr=0.003)
elif iter == 1:
optim = ScaledAdam(m.named_parameters(), lr=0.03, clipping_scale=2.0)
scheduler = Eden(optim, lr_batches=200, lr_epochs=5, verbose=False)
start = timeit.default_timer()
avg_loss = 0.0
for epoch in range(180):
scheduler.step_epoch()
# if epoch == 100 and iter in [2,3]:
# optim.reset_speedup() # check it doesn't crash.
# if epoch == 130:
# opts = diagnostics.TensorDiagnosticOptions(
# 512
# ) # allow 4 megabytes per sub-module
# diagnostic = diagnostics.attach_diagnostics(m, opts)
for n, (x, y) in enumerate(train_pairs):
y_out = m(x)
loss = ((y_out - y) ** 2).mean() * 100.0
if epoch == 0 and n == 0:
avg_loss = loss.item()
else:
avg_loss = 0.98 * avg_loss + 0.02 * loss.item()
if n == 0 and epoch % 5 == 0:
# norm1 = '%.2e' % (m[0].weight**2).mean().sqrt().item()
# norm1b = '%.2e' % (m[0].bias**2).mean().sqrt().item()
# norm2 = '%.2e' % (m[2].weight**2).mean().sqrt().item()
# norm2b = '%.2e' % (m[2].bias**2).mean().sqrt().item()
# scale1 = '%.2e' % (m[0].weight_scale.exp().item())
# scale1b = '%.2e' % (m[0].bias_scale.exp().item())
# scale2 = '%.2e' % (m[2].weight_scale.exp().item())
# scale2b = '%.2e' % (m[2].bias_scale.exp().item())
lr = scheduler.get_last_lr()[0]
logging.info(
f"Iter {iter}, epoch {epoch}, batch {n}, avg_loss {avg_loss:.4g}, lr={lr:.4e}"
) # , norms={norm1,norm1b,norm2,norm2b}") # scales={scale1,scale1b,scale2,scale2b}
loss.log().backward()
optim.step()
optim.zero_grad()
scheduler.step_batch()
# diagnostic.print_diagnostics()
stop = timeit.default_timer()
logging.info(f"Iter={iter}, Time taken: {stop - start}")
logging.info(f"last lr = {scheduler.get_last_lr()}")
# logging.info("state dict = ", scheduler.state_dict())
# logging.info("optim state_dict = ", optim.state_dict())
logging.info(f"input_magnitudes = {input_magnitudes}")
logging.info(f"output_magnitudes = {output_magnitudes}")
if __name__ == "__main__":
torch.set_num_threads(1)
torch.set_num_interop_threads(1)
logging.getLogger().setLevel(logging.INFO)
import subprocess
s = subprocess.check_output(
"git status -uno .; git log -1; git diff HEAD .", shell=True
)
logging.info(s)
import sys
if len(sys.argv) > 1:
hidden_dim = int(sys.argv[1])
else:
hidden_dim = 200
_test_scaled_adam(hidden_dim)
_test_eden()