240 lines
8.3 KiB
Python
240 lines
8.3 KiB
Python
"""resnet in pytorch
|
|
|
|
|
|
|
|
[1] Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun.
|
|
|
|
Deep Residual Learning for Image Recognition
|
|
https://arxiv.org/abs/1512.03385v1
|
|
"""
|
|
|
|
import torch
|
|
import torch.nn as nn
|
|
import torch.nn.functional as F
|
|
|
|
from trainer.networks import register_model
|
|
|
|
|
|
class BasicBlock(nn.Module):
|
|
"""Basic Block for resnet 18 and resnet 34
|
|
|
|
"""
|
|
|
|
#BasicBlock and BottleNeck block
|
|
#have different output size
|
|
#we use class attribute expansion
|
|
#to distinct
|
|
expansion = 1
|
|
|
|
def __init__(self, in_channels, out_channels, stride=1):
|
|
super().__init__()
|
|
|
|
#residual function
|
|
self.residual_function = nn.Sequential(
|
|
nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=stride, padding=1, bias=False),
|
|
nn.BatchNorm2d(out_channels),
|
|
nn.ReLU(inplace=True),
|
|
nn.Conv2d(out_channels, out_channels * BasicBlock.expansion, kernel_size=3, padding=1, bias=False),
|
|
nn.BatchNorm2d(out_channels * BasicBlock.expansion)
|
|
)
|
|
|
|
#shortcut
|
|
self.shortcut = nn.Sequential()
|
|
|
|
#the shortcut output dimension is not the same with residual function
|
|
#use 1*1 convolution to match the dimension
|
|
if stride != 1 or in_channels != BasicBlock.expansion * out_channels:
|
|
self.shortcut = nn.Sequential(
|
|
nn.Conv2d(in_channels, out_channels * BasicBlock.expansion, kernel_size=1, stride=stride, bias=False),
|
|
nn.BatchNorm2d(out_channels * BasicBlock.expansion)
|
|
)
|
|
|
|
def forward(self, x):
|
|
return nn.ReLU(inplace=True)(self.residual_function(x) + self.shortcut(x))
|
|
|
|
|
|
class BottleNeck(nn.Module):
|
|
"""Residual block for resnet over 50 layers
|
|
|
|
"""
|
|
expansion = 4
|
|
def __init__(self, in_channels, out_channels, stride=1):
|
|
super().__init__()
|
|
self.residual_function = nn.Sequential(
|
|
nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False),
|
|
nn.BatchNorm2d(out_channels),
|
|
nn.ReLU(inplace=True),
|
|
nn.Conv2d(out_channels, out_channels, stride=stride, kernel_size=3, padding=1, bias=False),
|
|
nn.BatchNorm2d(out_channels),
|
|
nn.ReLU(inplace=True),
|
|
nn.Conv2d(out_channels, out_channels * BottleNeck.expansion, kernel_size=1, bias=False),
|
|
nn.BatchNorm2d(out_channels * BottleNeck.expansion),
|
|
)
|
|
|
|
self.shortcut = nn.Sequential()
|
|
|
|
if stride != 1 or in_channels != out_channels * BottleNeck.expansion:
|
|
self.shortcut = nn.Sequential(
|
|
nn.Conv2d(in_channels, out_channels * BottleNeck.expansion, stride=stride, kernel_size=1, bias=False),
|
|
nn.BatchNorm2d(out_channels * BottleNeck.expansion)
|
|
)
|
|
|
|
def forward(self, x):
|
|
return nn.ReLU(inplace=True)(self.residual_function(x) + self.shortcut(x))
|
|
|
|
|
|
class ResNet(nn.Module):
|
|
|
|
def __init__(self, block, num_block, num_classes=100):
|
|
super().__init__()
|
|
|
|
self.in_channels = 32
|
|
|
|
self.conv1 = nn.Sequential(
|
|
nn.Conv2d(3, 32, kernel_size=3, padding=1, bias=False),
|
|
nn.BatchNorm2d(32),
|
|
nn.ReLU(inplace=True))
|
|
#we use a different inputsize than the original paper
|
|
#so conv2_x's stride is 1
|
|
self.conv2_x = self._make_layer(block, 32, num_block[0], 1)
|
|
self.conv3_x = self._make_layer(block, 64, num_block[1], 2)
|
|
self.conv4_x = self._make_layer(block, 128, num_block[2], 2)
|
|
self.conv5_x = self._make_layer(block, 256, num_block[3], 2)
|
|
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
|
|
self.fc = nn.Linear(256 * block.expansion, num_classes)
|
|
|
|
def _make_layer(self, block, out_channels, num_blocks, stride):
|
|
"""make resnet layers(by layer i didnt mean this 'layer' was the
|
|
same as a neuron netowork layer, ex. conv layer), one layer may
|
|
contain more than one residual block
|
|
|
|
Args:
|
|
block: block type, basic block or bottle neck block
|
|
out_channels: output depth channel number of this layer
|
|
num_blocks: how many blocks per layer
|
|
stride: the stride of the first block of this layer
|
|
|
|
Return:
|
|
return a resnet layer
|
|
"""
|
|
|
|
# we have num_block blocks per layer, the first block
|
|
# could be 1 or 2, other blocks would always be 1
|
|
strides = [stride] + [1] * (num_blocks - 1)
|
|
layers = []
|
|
for stride in strides:
|
|
layers.append(block(self.in_channels, out_channels, stride))
|
|
self.in_channels = out_channels * block.expansion
|
|
|
|
return nn.Sequential(*layers)
|
|
|
|
def forward(self, x):
|
|
output = self.conv1(x)
|
|
output = self.conv2_x(output)
|
|
output = self.conv3_x(output)
|
|
output = self.conv4_x(output)
|
|
output = self.conv5_x(output)
|
|
output = self.avgpool(output)
|
|
output = output.view(output.size(0), -1)
|
|
output = self.fc(output)
|
|
|
|
return output
|
|
|
|
|
|
class SymbolicLoss:
|
|
def __init__(self, category_depths=[3,5,5,3], convergence_weighting=[1,.6,.3,.1], divergence_weighting=[.1,.3,.6,1]):
|
|
self.depths = category_depths
|
|
self.total_classes = 1
|
|
for c in category_depths:
|
|
self.total_classes *= c
|
|
self.elements_per_level = []
|
|
m = 1
|
|
for c in category_depths[1:]:
|
|
m *= c
|
|
self.elements_per_level.append(self.total_classes // m)
|
|
self.elements_per_level = self.elements_per_level + [1]
|
|
self.convergence_weighting = convergence_weighting
|
|
self.divergence_weighting = divergence_weighting
|
|
# TODO: improve the above logic, I'm sure it can be done better.
|
|
|
|
def __call__(self, logits, collaboratorLabels):
|
|
"""
|
|
Computes the symbolic loss.
|
|
:param logits: Nested level scores for the network under training.
|
|
:param collaboratorLabels: level labels from the collaborator network.
|
|
:return: Convergence loss & divergence loss.
|
|
"""
|
|
b, l = logits.shape
|
|
assert l == self.total_classes, f"Expected {self.total_classes} predictions, got {l}"
|
|
|
|
convergence_loss = 0
|
|
divergence_loss = 0
|
|
for epc, cw, dw in zip(self.elements_per_level, self.convergence_weighting, self.divergence_weighting):
|
|
level_logits = logits.view(b, l//epc, epc)
|
|
level_logits = level_logits.sum(dim=-1)
|
|
level_labels = collaboratorLabels.div(epc, rounding_mode='trunc')
|
|
# Convergence
|
|
convergence_loss = convergence_loss + F.cross_entropy(level_logits, level_labels) * cw
|
|
# Divergence
|
|
div_label_indices = level_logits.argmax(dim=-1)
|
|
# TODO: find the torch-y way of doing this.
|
|
dp = []
|
|
for bi, i in enumerate(div_label_indices):
|
|
dp.append(level_logits[:, i])
|
|
div_preds = torch.stack(dp, dim=0)
|
|
div_labels = torch.arange(0, b, device=logits.device)
|
|
divergence_loss = divergence_loss + F.cross_entropy(div_preds, div_labels)
|
|
return convergence_loss, divergence_loss
|
|
|
|
|
|
if __name__ == '__main__':
|
|
sl = SymbolicLoss()
|
|
logits = torch.randn(5, sl.total_classes)
|
|
labels = torch.randint(0, sl.total_classes, (5,))
|
|
sl(logits, labels)
|
|
|
|
|
|
class TwinnedCifar(nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.loss = SymbolicLoss()
|
|
self.netA = ResNet(BasicBlock, [2, 2, 2, 2], num_classes=self.loss.total_classes)
|
|
self.netB = ResNet(BasicBlock, [2, 2, 2, 2], num_classes=self.loss.total_classes)
|
|
|
|
def forward(self, x):
|
|
y1 = self.netA(x)
|
|
y2 = self.netB(x)
|
|
b = x.shape[0]
|
|
convergenceA, divergenceA = self.loss(y1[:b//2], y2.argmax(dim=-1)[:b//2])
|
|
convergenceB, divergenceB = self.loss(y2[b//2:], y1.argmax(dim=-1)[b//2:])
|
|
return convergenceA + convergenceB, divergenceA + divergenceB
|
|
|
|
|
|
@register_model
|
|
def register_twin_cifar(opt_net, opt):
|
|
""" return a ResNet 18 object
|
|
"""
|
|
return TwinnedCifar()
|
|
|
|
def resnet34():
|
|
""" return a ResNet 34 object
|
|
"""
|
|
return ResNet(BasicBlock, [3, 4, 6, 3])
|
|
|
|
def resnet50():
|
|
""" return a ResNet 50 object
|
|
"""
|
|
return ResNet(BottleNeck, [3, 4, 6, 3])
|
|
|
|
def resnet101():
|
|
""" return a ResNet 101 object
|
|
"""
|
|
return ResNet(BottleNeck, [3, 4, 23, 3])
|
|
|
|
def resnet152():
|
|
""" return a ResNet 152 object
|
|
"""
|
|
return ResNet(BottleNeck, [3, 8, 36, 3])
|
|
|
|
|