zdou0830's picture
desco
749745d
raw
history blame
5.01 kB
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
import torch
from torch import nn
import torch.distributed as dist
import maskrcnn_benchmark.utils.comm as comm
from torch.autograd.function import Function
class FrozenBatchNorm2d(nn.Module):
"""
BatchNorm2d where the batch statistics and the affine parameters
are fixed
"""
def __init__(self, n):
super(FrozenBatchNorm2d, self).__init__()
self.register_buffer("weight", torch.ones(n))
self.register_buffer("bias", torch.zeros(n))
self.register_buffer("running_mean", torch.zeros(n))
self.register_buffer("running_var", torch.ones(n))
def forward(self, x):
scale = self.weight * self.running_var.rsqrt()
bias = self.bias - self.running_mean * scale
scale = scale.reshape(1, -1, 1, 1)
bias = bias.reshape(1, -1, 1, 1)
return x * scale + bias
class AllReduce(Function):
@staticmethod
def forward(ctx, input):
input_list = [torch.zeros_like(input) for k in range(dist.get_world_size())]
# Use allgather instead of allreduce since I don't trust in-place operations ..
dist.all_gather(input_list, input, async_op=False)
inputs = torch.stack(input_list, dim=0)
return torch.sum(inputs, dim=0)
@staticmethod
def backward(ctx, grad_output):
dist.all_reduce(grad_output, async_op=False)
return grad_output
class NaiveSyncBatchNorm2d(nn.BatchNorm2d):
"""
In PyTorch<=1.5, ``nn.SyncBatchNorm`` has incorrect gradient
when the batch size on each worker is different.
(e.g., when scale augmentation is used, or when it is applied to mask head).
This is a slower but correct alternative to `nn.SyncBatchNorm`.
Note:
There isn't a single definition of Sync BatchNorm.
When ``stats_mode==""``, this module computes overall statistics by using
statistics of each worker with equal weight. The result is true statistics
of all samples (as if they are all on one worker) only when all workers
have the same (N, H, W). This mode does not support inputs with zero batch size.
When ``stats_mode=="N"``, this module computes overall statistics by weighting
the statistics of each worker by their ``N``. The result is true statistics
of all samples (as if they are all on one worker) only when all workers
have the same (H, W). It is slower than ``stats_mode==""``.
Even though the result of this module may not be the true statistics of all samples,
it may still be reasonable because it might be preferrable to assign equal weights
to all workers, regardless of their (H, W) dimension, instead of putting larger weight
on larger images. From preliminary experiments, little difference is found between such
a simplified implementation and an accurate computation of overall mean & variance.
"""
def __init__(self, *args, stats_mode="", **kwargs):
super().__init__(*args, **kwargs)
assert stats_mode in ["", "N"]
self._stats_mode = stats_mode
def forward(self, input):
if comm.get_world_size() == 1 or not self.training:
return super().forward(input)
B, C = input.shape[0], input.shape[1]
mean = torch.mean(input, dim=[0, 2, 3])
meansqr = torch.mean(input * input, dim=[0, 2, 3])
if self._stats_mode == "":
assert B > 0, 'SyncBatchNorm(stats_mode="") does not support zero batch size.'
vec = torch.cat([mean, meansqr], dim=0)
vec = AllReduce.apply(vec) * (1.0 / dist.get_world_size())
mean, meansqr = torch.split(vec, C)
momentum = self.momentum
else:
if B == 0:
vec = torch.zeros([2 * C + 1], device=mean.device, dtype=mean.dtype)
vec = vec + input.sum() # make sure there is gradient w.r.t input
else:
vec = torch.cat([mean, meansqr, torch.ones([1], device=mean.device, dtype=mean.dtype)], dim=0)
vec = AllReduce.apply(vec * B)
total_batch = vec[-1].detach()
momentum = total_batch.clamp(max=1) * self.momentum # no update if total_batch is 0
total_batch = torch.max(total_batch, torch.ones_like(total_batch)) # avoid div-by-zero
mean, meansqr, _ = torch.split(vec / total_batch, C)
var = meansqr - mean * mean
invstd = torch.rsqrt(var + self.eps)
scale = self.weight * invstd
bias = self.bias - mean * scale
scale = scale.reshape(1, -1, 1, 1)
bias = bias.reshape(1, -1, 1, 1)
self.running_mean += momentum * (mean.detach() - self.running_mean)
self.running_var += momentum * (var.detach() - self.running_var)
return input * scale + bias