oil/architectures/img_classifiers/shake_shake.py
# coding: utf-8
# Copyright (c) 2018, Curious AI Ltd. All rights reserved.
#
# This work is licensed under the Creative Commons Attribution-NonCommercial
# 4.0 International License. To view a copy of this license, visit
# http://creativecommons.org/licenses/by-nc/4.0/ or send a letter to
# Creative Commons, PO Box 1866, Mountain View, CA 94042, USA.
"""
ShakeShake model definition
ported from https://github.com/CuriousAI/mean-teacher/blob/master/pytorch/mean_teacher/architectures.py
"""
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Function
from ...utils.utils import export,Named
import itertools
__all__ = ['ShakeShake26','ResNext152']
class ResNet224x224(nn.Module,metaclass=Named):
def __init__(self, block, layers, channels, groups=1, num_targets=1000, downsample='basic'):
super().__init__()
assert len(layers) == 4
self.downsample_mode = downsample
self.inplanes = 64
self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=7, stride=2, padding=3,
bias=False)
self.bn1 = nn.BatchNorm2d(self.inplanes)
self.relu = nn.ReLU(inplace=True)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
self.layer1 = self._make_layer(block, channels, groups, layers[0])
self.layer2 = self._make_layer(
block, channels * 2, groups, layers[1], stride=2)
self.layer3 = self._make_layer(
block, channels * 4, groups, layers[2], stride=2)
self.layer4 = self._make_layer(
block, channels * 8, groups, layers[3], stride=2)
self.avgpool = nn.AvgPool2d(7)
self.fc1 = nn.Linear(block.out_channels(
channels * 8, groups), num_targets)
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, nn.BatchNorm2d):
m.weight.data.fill_(1)
m.bias.data.zero_()
def _make_layer(self, block, planes, groups, blocks, stride=1):
downsample = None
if stride != 1 or self.inplanes != block.out_channels(planes, groups):
if self.downsample_mode == 'basic' or stride == 1:
downsample = nn.Sequential(
nn.Conv2d(self.inplanes, block.out_channels(planes, groups),
kernel_size=1, stride=stride, bias=False),
nn.BatchNorm2d(block.out_channels(planes, groups)),
)
elif self.downsample_mode == 'shift_conv':
downsample = ShiftConvDownsample(in_channels=self.inplanes,
out_channels=block.out_channels(planes, groups))
else:
assert False
layers = []
layers.append(block(self.inplanes, planes, groups, stride, downsample))
self.inplanes = block.out_channels(planes, groups)
for i in range(1, blocks):
layers.append(block(self.inplanes, planes, groups))
return nn.Sequential(*layers)
def forward(self, x):
x = self.conv1(x)
x = self.bn1(x)
x = self.relu(x)
x = self.maxpool(x)
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.layer4(x)
x = self.avgpool(x)
x = x.view(x.size(0), -1)
return self.fc1(x)
class ResNet32x32(nn.Module,metaclass=Named):
def __init__(self, block, layers, channels, groups=1, num_targets=1000, downsample='basic'):
super().__init__()
assert len(layers) == 3
self.downsample_mode = downsample
self.inplanes = 16
self.conv1 = nn.Conv2d(3, 16, kernel_size=3, stride=1,
padding=1, bias=False)
self.layer1 = self._make_layer(block, channels, groups, layers[0])
self.layer2 = self._make_layer(
block, channels * 2, groups, layers[1], stride=2)
self.layer3 = self._make_layer(
block, channels * 4, groups, layers[2], stride=2)
self.avgpool = nn.AvgPool2d(8)
self.fc1 = nn.Linear(block.out_channels(
channels * 4, groups), num_targets)
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, nn.BatchNorm2d):
m.weight.data.fill_(1)
m.bias.data.zero_()
def _make_layer(self, block, planes, groups, blocks, stride=1):
downsample = None
if stride != 1 or self.inplanes != block.out_channels(planes, groups):
if self.downsample_mode == 'basic' or stride == 1:
downsample = nn.Sequential(
nn.Conv2d(self.inplanes, block.out_channels(planes, groups),
kernel_size=1, stride=stride, bias=False),
nn.BatchNorm2d(block.out_channels(planes, groups)),
)
elif self.downsample_mode == 'shift_conv':
downsample = ShiftConvDownsample(in_channels=self.inplanes,
out_channels=block.out_channels(planes, groups))
else:
assert False
layers = []
layers.append(block(self.inplanes, planes, groups, stride, downsample))
self.inplanes = block.out_channels(planes, groups)
for i in range(1, blocks):
layers.append(block(self.inplanes, planes, groups))
return nn.Sequential(*layers)
def forward(self, x):
x = self.conv1(x)
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.avgpool(x)
x = x.view(x.size(0), -1)
return self.fc1(x)
def conv3x3(in_planes, out_planes, stride=1):
"3x3 convolution with padding"
return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
padding=1, bias=False)
class BottleneckBlock(nn.Module):
@classmethod
def out_channels(cls, planes, groups):
if groups > 1:
return 2 * planes
else:
return 4 * planes
def __init__(self, inplanes, planes, groups, stride=1, downsample=None):
super().__init__()
self.relu = nn.ReLU(inplace=True)
self.conv_a1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False)
self.bn_a1 = nn.BatchNorm2d(planes)
self.conv_a2 = nn.Conv2d(
planes, planes, kernel_size=3, stride=stride, padding=1, bias=False, groups=groups)
self.bn_a2 = nn.BatchNorm2d(planes)
self.conv_a3 = nn.Conv2d(planes, self.out_channels(
planes, groups), kernel_size=1, bias=False)
self.bn_a3 = nn.BatchNorm2d(self.out_channels(planes, groups))
self.downsample = downsample
self.stride = stride
def forward(self, x):
a, residual = x, x
a = self.conv_a1(a)
a = self.bn_a1(a)
a = self.relu(a)
a = self.conv_a2(a)
a = self.bn_a2(a)
a = self.relu(a)
a = self.conv_a3(a)
a = self.bn_a3(a)
if self.downsample is not None:
residual = self.downsample(residual)
return self.relu(residual + a)
class ShakeShakeBlock(nn.Module):
@classmethod
def out_channels(cls, planes, groups):
assert groups == 1
return planes
def __init__(self, inplanes, planes, groups, stride=1, downsample=None):
super().__init__()
assert groups == 1
self.conv_a1 = conv3x3(inplanes, planes, stride)
self.bn_a1 = nn.BatchNorm2d(planes)
self.conv_a2 = conv3x3(planes, planes)
self.bn_a2 = nn.BatchNorm2d(planes)
self.conv_b1 = conv3x3(inplanes, planes, stride)
self.bn_b1 = nn.BatchNorm2d(planes)
self.conv_b2 = conv3x3(planes, planes)
self.bn_b2 = nn.BatchNorm2d(planes)
self.downsample = downsample
self.stride = stride
def forward(self, x):
a, b, residual = x, x, x
a = F.relu(a, inplace=False)
a = self.conv_a1(a)
a = self.bn_a1(a)
a = F.relu(a, inplace=True)
a = self.conv_a2(a)
a = self.bn_a2(a)
b = F.relu(b, inplace=False)
b = self.conv_b1(b)
b = self.bn_b1(b)
b = F.relu(b, inplace=True)
b = self.conv_b2(b)
b = self.bn_b2(b)
ab = shake(a, b, training=self.training)
if self.downsample is not None:
residual = self.downsample(x)
return residual + ab
class Shake(Function):
@classmethod
def forward(cls, ctx, inp1, inp2, training):
assert inp1.size() == inp2.size()
gate_size = [inp1.size()[0], *itertools.repeat(1, inp1.dim() - 1)]
gate = inp1.new(*gate_size)
if training:
gate.uniform_(0, 1)
else:
gate.fill_(0.5)
return inp1 * gate + inp2 * (1. - gate)
@classmethod
def backward(cls, ctx, grad_output):
grad_inp1 = grad_inp2 = grad_training = None
gate_size = [grad_output.size()[0], *itertools.repeat(1,
grad_output.dim() - 1)]
gate = Variable(grad_output.data.new(*gate_size).uniform_(0, 1))
if ctx.needs_input_grad[0]:
grad_inp1 = grad_output * gate
if ctx.needs_input_grad[1]:
grad_inp2 = grad_output * (1 - gate)
assert not ctx.needs_input_grad[2]
return grad_inp1, grad_inp2, grad_training
def shake(inp1, inp2, training=False):
return Shake.apply(inp1, inp2, training)
class ShiftConvDownsample(nn.Module):
def __init__(self, in_channels, out_channels):
super().__init__()
self.relu = nn.ReLU(inplace=True)
self.conv = nn.Conv2d(in_channels=2 * in_channels,
out_channels=out_channels,
kernel_size=1,
groups=2)
self.bn = nn.BatchNorm2d(out_channels)
def forward(self, x):
x = torch.cat((x[:, :, 0::2, 0::2],
x[:, :, 1::2, 1::2]), dim=1)
x = self.relu(x)
x = self.conv(x)
x = self.bn(x)
return x
class ShakeShake26(ResNet32x32):
def __init__(self,num_targets=10):
super().__init__(ShakeShakeBlock,
layers=[4, 4, 4],
channels=96,
downsample='shift_conv', num_targets=num_targets)
class ResNext152(ResNet224x224):
def __init__(self,num_targets=10):
super().__init__(BottleneckBlock,
layers=[3, 8, 36, 3],
channels=32 * 4,
groups=32,
downsample='basic', num_targets=num_targets)