xception.py
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.model_zoo as model_zoo
from utils.batchnorm import SynchronizedBatchNorm2d
def fixed_padding(inputs, kernel_size, dilation):
kernel_size_effective = kernel_size + (kernel_size - 1) * (dilation - 1)
pad_total = kernel_size_effective - 1
pad_beg = pad_total // 2
pad_end = pad_total - pad_beg
padded_inputs = F.pad(inputs, (pad_beg, pad_end, pad_beg, pad_end))
return padded_inputs
class SeparableConv2d(nn.Module):
def __init__(self, inplanes, planes, kernel_size=3, stride=1, dilation=1, bias=False, BatchNorm=None):
super(SeparableConv2d, self).__init__()
self.conv1 = nn.Conv2d(inplanes, inplanes, kernel_size, stride, 0, dilation,
groups=inplanes, bias=bias)
self.bn = BatchNorm(inplanes)
self.pointwise = nn.Conv2d(inplanes, planes, 1, 1, 0, 1, 1, bias=bias)
def forward(self, x):
x = fixed_padding(x, self.conv1.kernel_size[0], dilation=self.conv1.dilation[0])
x = self.conv1(x)
x = self.bn(x)
x = self.pointwise(x)
return x
class Block(nn.Module):
def __init__(self, inplanes, planes, reps, stride=1, dilation=1, BatchNorm=None,
start_with_relu=True, grow_first=True, is_last=False):
super(Block, self).__init__()
if planes != inplanes or stride != 1:
self.skip = nn.Conv2d(inplanes, planes, 1, stride=stride, bias=False)
self.skipbn = BatchNorm(planes)
else:
self.skip = None
self.relu = nn.ReLU(inplace=True)
rep = []
filters = inplanes
if grow_first:
rep.append(self.relu)
rep.append(SeparableConv2d(inplanes, planes, 3, 1, dilation, BatchNorm=BatchNorm))
rep.append(BatchNorm(planes))
filters = planes
for i in range(reps - 1):
rep.append(self.relu)
rep.append(SeparableConv2d(filters, filters, 3, 1, dilation, BatchNorm=BatchNorm))
rep.append(BatchNorm(filters))
if not grow_first:
rep.append(self.relu)
rep.append(SeparableConv2d(inplanes, planes, 3, 1, dilation, BatchNorm=BatchNorm))
rep.append(BatchNorm(planes))
if stride != 1:
rep.append(self.relu)
rep.append(SeparableConv2d(planes, planes, 3, 2, BatchNorm=BatchNorm))
rep.append(BatchNorm(planes))
if stride == 1 and is_last:
rep.append(self.relu)
rep.append(SeparableConv2d(planes, planes, 3, 1, BatchNorm=BatchNorm))
rep.append(BatchNorm(planes))
if not start_with_relu:
rep = rep[1:]
self.rep = nn.Sequential(*rep)
def forward(self, inp):
x = self.rep(inp)
if self.skip is not None:
skip = self.skip(inp)
skip = self.skipbn(skip)
else:
skip = inp
x = x + skip
return x
class AlignedXception(nn.Module):
"""
Modified Alighed Xception
"""
def __init__(self, output_stride, BatchNorm,
pretrained=True):
super(AlignedXception, self).__init__()
if output_stride == 16:
entry_block3_stride = 2
middle_block_dilation = 1
exit_block_dilations = (1, 2)
elif output_stride == 8:
entry_block3_stride = 1
middle_block_dilation = 2
exit_block_dilations = (2, 4)
else:
raise NotImplementedError
# Entry flow
self.conv1 = nn.Conv2d(3, 32, 3, stride=1, padding=1, bias=False)
self.bn1 = BatchNorm(32)
self.relu = nn.ReLU(inplace=True)
#
# self.conv2 = nn.Conv2d(32, 64, 3, stride=1, padding=1, bias=False)
# self.bn2 = BatchNorm(64)
self.block1 = Block(32, 64, reps=2, stride=1, BatchNorm=BatchNorm, start_with_relu=False)
self.block2 = Block(64, 128, reps=2, stride=1, BatchNorm=BatchNorm, start_with_relu=False,grow_first=True)
self.block3 = Block(128, 256, reps=2, stride=entry_block3_stride, BatchNorm=BatchNorm,
start_with_relu=True, grow_first=True, is_last=True)
self.convlast1 = SeparableConv2d(256, 128, 3, stride=1, dilation=exit_block_dilations[1], BatchNorm=BatchNorm)
self.bnlast1 = BatchNorm(128)
self.convlast2 = SeparableConv2d(128, 64, 3, stride=1, dilation=exit_block_dilations[1], BatchNorm=BatchNorm)
self.bnlast2 = BatchNorm(64)
# Init weights
self._init_weight()
def forward(self, x):
# Entry flow
x = self.conv1(x)
x = self.bn1(x)
x = self.relu(x)
# x = self.conv2(x)
# x = self.bn2(x)
# x = self.relu(x)
x = self.block1(x)
# add relu here
x = self.relu(x)
low_level_feat = x
x = self.block2(x)
x = self.block3(x)
#--------------------
x = self.convlast1(x)
x = self.bnlast1(x)
x = self.relu(x)
x = self.convlast2(x)
x = self.bnlast2(x)
x = self.relu(x)
return x, low_level_feat
def _init_weight(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
m.weight.data.normal_(0, math.sqrt(2. / n))
elif isinstance(m, SynchronizedBatchNorm2d):
m.weight.data.fill_(1)
m.bias.data.zero_()
elif isinstance(m, nn.BatchNorm2d):
m.weight.data.fill_(1)
m.bias.data.zero_()
if __name__ == "__main__":
import torch
model = AlignedXception(BatchNorm=nn.BatchNorm2d, pretrained=False, output_stride=16)
input = torch.rand(1, 3, 32, 32)
output, low_level_feat = model(input)
print(output.size())
print(low_level_feat.size())
# print(model)
total = sum([param.nelement() for param in model.parameters()])
print("Number of parameter: %.2f M " % (total / 1e6))
batchnorm.py
# -*- coding: utf-8 -*-
# File : batchnorm.py
# Author : Jiayuan Mao
# Email : maojiayuan@gmail.com
# Date : 27/01/2018
#
# This file is part of Synchronized-BatchNorm-PyTorch.
# https://github.com/vacancy/Synchronized-BatchNorm-PyTorch
# Distributed under MIT License.
import collections
import torch
import torch.nn.functional as F
from torch.nn.modules.batchnorm import _BatchNorm
from torch.nn.parallel._functions import ReduceAddCoalesced, Broadcast
from .comm import SyncMaster
__all__ = ['SynchronizedBatchNorm1d', 'SynchronizedBatchNorm2d', 'SynchronizedBatchNorm3d']
def _sum_ft(tensor):
"""sum over the first and last dimention"""
return tensor.sum(dim=0).sum(dim=-1)
def _unsqueeze_ft(tensor):
"""add new dementions at the front and the tail"""
return tensor.unsqueeze(0).unsqueeze(-1)
_ChildMessage = collections.namedtuple('_ChildMessage', ['sum', 'ssum', 'sum_size'])
_MasterMessage = collections.namedtuple('_MasterMessage', ['sum', 'inv_std'])
class _SynchronizedBatchNorm(_BatchNorm):
def __init__(self, num_features, eps=1e-5, momentum=0.1, affine=True):
super(_SynchronizedBatchNorm, self).__init__(num_features, eps=eps, momentum=momentum, affine=affine)
self._sync_master = SyncMaster(self._data_parallel_master)
self._is_parallel = False
self._parallel_id = None
self._slave_pipe = None
def forward(self, input):
# If it is not parallel computation or is in evaluation mode, use PyTorch's implementation.
if not (self._is_parallel and self.training):
return F.batch_norm(
input, self.running_mean, self.running_var, self.weight, self.bias,
self.training, self.momentum, self.eps)
# Resize the input to (B, C, -1).
input_shape = input.size()
input = input.view(input.size(0), self.num_features, -1)
# Compute the sum and square-sum.
sum_size = input.size(0) * input.size(2)
input_sum = _sum_ft(input)
input_ssum = _sum_ft(input ** 2)
# Reduce-and-broadcast the statistics.
if self._parallel_id == 0:
mean, inv_std = self._sync_master.run_master(_ChildMessage(input_sum, input_ssum, sum_size))
else:
mean, inv_std = self._slave_pipe.run_slave(_ChildMessage(input_sum, input_ssum, sum_size))
# Compute the output.
if self.affine:
# MJY:: Fuse the multiplication for speed.
output = (input - _unsqueeze_ft(mean)) * _unsqueeze_ft(inv_std * self.weight) + _unsqueeze_ft(self.bias)
else:
output = (input - _unsqueeze_ft(mean)) * _unsqueeze_ft(inv_std)
# Reshape it.
return output.view(input_shape)
def __data_parallel_replicate__(self, ctx, copy_id):
self._is_parallel = True
self._parallel_id = copy_id
# parallel_id == 0 means master device.
if self._parallel_id == 0:
ctx.sync_master = self._sync_master
else:
self._slave_pipe = ctx.sync_master.register_slave(copy_id)
def _data_parallel_master(self, intermediates):
"""Reduce the sum and square-sum, compute the statistics, and broadcast it."""
# Always using same "device order" makes the ReduceAdd operation faster.
# Thanks to:: Tete Xiao (http://tetexiao.com/)
intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device())
to_reduce = [i[1][:2] for i in intermediates]
to_reduce = [j for i in to_reduce for j in i] # flatten
target_gpus = [i[1].sum.get_device() for i in intermediates]
sum_size = sum([i[1].sum_size for i in intermediates])
sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce)
mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size)
broadcasted = Broadcast.apply(target_gpus, mean, inv_std)
outputs = []
for i, rec in enumerate(intermediates):
outputs.append((rec[0], _MasterMessage(*broadcasted[i * 2:i * 2 + 2])))
return outputs
def _compute_mean_std(self, sum_, ssum, size):
"""Compute the mean and standard-deviation with sum and square-sum. This method
also maintains the moving average on the master device."""
assert size > 1, 'BatchNorm computes unbiased standard-deviation, which requires size > 1.'
mean = sum_ / size
sumvar = ssum - sum_ * mean
unbias_var = sumvar / (size - 1)
bias_var = sumvar / size
self.running_mean = (1 - self.momentum) * self.running_mean + self.momentum * mean.data
self.running_var = (1 - self.momentum) * self.running_var + self.momentum * unbias_var.data
return mean, bias_var.clamp(self.eps) ** -0.5
class SynchronizedBatchNorm1d(_SynchronizedBatchNorm):
r"""Applies Synchronized Batch Normalization over a 2d or 3d input that is seen as a
mini-batch.
.. math::
y = \frac{x - mean[x]}{ \sqrt{Var[x] + \epsilon}} * gamma + beta
This module differs from the built-in PyTorch BatchNorm1d as the mean and
standard-deviation are reduced across all devices during training.
For example, when one uses `nn.DataParallel` to wrap the network during
training, PyTorch's implementation normalize the tensor on each device using
the statistics only on that device, which accelerated the computation and
is also easy to implement, but the statistics might be inaccurate.
Instead, in this synchronized version, the statistics will be computed
over all training samples distributed on multiple devices.
Note that, for one-GPU or CPU-only case, this module behaves exactly same
as the built-in PyTorch implementation.
The mean and standard-deviation are calculated per-dimension over
the mini-batches and gamma and beta are learnable parameter vectors
of size C (where C is the input size).
During training, this layer keeps a running estimate of its computed mean
and variance. The running sum is kept with a default momentum of 0.1.
During evaluation, this running mean/variance is used for normalization.
Because the BatchNorm is done over the `C` dimension, computing statistics
on `(N, L)` slices, it's common terminology to call this Temporal BatchNorm
Args:
num_features: num_features from an expected input of size
`batch_size x num_features [x width]`
eps: a value added to the denominator for numerical stability.
Default: 1e-5
momentum: the value used for the running_mean and running_var
computation. Default: 0.1
affine: a boolean value that when set to ``True``, gives the layer learnable
affine parameters. Default: ``True``
Shape:
- Input: :math:`(N, C)` or :math:`(N, C, L)`
- Output: :math:`(N, C)` or :math:`(N, C, L)` (same shape as input)
Examples:
>>> # With Learnable Parameters
>>> m = SynchronizedBatchNorm1d(100)
>>> # Without Learnable Parameters
>>> m = SynchronizedBatchNorm1d(100, affine=False)
>>> input = torch.autograd.Variable(torch.randn(20, 100))
>>> output = m(input)
"""
def _check_input_dim(self, input):
if input.dim() != 2 and input.dim() != 3:
raise ValueError('expected 2D or 3D input (got {}D input)'
.format(input.dim()))
super(SynchronizedBatchNorm1d, self)._check_input_dim(input)
class SynchronizedBatchNorm2d(_SynchronizedBatchNorm):
r"""Applies Batch Normalization over a 4d input that is seen as a mini-batch
of 3d inputs
.. math::
y = \frac{x - mean[x]}{ \sqrt{Var[x] + \epsilon}} * gamma + beta
This module differs from the built-in PyTorch BatchNorm2d as the mean and
standard-deviation are reduced across all devices during training.
For example, when one uses `nn.DataParallel` to wrap the network during
training, PyTorch's implementation normalize the tensor on each device using
the statistics only on that device, which accelerated the computation and
is also easy to implement, but the statistics might be inaccurate.
Instead, in this synchronized version, the statistics will be computed
over all training samples distributed on multiple devices.
Note that, for one-GPU or CPU-only case, this module behaves exactly same
as the built-in PyTorch implementation.
The mean and standard-deviation are calculated per-dimension over
the mini-batches and gamma and beta are learnable parameter vectors
of size C (where C is the input size).
During training, this layer keeps a running estimate of its computed mean
and variance. The running sum is kept with a default momentum of 0.1.
During evaluation, this running mean/variance is used for normalization.
Because the BatchNorm is done over the `C` dimension, computing statistics
on `(N, H, W)` slices, it's common terminology to call this Spatial BatchNorm
Args:
num_features: num_features from an expected input of
size batch_size x num_features x height x width
eps: a value added to the denominator for numerical stability.
Default: 1e-5
momentum: the value used for the running_mean and running_var
computation. Default: 0.1
affine: a boolean value that when set to ``True``, gives the layer learnable
affine parameters. Default: ``True``
Shape:
- Input: :math:`(N, C, H, W)`
- Output: :math:`(N, C, H, W)` (same shape as input)
Examples:
>>> # With Learnable Parameters
>>> m = SynchronizedBatchNorm2d(100)
>>> # Without Learnable Parameters
>>> m = SynchronizedBatchNorm2d(100, affine=False)
>>> input = torch.autograd.Variable(torch.randn(20, 100, 35, 45))
>>> output = m(input)
"""
def _check_input_dim(self, input):
if input.dim() != 4:
raise ValueError('expected 4D input (got {}D input)'
.format(input.dim()))
super(SynchronizedBatchNorm2d, self)._check_input_dim(input)
class SynchronizedBatchNorm3d(_SynchronizedBatchNorm):
r"""Applies Batch Normalization over a 5d input that is seen as a mini-batch
of 4d inputs
.. math::
y = \frac{x - mean[x]}{ \sqrt{Var[x] + \epsilon}} * gamma + beta
This module differs from the built-in PyTorch BatchNorm3d as the mean and
standard-deviation are reduced across all devices during training.
For example, when one uses `nn.DataParallel` to wrap the network during
training, PyTorch's implementation normalize the tensor on each device using
the statistics only on that device, which accelerated the computation and
is also easy to implement, but the statistics might be inaccurate.
Instead, in this synchronized version, the statistics will be computed
over all training samples distributed on multiple devices.
Note that, for one-GPU or CPU-only case, this module behaves exactly same
as the built-in PyTorch implementation.
The mean and standard-deviation are calculated per-dimension over
the mini-batches and gamma and beta are learnable parameter vectors
of size C (where C is the input size).
During training, this layer keeps a running estimate of its computed mean
and variance. The running sum is kept with a default momentum of 0.1.
During evaluation, this running mean/variance is used for normalization.
Because the BatchNorm is done over the `C` dimension, computing statistics
on `(N, D, H, W)` slices, it's common terminology to call this Volumetric BatchNorm
or Spatio-temporal BatchNorm
Args:
num_features: num_features from an expected input of
size batch_size x num_features x depth x height x width
eps: a value added to the denominator for numerical stability.
Default: 1e-5
momentum: the value used for the running_mean and running_var
computation. Default: 0.1
affine: a boolean value that when set to ``True``, gives the layer learnable
affine parameters. Default: ``True``
Shape:
- Input: :math:`(N, C, D, H, W)`
- Output: :math:`(N, C, D, H, W)` (same shape as input)
Examples:
>>> # With Learnable Parameters
>>> m = SynchronizedBatchNorm3d(100)
>>> # Without Learnable Parameters
>>> m = SynchronizedBatchNorm3d(100, affine=False)
>>> input = torch.autograd.Variable(torch.randn(20, 100, 35, 45, 10))
>>> output = m(input)
"""
def _check_input_dim(self, input):
if input.dim() != 5:
raise ValueError('expected 5D input (got {}D input)'
.format(input.dim()))
super(SynchronizedBatchNorm3d, self)._check_input_dim(input)
comm.py
# -*- coding: utf-8 -*-
# File : comm.py
# Author : Jiayuan Mao
# Email : maojiayuan@gmail.com
# Date : 27/01/2018
#
# This file is part of Synchronized-BatchNorm-PyTorch.
# https://github.com/vacancy/Synchronized-BatchNorm-PyTorch
# Distributed under MIT License.
import queue
import collections
import threading
__all__ = ['FutureResult', 'SlavePipe', 'SyncMaster']
class FutureResult(object):
"""A thread-safe future implementation. Used only as one-to-one pipe."""
def __init__(self):
self._result = None
self._lock = threading.Lock()
self._cond = threading.Condition(self._lock)
def put(self, result):
with self._lock:
assert self._result is None, 'Previous result has\'t been fetched.'
self._result = result
self._cond.notify()
def get(self):
with self._lock:
if self._result is None:
self._cond.wait()
res = self._result
self._result = None
return res
_MasterRegistry = collections.namedtuple('MasterRegistry', ['result'])
_SlavePipeBase = collections.namedtuple('_SlavePipeBase', ['identifier', 'queue', 'result'])
class SlavePipe(_SlavePipeBase):
"""Pipe for master-slave communication."""
def run_slave(self, msg):
self.queue.put((self.identifier, msg))
ret = self.result.get()
self.queue.put(True)
return ret
class SyncMaster(object):
"""An abstract `SyncMaster` object.
- During the replication, as the data parallel will trigger an callback of each module, all slave devices should
call `register(id)` and obtain an `SlavePipe` to communicate with the master.
- During the forward pass, master device invokes `run_master`, all messages from slave devices will be collected,
and passed to a registered callback.
- After receiving the messages, the master device should gather the information and determine to message passed
back to each slave devices.
"""
def __init__(self, master_callback):
"""
Args:
master_callback: a callback to be invoked after having collected messages from slave devices.
"""
self._master_callback = master_callback
self._queue = queue.Queue()
self._registry = collections.OrderedDict()
self._activated = False
def __getstate__(self):
return {'master_callback': self._master_callback}
def __setstate__(self, state):
self.__init__(state['master_callback'])
def register_slave(self, identifier):
"""
Register an slave device.
Args:
identifier: an identifier, usually is the device id.
Returns: a `SlavePipe` object which can be used to communicate with the master device.
"""
if self._activated:
assert self._queue.empty(), 'Queue is not clean before next initialization.'
self._activated = False
self._registry.clear()
future = FutureResult()
self._registry[identifier] = _MasterRegistry(future)
return SlavePipe(identifier, self._queue, future)
def run_master(self, master_msg):
"""
Main entry for the master device in each forward pass.
The messages were first collected from each devices (including the master device), and then
an callback will be invoked to compute the message to be sent back to each devices
(including the master device).
Args:
master_msg: the message that the master want to send to itself. This will be placed as the first
message when calling `master_callback`. For detailed usage, see `_SynchronizedBatchNorm` for an example.
Returns: the message to be sent back to the master device.
"""
self._activated = True
intermediates = [(0, master_msg)]
for i in range(self.nr_slaves):
intermediates.append(self._queue.get())
results = self._master_callback(intermediates)
assert results[0][0] == 0, 'The first result should belongs to the master.'
for i, res in results:
if i == 0:
continue
self._registry[i].result.put(res)
for i in range(self.nr_slaves):
assert self._queue.get() is True
return results[0][1]
@property
def nr_slaves(self):
return len(self._registry)
4943
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
