|
from torch.nn import Linear, Conv2d, BatchNorm1d, BatchNorm2d, PReLU, ReLU, Sigmoid, Dropout2d, Dropout, AvgPool2d, \ |
|
MaxPool2d, AdaptiveAvgPool2d, Sequential, Module, Parameter |
|
import torch.nn.functional as F |
|
import torch |
|
from collections import namedtuple |
|
import math |
|
import pdb |
|
|
|
|
|
|
|
|
|
class Flatten(Module): |
|
def forward(self, input): |
|
return input.view(input.size(0), -1) |
|
|
|
|
|
def l2_norm(input, axis=1): |
|
norm = torch.norm(input, 2, axis, True) |
|
output = torch.div(input, norm) |
|
return output |
|
|
|
|
|
class SEModule(Module): |
|
def __init__(self, channels, reduction): |
|
super(SEModule, self).__init__() |
|
self.avg_pool = AdaptiveAvgPool2d(1) |
|
self.fc1 = Conv2d( |
|
channels, channels // reduction, kernel_size=1, padding=0, bias=False) |
|
self.relu = ReLU(inplace=True) |
|
self.fc2 = Conv2d( |
|
channels // reduction, channels, kernel_size=1, padding=0, bias=False) |
|
self.sigmoid = Sigmoid() |
|
|
|
def forward(self, x): |
|
module_input = x |
|
x = self.avg_pool(x) |
|
x = self.fc1(x) |
|
x = self.relu(x) |
|
x = self.fc2(x) |
|
x = self.sigmoid(x) |
|
return module_input * x |
|
|
|
|
|
|
|
|
|
class bottleneck_IR(Module): |
|
def __init__(self, in_channel, depth, stride): |
|
super(bottleneck_IR, self).__init__() |
|
if in_channel == depth: |
|
self.shortcut_layer = MaxPool2d(1, stride) |
|
else: |
|
self.shortcut_layer = Sequential( |
|
Conv2d(in_channel, depth, (1, 1), stride, bias=False), BatchNorm2d(depth)) |
|
self.res_layer = Sequential( |
|
BatchNorm2d(in_channel), |
|
Conv2d(in_channel, depth, (3, 3), (1, 1), 1, bias=False), PReLU(depth), |
|
Conv2d(depth, depth, (3, 3), stride, 1, bias=False), BatchNorm2d(depth)) |
|
i = 0 |
|
|
|
def forward(self, x): |
|
shortcut = self.shortcut_layer(x) |
|
|
|
|
|
res = self.res_layer(x) |
|
|
|
|
|
|
|
|
|
|
|
return res + shortcut |
|
|
|
|
|
class bottleneck_IR_SE(Module): |
|
def __init__(self, in_channel, depth, stride): |
|
super(bottleneck_IR_SE, self).__init__() |
|
if in_channel == depth: |
|
self.shortcut_layer = MaxPool2d(1, stride) |
|
else: |
|
self.shortcut_layer = Sequential( |
|
Conv2d(in_channel, depth, (1, 1), stride, bias=False), |
|
BatchNorm2d(depth)) |
|
self.res_layer = Sequential( |
|
BatchNorm2d(in_channel), |
|
Conv2d(in_channel, depth, (3, 3), (1, 1), 1, bias=False), |
|
PReLU(depth), |
|
Conv2d(depth, depth, (3, 3), stride, 1, bias=False), |
|
BatchNorm2d(depth), |
|
SEModule(depth, 16) |
|
) |
|
|
|
def forward(self, x): |
|
shortcut = self.shortcut_layer(x) |
|
res = self.res_layer(x) |
|
return res + shortcut |
|
|
|
|
|
class Bottleneck(namedtuple('Block', ['in_channel', 'depth', 'stride'])): |
|
'''A named tuple describing a ResNet block.''' |
|
|
|
|
|
|
|
def get_block(in_channel, depth, num_units, stride=2): |
|
return [Bottleneck(in_channel, depth, stride)] + [Bottleneck(depth, depth, 1) for i in range(num_units - 1)] |
|
|
|
|
|
def get_blocks(num_layers): |
|
if num_layers == 50: |
|
blocks1 = [ |
|
get_block(in_channel=64, depth=64, num_units=3), |
|
|
|
|
|
|
|
] |
|
blocks2 = [ |
|
|
|
get_block(in_channel=64, depth=128, num_units=4), |
|
|
|
|
|
] |
|
blocks3 = [ |
|
|
|
|
|
get_block(in_channel=128, depth=256, num_units=14), |
|
|
|
] |
|
|
|
elif num_layers == 100: |
|
blocks = [ |
|
get_block(in_channel=64, depth=64, num_units=3), |
|
get_block(in_channel=64, depth=128, num_units=13), |
|
get_block(in_channel=128, depth=256, num_units=30), |
|
get_block(in_channel=256, depth=512, num_units=3) |
|
] |
|
elif num_layers == 152: |
|
blocks = [ |
|
get_block(in_channel=64, depth=64, num_units=3), |
|
get_block(in_channel=64, depth=128, num_units=8), |
|
get_block(in_channel=128, depth=256, num_units=36), |
|
get_block(in_channel=256, depth=512, num_units=3) |
|
] |
|
return blocks1, blocks2, blocks3 |
|
|
|
|
|
class Backbone(Module): |
|
def __init__(self, num_layers, drop_ratio, mode='ir'): |
|
super(Backbone, self).__init__() |
|
|
|
assert mode in ['ir', 'ir_se'], 'mode should be ir or ir_se' |
|
blocks1, blocks2, blocks3 = get_blocks(num_layers) |
|
|
|
if mode == 'ir': |
|
unit_module = bottleneck_IR |
|
elif mode == 'ir_se': |
|
unit_module = bottleneck_IR_SE |
|
self.input_layer = Sequential(Conv2d(3, 64, (3, 3), 1, 1, bias=False), |
|
BatchNorm2d(64), |
|
PReLU(64)) |
|
self.output_layer = Sequential(BatchNorm2d(512), |
|
Dropout(drop_ratio), |
|
Flatten(), |
|
Linear(512 * 7 * 7, 512), |
|
BatchNorm1d(512)) |
|
modules1 = [] |
|
for block in blocks1: |
|
for bottleneck in block: |
|
modules1.append( |
|
unit_module(bottleneck.in_channel, |
|
bottleneck.depth, |
|
bottleneck.stride)) |
|
|
|
modules2 = [] |
|
for block in blocks2: |
|
for bottleneck in block: |
|
modules2.append( |
|
unit_module(bottleneck.in_channel, |
|
bottleneck.depth, |
|
bottleneck.stride)) |
|
|
|
modules3 = [] |
|
for block in blocks3: |
|
for bottleneck in block: |
|
modules3.append( |
|
unit_module(bottleneck.in_channel, |
|
bottleneck.depth, |
|
bottleneck.stride)) |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
self.body1 = Sequential(*modules1) |
|
self.body2 = Sequential(*modules2) |
|
self.body3 = Sequential(*modules3) |
|
|
|
|
|
def forward(self, x): |
|
x = F.interpolate(x, size=112) |
|
x = self.input_layer(x) |
|
x1 = self.body1(x) |
|
x2 = self.body2(x1) |
|
x3 = self.body3(x2) |
|
|
|
|
|
|
|
|
|
return x1, x2, x3 |
|
|
|
def load_pretrained_weights(model, checkpoint): |
|
import collections |
|
if 'state_dict' in checkpoint: |
|
state_dict = checkpoint['state_dict'] |
|
else: |
|
state_dict = checkpoint |
|
model_dict = model.state_dict() |
|
new_state_dict = collections.OrderedDict() |
|
matched_layers, discarded_layers = [], [] |
|
for i, (k, v) in enumerate(state_dict.items()): |
|
|
|
|
|
|
|
|
|
if k.startswith('module.'): |
|
k = k[7:] |
|
if k in model_dict and model_dict[k].size() == v.size(): |
|
|
|
new_state_dict[k] = v |
|
matched_layers.append(k) |
|
else: |
|
|
|
discarded_layers.append(k) |
|
|
|
model_dict.update(new_state_dict) |
|
model.load_state_dict(model_dict) |
|
print('load_weight', len(matched_layers)) |
|
return model |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|