mirror of
https://github.com/k2-fsa/icefall.git
synced 2025-09-07 08:04:18 +00:00
972 lines
39 KiB
Python
972 lines
39 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.
|
|
|
|
|
|
from typing import List, Optional, Union, Tuple, List
|
|
|
|
import torch
|
|
from torch import Tensor
|
|
from torch.optim import Optimizer
|
|
|
|
|
|
|
|
def _product(*args):
|
|
ans = args[0]
|
|
for x in args[1:]:
|
|
ans = ans * x
|
|
return ans
|
|
|
|
def _sum(*args):
|
|
ans = args[0]
|
|
for x in args[1:]:
|
|
ans = ans + x
|
|
return ans
|
|
|
|
|
|
def _mean_like(x: Tensor, size) -> Tensor:
|
|
"""
|
|
Returns a mean of x with the shape `size`, summing
|
|
over all dimensions in which x.shape[dim] != 1 and size[dim] == 1.
|
|
Args:
|
|
x: Tensor to compute mean over some dims.
|
|
size: a sequence of integers or torch.Size, with
|
|
len(size) == len(x.size), that
|
|
broadcasts with x and elementwise <= x.size.
|
|
"""
|
|
ndim = x.ndim
|
|
dims_to_sum = [i for i in range(ndim) if x.shape[i] != size[i]]
|
|
return x.mean(dims_to_sum, keepdim=True)
|
|
|
|
def _sum_like(x: Tensor, size) -> Tensor:
|
|
"""
|
|
Returns a sum of values of x with the shape `size`, summing
|
|
over all dimensions in which x.shape[dim] != 1 and size[dim] == 1.
|
|
Args:
|
|
x: Tensor to compute mean over some dims.
|
|
size: a sequence of integers or torch.Size, with
|
|
len(size) == len(x.size), that
|
|
broadcasts with x and elementwise <= x.size.
|
|
"""
|
|
ndim = x.ndim
|
|
dims_to_sum = [i for i in range(ndim) if x.shape[i] != size[i]]
|
|
return x.sum(dims_to_sum, keepdim=True)
|
|
|
|
|
|
|
|
|
|
|
|
def _get_factor_shapes(x_shape: List) -> List[List]:
|
|
"""
|
|
Returns a list of the shapes of factors of a tensor with shape
|
|
x_shape, e.g.:
|
|
_get_factor_shapes([3,4]) = [[3,1], [1,4]]
|
|
_get_factor_shapes([5]) = [[1]]
|
|
"""
|
|
base_shape = [1] * len(x_shape)
|
|
shapes = []
|
|
numel = 1
|
|
for n in x_shape:
|
|
numel *= n
|
|
for dim,size in enumerate(x_shape):
|
|
if size != 0 and size != numel:
|
|
shape = list(base_shape)
|
|
shape[dim] = size
|
|
shapes.append(shape)
|
|
if len(shapes) == 0:
|
|
shapes.append(base_shape)
|
|
return shapes
|
|
|
|
def _update_factorization(x: Tensor, x_factorized: Tensor,
|
|
speed: float, eps: float) -> None:
|
|
"""
|
|
This function is like _update_factors but instead of operating
|
|
on the factors directly it operates on their product.
|
|
|
|
Args:
|
|
x: the Tensor to be normalized; for purposes of understanding
|
|
what this function does, assume that x is drawn from some fixed
|
|
distribution. Require x.ndim >= 2 and that at least 2 dims
|
|
have size != 1.
|
|
x_factorized: a Tensor with the same shape as x, but it is
|
|
a product of factors, one factor for each axis i that has
|
|
x.shape[i] != 1 and x.shape[i] != x.numel(); if no axes
|
|
remain, we have a single scalar factor. This function
|
|
modifies x_factorized to be closer to a model of the stddevs
|
|
of elements of x.
|
|
speed: update speed, e.g 0.02 (consider this as 1-beta_2
|
|
where beta_2 is the 2nd beta from Adam). Must satisfy
|
|
speed > 0 and speed * x.ndim < 1.
|
|
eps: small value intended to prevent zero values in
|
|
x_factorized; you can think of it as a floor on elements
|
|
of x_factorized.
|
|
"""
|
|
x_norm_var = (x / x_factorized) ** 2 + eps*eps
|
|
shapes = _get_factor_shapes(list(x.shape))
|
|
factors = []
|
|
for shape in shapes:
|
|
# elements of this_mean will be >1.0 if the squares of elements in x are
|
|
# larger than predicted by x_factorized; and <1.0 otherwise.
|
|
this_mean = _mean_like(x_norm_var, shape)
|
|
f = ((1.0 - speed) + speed * this_mean)
|
|
factors.append(f)
|
|
x_factorized *= _product(*factors)
|
|
|
|
def _get_factor_grads(x: Tensor, x_grad: Tensor) -> List[Tensor]:
|
|
"""
|
|
Get gradients w.r.t. the factors in a factorized representation of x.
|
|
Consider the example where
|
|
x_norm.shape == (3, 4), x_factor1.shape == (3, 1), x_factor2.shape == (1, 4),
|
|
and:
|
|
x = x_norm * x_factor1.exp() * x_factor2.exp()
|
|
Given x and x_grad, this function will return the (x_factor1_grad, x_factor2_grad),
|
|
i.e. the loss-function derivatives w.r.t. x_factor1 and x_factor2.
|
|
|
|
The specific values of the factors are not needed here; their
|
|
product is enough.
|
|
"""
|
|
shapes = _get_factor_shapes(list(x.shape))
|
|
prod = x * x_grad
|
|
return [ _sum_like(prod, shape) for shape in shapes ]
|
|
|
|
|
|
|
|
|
|
def _init_factors(x: Tensor, eps: float = 1.0e-04) -> Tuple[Tensor]:
|
|
"""
|
|
Initialize some factors which we will use to normalize the variance of x.
|
|
For example: for a Tensor of shape (3, 4) we will return factors of shapes
|
|
(3, 1) and (1, 4) having the property that x / (factor1 * factor2) has quite
|
|
close to unit variance when indexed on either dimension, i.e. that
|
|
(x[i,:]**2).mean() is close to 1 and (x[:,j]**2).mean() is close to 1.
|
|
"""
|
|
assert x.numel() > 1
|
|
factors = []
|
|
x2 = x**2 + eps*eps
|
|
for d in range(x.ndim):
|
|
if x.shape[d] != x.numel() and x.shape[d] != 1:
|
|
shape = [1] * x.ndim
|
|
shape[d] = x.shape[d]
|
|
factor = _mean_like(x2, shape)
|
|
x2 = x2 / factor
|
|
factors.append(factor ** 0.5)
|
|
# We need at least one factor, to capture the scalar magnitude.
|
|
if len(factors) == 0:
|
|
factors.append(x2.mean(dim=tuple(range(x.ndim)),
|
|
keepdims=True) ** 0.5)
|
|
return factors
|
|
|
|
|
|
def _init_factorization(x: Tensor, eps: float = 1.0e-04) -> Tensor:
|
|
return _product(*_init_factors(x, eps))
|
|
|
|
class Abel(Optimizer):
|
|
"""
|
|
Implements the Abel algorithm. You can think of this as a modified version
|
|
of the Adam algorithm in which the parameters have an implicit factorization
|
|
that is maintained in a special way. (Actually this can also be considered
|
|
as a refinement of the "Eve" algorithm, see below).
|
|
Suppose we have a parameter `x` of
|
|
shape (40, 50). Imagine that we factorize x as follows:
|
|
|
|
x = x_norm * x_factor1.exp() * x_factor2.exp()
|
|
|
|
where (x_norm, x_factor1, x_factor2) are of shapes: (40, 50), (40, 1), (1, 50),
|
|
and all 3 of x_norm, x_factor1 and x_factor2 are learned by Adam. Also imagine
|
|
that we frequently reconstruct x and then refactorize it in such a way that
|
|
the standard deviation of elements of x, when restricted to any specific rows or column,
|
|
equals `target_rms == 0.1`. So it's a bit like we constrain the elements of x_norm
|
|
to have a specific standard deviation on any row or column, and instead learn the
|
|
magnitudes of the rows and columns in log-space as x_factor1 and x_factor2.
|
|
But this is all done inside the optimizer. To simplify things a bit when updating
|
|
the factorization, we assume in the explanation below and in most of the code that
|
|
target_rms equals 1.0; it then simply becomes a factor of `target_rms` in the
|
|
learning rate when we update x (but not the part of the step that comes from
|
|
the udpate of x_factor1 and x_factor2).
|
|
|
|
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-08).
|
|
target_rms (float, optional): target root-mean-square value of
|
|
x_norm in factorization (conceptually). Actually this now just becomes
|
|
a factor in the learning rate, we may remove it at some point.
|
|
|
|
|
|
.. _Adam\: A Method for Stochastic Optimization:
|
|
https://arxiv.org/abs/1412.6980
|
|
.. _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-04,
|
|
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 < target_rms <= 1.0:
|
|
raise ValueError("Invalid target_rms value: {}".format(target_rms))
|
|
defaults = dict(
|
|
lr=lr,
|
|
betas=betas,
|
|
eps=eps,
|
|
target_rms=target_rms,
|
|
)
|
|
super(Abel, self).__init__(params, defaults)
|
|
|
|
def __setstate__(self, state):
|
|
super(Abel, 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]
|
|
eps = group["eps"]
|
|
|
|
# State initialization
|
|
if len(state) == 0:
|
|
state["step"] = 0
|
|
|
|
# The change in parameter p from one step to the next; this
|
|
# is the moving average of the steps before taking momentum
|
|
# (beta) into account. We implement momentum this way,
|
|
# instead of the "exp_avg" method as in Adam, to save a
|
|
# little memory and complexity because this way the momentum for the
|
|
# various fctors doesn't have to be kept track of
|
|
# separately.
|
|
state["delta"] = torch.zeros_like(
|
|
p, memory_format=torch.preserve_format
|
|
)
|
|
|
|
# Exponential moving average of squared gradient values,
|
|
# w.r.t. x.
|
|
state["exp_avg_sq"] = torch.zeros_like(
|
|
p, memory_format=torch.preserve_format
|
|
)
|
|
|
|
if p.numel() > 1:
|
|
# exponential moving average of gradients w.r.t
|
|
# the factors x_factor1, x_factor2 etc.
|
|
state["factors_exp_avg_sq"] = [
|
|
torch.zeros(*shape, dtype=p.dtype, device=p.device)
|
|
for shape in _get_factor_shapes(list(p.shape))
|
|
]
|
|
# we call _init_factorization inside step().
|
|
state["factorization"] = torch.zeros_like(p)
|
|
|
|
exp_avg_sq = state["exp_avg_sq"]
|
|
delta, exp_avg_sq = state["delta"], state["exp_avg_sq"]
|
|
|
|
lr = group["lr"]
|
|
target_rms = group["target_rms"]
|
|
eps = group["eps"]
|
|
|
|
beta1, beta2 = group["betas"]
|
|
state["step"] += 1
|
|
step = state["step"]
|
|
# we don't bother with bias_correction1, this only affects the first few
|
|
# steps and anyway the slower-than-it-should be update probably helps stability.
|
|
bias_correction2 = 1 - beta2 ** step
|
|
|
|
# keep this_exp_avg_sq updated as an average variance of gradients.
|
|
def update_exp_avg_sq(this_grad, this_exp_avg_sq):
|
|
this_exp_avg_sq.mul_(beta2).addcmul_(this_grad, this_grad,
|
|
value=1 - beta2)
|
|
|
|
update_exp_avg_sq(grad, exp_avg_sq)
|
|
if p.numel() == 1:
|
|
# simpler way of setting this_delta; this is going to be equivalent
|
|
# to standard Adam.
|
|
denom = (exp_avg_sq/bias_correction2 + eps*eps).sqrt()
|
|
this_delta = grad / denom
|
|
|
|
# eventually will remove this special case, this is to deal
|
|
# with our model that has scaling factors.
|
|
lr_eff = lr * (bias_correction2 ** 0.5)
|
|
else:
|
|
# update that includes factorization-related stuff..
|
|
factors_exp_avg_sq, factorization = state["factors_exp_avg_sq"], state["factorization"]
|
|
|
|
# factor_grads: List[Tensor], len == num-factors; the gradients
|
|
# w.r.t. the factors of x (x_factor1, x_factor2..)
|
|
factor_grads = _get_factor_grads(p, grad)
|
|
|
|
if step < 10 or step % 10 == 1:
|
|
# note, the first step number we'll see is 1.
|
|
# update the factorization this only every 10 steps, to save time.
|
|
if step % 1000 == 1:
|
|
# Periodically refresh the factorization; this is
|
|
# out of concern that after a large number of
|
|
# multiplications, roundoff could cause it to drift
|
|
# significantly away from being a product of
|
|
# independent factors, one per dimension.
|
|
factorization[:] = _init_factorization(p, eps)
|
|
_update_factorization(p, factorization,
|
|
speed=0.1,
|
|
eps=eps)
|
|
|
|
|
|
factors_sum = None
|
|
numel = p.numel()
|
|
for g, e in zip(factor_grads, factors_exp_avg_sq):
|
|
this_numel = numel / g.numel()
|
|
update_exp_avg_sq(g, e)
|
|
# this_numel**0.5 will turn into this_numel**-0.25 when we sqrt() and
|
|
# divide by denom. This becomes a factor in the learning rate.
|
|
# The technique naturally makes us learn faster by (this_numel**0.5)
|
|
# in any of the parameter directions corresponding to rows or columns,
|
|
# and because we don't want to be to aggressive and risk instability or
|
|
# excess param noise, we reduce this to (this_numel**0.25) by dividing
|
|
# the effective LR by this_numel**0.25.
|
|
this_denom = (e*(this_numel**0.5)/bias_correction2 + eps*eps).sqrt()
|
|
assert g.shape == this_denom.shape
|
|
factor_delta = g / this_denom
|
|
factors_sum = (factor_delta if factors_sum is None
|
|
else factors_sum + factor_delta)
|
|
|
|
# at this point, factors_sum either has the same shape as p, or,
|
|
# if p has only one non-trivial dimension, it has a shape equal
|
|
# to (1,) * p.ndim. After multiplying by the learning rate
|
|
# and p, it will represent the change in parameter that arises
|
|
# from updating the factors (before considering momentum!)
|
|
|
|
|
|
denom = (exp_avg_sq/bias_correction2 +
|
|
eps*eps).sqrt()
|
|
|
|
# `grad * factorization / denom` represents the change in parameters x, only considering
|
|
# the change from `x_norm` in the factorization, and
|
|
# before multiplying by the learning rate or taking into account
|
|
# momentum. We compute this in the original space of `x`,
|
|
# simply because this is more convenient, but due to invariants
|
|
# of the Adam update this would be exactly the same if computed
|
|
# on `x_norm`, except for some small differences relating to
|
|
# `eps`.
|
|
# `p * factors_sum` is the contribution from changes in x_factor1
|
|
# and x_factor2: again, before taking into account the learning
|
|
# rate or momentum.
|
|
this_delta = ((grad * factorization / denom) + p * factors_sum)
|
|
lr_eff = lr * (bias_correction2 ** 0.5) / target_rms
|
|
|
|
delta.mul_(beta1).add_(this_delta, alpha=-lr_eff*(1-beta1))
|
|
|
|
p.add_(delta)
|
|
|
|
|
|
return loss
|
|
|
|
@torch.no_grad()
|
|
def reset(self):
|
|
"""
|
|
Resets parts of the state of the optimizer. This is to be called after
|
|
refactorization/orthogonalization of parameters. At these points we
|
|
discard any momentum because it is no longer accurate/relevant;
|
|
we reconstruct the factorization of x (i.e. x_factor1 and x_factor2
|
|
in the intro to this class); and we modify exp_avg_sq to
|
|
0.5 (exp_avg_sq + exp_avg_sq.mean()); this ensures that none of the
|
|
exp_avg_sq values will be substantially too small and prevents any
|
|
too-fast updates.
|
|
"""
|
|
for s in self.state.values():
|
|
s["delta"].zero_() # zero out momentum
|
|
s["step"] = 0 # will cause state["factorization"] to be reinitialized
|
|
s["exp_avg_sq"].zero_()
|
|
if "factors_exp_avg_sq" in s:
|
|
for e in s["factors_exp_avg_sq"]:
|
|
e.zero_()
|
|
|
|
|
|
|
|
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"]
|
|
|
|
# forget bias_correction1. We use normal momentum. exp_avg really
|
|
# just stores the moving-average gradient step.
|
|
bias_correction2 = 1 - beta2 ** self._step_for_bias_correction(state["step"])
|
|
|
|
grad = self._change_coordinates(grad, state, forward=True)
|
|
|
|
# Decay the second moment running average coefficient
|
|
exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
|
|
denom = (exp_avg_sq.sqrt()).add_(group["eps"])
|
|
|
|
|
|
step_size = group["lr"] # forget bias_correction1
|
|
target_rms = group["target_rms"]
|
|
weight_decay = group["weight_decay"]
|
|
|
|
this_delta = grad / denom
|
|
this_delta = self._change_coordinates(this_delta, state, forward=False)
|
|
|
|
# treat/repurpose exp_avg as moving-average of `this_delta`
|
|
exp_avg.mul_(beta1).add_(this_delta, alpha=-step_size*(1-beta1)*(bias_correction2 ** 0.5))
|
|
|
|
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.add_(exp_avg)
|
|
state["step"] += 1
|
|
|
|
return loss
|
|
|
|
def _change_coordinates(self,
|
|
grad: Tensor,
|
|
state: dict,
|
|
forward: bool):
|
|
"""
|
|
Possibly performs some magnitude-preserving changes of co-ordinates on `grad`
|
|
"""
|
|
ndim = grad.ndim
|
|
numel = grad.numel()
|
|
for dim in range(grad.ndim):
|
|
# see for each dim in turn whether we want to perform any changes in co-ordinates,
|
|
# or store any stats.
|
|
size = grad.shape[dim]
|
|
if size <= 3 or size > 1024 or size == numel:
|
|
# we don't do any such co-ordinate changes in dims that are too small (no point)
|
|
# or large (too slow)
|
|
continue
|
|
if dim == 0:
|
|
# FOR NOW: don't do such co-ordinate changes on output
|
|
# dims, which will generally be dimension zero. We can revisit this later.
|
|
continue
|
|
grad = self._change_coordinates_for_dim(grad, state, dim, forward)
|
|
return grad
|
|
|
|
|
|
def _change_coordinates_for_dim(self,
|
|
grad: Tensor,
|
|
state: dict,
|
|
dim: int,
|
|
forward: bool,
|
|
orig_dim: int = -1):
|
|
assert grad.ndim > 1
|
|
if not (grad.ndim == 2 and dim == 1):
|
|
# normalize the format and recurse.
|
|
dims_order = []
|
|
rev_dims_order = []
|
|
ndim = grad.ndim
|
|
for i in range(dim):
|
|
dims_order.append(i)
|
|
rev_dims_order.append(i)
|
|
rev_dims_order.append(ndim-1)
|
|
for i in range(dim+1, ndim):
|
|
dims_order.append(i)
|
|
rev_dims_order.append(i-1)
|
|
dims_order.append(dim)
|
|
# e.g. ndim=4, dim=1, dims_order=(0,2,3,1), rev_dims_order=(0,3,1,2)
|
|
new_grad = grad.permute(*dims_order)
|
|
assert new_grad.shape[-1] == grad.shape[dim]
|
|
new_shape = new_grad.shape
|
|
new_grad = new_grad.reshape(-1, new_grad.shape[-1])
|
|
new_grad = self._change_coordinates_for_dim(new_grad, state, 1, forward,
|
|
orig_dim=dim)
|
|
return new_grad.reshape(new_shape).permute(*rev_dims_order)
|
|
|
|
# OK: grad.ndim == 2 and dim == 1
|
|
size = grad.shape[1]
|
|
if orig_dim == -1:
|
|
orig_dim = dim
|
|
step = state["step"]
|
|
must_store_stats, must_zero_stats = self._must_store_stats(step)
|
|
if must_store_stats:
|
|
# store stats for 20 iters preceding when we estimate the
|
|
# transform.
|
|
stats_name = f"cov_{orig_dim}"
|
|
if not stats_name in state:
|
|
state[stats_name] = torch.zeros(size, size, dtype=grad.dtype,
|
|
device=grad.device)
|
|
cov = state[stats_name]
|
|
if must_zero_stats:
|
|
#print("zero")
|
|
cov.zero_()
|
|
cov += torch.matmul(grad.t(), grad) * (1/grad.shape[0])
|
|
|
|
transform_name = f"t_{orig_dim}"
|
|
if transform_name not in state:
|
|
# make sure they're all allocated at the start, may be better for
|
|
# fragmention and debugging reasons (although may not).
|
|
state[transform_name] = torch.eye(size, device=grad.device,
|
|
dtype=grad.dtype)
|
|
|
|
must_estimate_transform = self._must_estimate_transform(step)
|
|
if must_estimate_transform:
|
|
stats_name = f"cov_{orig_dim}"
|
|
cov = state[stats_name]
|
|
l, U = cov.symeig(eigenvectors=True)
|
|
#print("estimate, l = ", l[::20])
|
|
t = U.t() # t is indexed (new_dim, old_dim) a.k.a. (transformed_dim, real_dim)
|
|
state[transform_name][:] = t
|
|
state["exp_avg_sq"].zero_() # zero exp-avg-sq stats, we'll restart these from scratch.
|
|
|
|
t = state[transform_name]
|
|
if forward:
|
|
return torch.matmul(grad, t.t())
|
|
else:
|
|
return torch.matmul(grad, t)
|
|
|
|
def _must_store_stats(self, step: int) -> Tuple[bool, bool]:
|
|
"""
|
|
Returns (must_store_stats, must_zero_stats), where
|
|
must_store_stats: bool, is True for the 20 steps preceding
|
|
a step on which we will estimate the transform.
|
|
must_zero_stats: bool, is True for only the 1st of the 20
|
|
steps mentioned above.
|
|
"""
|
|
if step < 4000:
|
|
if step < 2000:
|
|
return (step % 500 >= 480, step % 500 == 480)
|
|
else:
|
|
return (step % 1000 >= 980, step % 1000 == 980)
|
|
else:
|
|
return (step % 2000 >= 1980, step % 2000 == 1980)
|
|
|
|
def _must_estimate_transform(self, step: int) -> bool:
|
|
"""
|
|
Returns True if we will re-estimate the transform on this step
|
|
"""
|
|
if step < 4000:
|
|
if step < 2000:
|
|
return step % 500 == 0 and step > 0
|
|
else:
|
|
return step % 1000 == 0 and step > 0
|
|
else:
|
|
return step % 2000 == 0
|
|
|
|
def _step_for_bias_correction(self, step: int) -> bool:
|
|
"""
|
|
Return a modified step for bias-correction (actually just to give us the
|
|
right denominator for the exp-avg-sq stats).. this needs to take into
|
|
account how long it has been since we re-estimated the transforms, because
|
|
we zero the exp-avg-sq stats at those times.
|
|
|
|
The + 1 is for the "current step". We post-increment the step so it
|
|
starts from 0.
|
|
"""
|
|
if step < 4000:
|
|
if step < 2000:
|
|
return step % 500 + 1
|
|
else:
|
|
return step % 1000 + 1
|
|
else:
|
|
return step % 2000 + 1
|
|
|
|
|
|
|
|
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("initial_lr", group["lr"])
|
|
|
|
self.base_lrs = [
|
|
group["initial_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 {
|
|
"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`.
|
|
"""
|
|
self.__dict__.update(state_dict)
|
|
|
|
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:
|
|
print(
|
|
f"Epoch={self.epoch}, batch={self.batch}: adjusting learning rate"
|
|
f" of group {group} to {lr:.4e}."
|
|
)
|
|
|
|
|
|
class Eden(LRScheduler):
|
|
"""
|
|
Eden scheduler.
|
|
lr = initial_lr * (((batch**2 + lr_batches**2) / lr_batches**2) ** -0.25 *
|
|
(((epoch**2 + lr_epochs**2) / lr_epochs**2) ** -0.25))
|
|
|
|
E.g. suggest initial-lr = 0.003 (passed to optimizer).
|
|
|
|
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],
|
|
verbose: bool = False,
|
|
):
|
|
super(Eden, self).__init__(optimizer, verbose)
|
|
self.lr_batches = lr_batches
|
|
self.lr_epochs = lr_epochs
|
|
|
|
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
|
|
)
|
|
return [x * factor for x in self.base_lrs]
|
|
|
|
|
|
def _test_eden():
|
|
m = torch.nn.Linear(100, 100)
|
|
optim = Eve(m.parameters(), lr=0.003)
|
|
|
|
scheduler = Eden(optim, lr_batches=30, 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()
|
|
|
|
print("last lr = ", scheduler.get_last_lr())
|
|
print("state dict = ", scheduler.state_dict())
|
|
|
|
|
|
def _test_abel():
|
|
import timeit
|
|
from scaling import ScaledLinear
|
|
E = 100
|
|
B = 4
|
|
T = 2
|
|
print("in test_abel")
|
|
device = torch.device('cuda')
|
|
dtype = torch.float32
|
|
|
|
# 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]:
|
|
Linear = torch.nn.Linear if iter == 0 else ScaledLinear
|
|
m = torch.nn.Sequential(Linear(E, 200),
|
|
torch.nn.ReLU(),
|
|
Linear(200, 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 = Abel(m.parameters(), lr=0.003)
|
|
else: optim = Eve(m.parameters(), lr=0.003)
|
|
scheduler = Eden(optim, lr_batches=300, lr_epochs=20, verbose=False)
|
|
|
|
start = timeit.default_timer()
|
|
for epoch in range(150):
|
|
scheduler.step_epoch()
|
|
if isinstance(optim, Abel) and epoch % 5 == 0: # this is every 100 minibatches.
|
|
optim.reset() # just test that calling reset() doesn't
|
|
# break things, you wouldn't call it every
|
|
# epoch like this.
|
|
for n, (x,y) in enumerate(train_pairs):
|
|
y_out = m(x)
|
|
loss = ((y_out - y)**2).mean() * 100.0
|
|
if n == 0 and epoch % 10 == 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())
|
|
print(f"Iter {iter}, epoch {epoch}, batch {n}, loss {loss.item()}, norms={norm1,norm1b,norm2,norm2b}") # scales={scale1,scale1b,scale2,scale2b}
|
|
loss.log().backward()
|
|
optim.step()
|
|
optim.zero_grad()
|
|
scheduler.step_batch()
|
|
|
|
stop = timeit.default_timer()
|
|
print(f"Iter={iter}, Time taken: {stop - start}")
|
|
|
|
print("last lr = ", scheduler.get_last_lr())
|
|
#print("state dict = ", scheduler.state_dict())
|
|
#print("optim state_dict = ", optim.state_dict())
|
|
print("input_magnitudes = ", input_magnitudes)
|
|
print("output_magnitudes = ", output_magnitudes)
|
|
|
|
def stddev(x):
|
|
return ((x-x.mean())**2).mean().sqrt()
|
|
print("Un-normalized input col magnitudes log-stddev: ", stddev((m[0].weight**2).sum(dim=0).sqrt().log()))
|
|
print("Normalized input col magnitudes log-stddev: ", stddev(((m[0].weight**2).sum(dim=0).sqrt() * input_magnitudes).log()))
|
|
|
|
print("Un-normalized 0-output row magnitudes log-stddev: ", stddev((m[0].weight**2).sum(dim=1).sqrt().log()))
|
|
print("Un-normalized 2-input col magnitudes log-stddev: ", stddev((m[2].weight**2).sum(dim=0).sqrt().log()))
|
|
print("Un-normalized 2-output row magnitudes log-stddev: ", stddev((m[2].weight**2).sum(dim=1).sqrt().log()))
|
|
|
|
print("Normalized output row magnitudes log-stddev: ", stddev(((m[2].weight**2).sum(dim=1).sqrt() / output_magnitudes).log()))
|
|
|
|
|
|
|
|
if __name__ == "__main__":
|
|
_test_abel()
|
|
#_test_eden()
|