281 lines
10 KiB
Python
281 lines
10 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.distributed as dist
|
|
|
|
from models.switched_conv.switched_conv_hard_routing import SwitchNorm, RouteTop1
|
|
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 ResNetTail(nn.Module):
|
|
def __init__(self, block, num_block, num_classes=100):
|
|
super().__init__()
|
|
|
|
self.in_channels = 64
|
|
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.avg_pool = nn.AdaptiveAvgPool2d((1, 1))
|
|
self.fc = nn.Linear(256 * block.expansion, num_classes)
|
|
|
|
def _make_layer(self, block, out_channels, num_blocks, stride):
|
|
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.conv4_x(x)
|
|
output = self.conv5_x(output)
|
|
output = self.avg_pool(output)
|
|
output = output.view(output.size(0), -1)
|
|
output = self.fc(output)
|
|
|
|
return output
|
|
|
|
|
|
class DropoutNorm(SwitchNorm):
|
|
def __init__(self, group_size, dropout_rate, accumulator_size=256, eps=1e-6):
|
|
super().__init__(group_size, accumulator_size)
|
|
self.accumulator_desired_size = accumulator_size
|
|
self.group_size = group_size
|
|
self.dropout_rate = dropout_rate
|
|
self.register_buffer("accumulator_index", torch.zeros(1, dtype=torch.long, device='cpu'))
|
|
self.register_buffer("accumulator_filled", torch.zeros(1, dtype=torch.long, device='cpu'))
|
|
self.register_buffer("accumulator", torch.zeros(accumulator_size, group_size))
|
|
self.eps = eps
|
|
|
|
def add_norm_to_buffer(self, x):
|
|
flatten_dims = [0] + [k+2 for k in range(len(x.shape)-2)]
|
|
flat = x.mean(dim=flatten_dims)
|
|
|
|
self.accumulator[self.accumulator_index] = flat.detach().clone()
|
|
self.accumulator_index += 1
|
|
if self.accumulator_index >= self.accumulator_desired_size:
|
|
self.accumulator_index *= 0
|
|
if self.accumulator_filled <= 0:
|
|
self.accumulator_filled += 1
|
|
|
|
# Input into forward is a switching tensor of shape (batch,groups,<misc>)
|
|
def forward(self, x: torch.Tensor):
|
|
assert len(x.shape) >= 2
|
|
|
|
if not self.training:
|
|
return x
|
|
|
|
# Only accumulate the "winning" switch slots.
|
|
mask = torch.nn.functional.one_hot(x.argmax(dim=1), num_classes=x.shape[1])
|
|
if len(x.shape) > 2:
|
|
mask = mask.permute(0, 3, 1, 2) # TODO: Make this more extensible.
|
|
xtop = torch.ones_like(x)
|
|
xtop[mask != 1] = 0
|
|
|
|
# Push the accumulator to the right device on the first iteration.
|
|
if self.accumulator.device != xtop.device:
|
|
self.accumulator = self.accumulator.to(xtop.device)
|
|
self.add_norm_to_buffer(xtop)
|
|
|
|
# Reduce across all distributed entities, if needed
|
|
if dist.is_available() and dist.is_initialized():
|
|
dist.all_reduce(self.accumulator, op=dist.ReduceOp.SUM)
|
|
self.accumulator /= dist.get_world_size()
|
|
|
|
# Compute the dropout probabilities. This module is a no-op before the accumulator is initialized.
|
|
if self.accumulator_filled > 0:
|
|
with torch.no_grad():
|
|
probs = torch.mean(self.accumulator, dim=0) * self.dropout_rate
|
|
bs, br = x.shape[:2]
|
|
drop = torch.rand((bs, br), device=x.device) > probs.unsqueeze(0)
|
|
# Ensure that there is always at least one switch left un-dropped out
|
|
fix_blank = (drop.sum(dim=1, keepdim=True) == 0).repeat(1, br)
|
|
drop = drop.logical_or(fix_blank)
|
|
x_dropped = drop * x + ~drop * -1e20
|
|
x = x_dropped
|
|
|
|
return x
|
|
|
|
|
|
class HardRoutingGate(nn.Module):
|
|
def __init__(self, breadth, fade_steps=10000, dropout_rate=.8):
|
|
super().__init__()
|
|
self.norm = DropoutNorm(breadth, dropout_rate, accumulator_size=128)
|
|
self.fade_steps = fade_steps
|
|
self.register_buffer("last_step", torch.zeros(1, dtype=torch.long, device='cpu'))
|
|
|
|
def forward(self, x):
|
|
if self.last_step < self.fade_steps:
|
|
x = torch.randn_like(x) * (self.fade_steps - self.last_step) / self.fade_steps + \
|
|
x * self.last_step / self.fade_steps
|
|
self.last_step = self.last_step + 1
|
|
soft = nn.functional.softmax(self.norm(x), dim=1)
|
|
return RouteTop1.apply(soft)
|
|
|
|
|
|
class ResNet(nn.Module):
|
|
|
|
def __init__(self, block, num_block, num_classes=100, num_tails=8):
|
|
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))
|
|
|
|
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.tails = nn.ModuleList([ResNetTail(block, num_block, 256) for _ in range(num_tails)])
|
|
self.selector = ResNetTail(block, num_block, num_tails)
|
|
self.selector_gate = nn.Linear(256, 1)
|
|
self.gate = HardRoutingGate(num_tails, dropout_rate=2)
|
|
self.final_linear = nn.Linear(256, num_classes)
|
|
|
|
def _make_layer(self, block, out_channels, num_blocks, stride):
|
|
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 get_debug_values(self, step, __):
|
|
logs = {'histogram_switch_usage': self.latest_masks}
|
|
return logs
|
|
|
|
def forward(self, x, coarse_label, return_selector=False):
|
|
output = self.conv1(x)
|
|
output = self.conv2_x(output)
|
|
output = self.conv3_x(output)
|
|
|
|
keys = []
|
|
for t in self.tails:
|
|
keys.append(t(output))
|
|
keys = torch.stack(keys, dim=1)
|
|
|
|
query = self.selector(output).unsqueeze(2)
|
|
selector = self.selector_gate(query * keys).squeeze(-1)
|
|
selector = self.gate(selector)
|
|
self.latest_masks = (selector.max(dim=1, keepdim=True)[0].repeat(1,8) == selector).float().argmax(dim=1)
|
|
values = self.final_linear(selector.unsqueeze(-1) * keys)
|
|
|
|
if return_selector:
|
|
return values.sum(dim=1), selector
|
|
else:
|
|
return values.sum(dim=1)
|
|
|
|
#bs = output.shape[0]
|
|
#return (tailouts[coarse_label] * torch.eye(n=bs, device=x.device).view(bs,bs,1)).sum(dim=1)
|
|
|
|
@register_model
|
|
def register_cifar_resnet18_branched(opt_net, opt):
|
|
""" return a ResNet 18 object
|
|
"""
|
|
return ResNet(BasicBlock, [2, 2, 2, 2])
|
|
|
|
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])
|
|
|
|
|
|
if __name__ == '__main__':
|
|
model = ResNet(BasicBlock, [2,2,2,2])
|
|
for j in range(10):
|
|
v = model(torch.randn(256,3,32,32), None)
|
|
print(model.get_debug_values(0, None))
|
|
print(v.shape)
|
|
l = nn.MSELoss()(v, torch.randn_like(v))
|
|
l.backward()
|
|
|