import torch as t
import numpy as np
import torch.nn as nn
from torch.nn.utils.weight_norm import weight_norm
from .base import NeuralNetworkModule
def conv1x1(in_planes, out_planes, stride=1):
"""
Create a 1x1 2d convolution block
"""
return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False)
def conv3x3(in_planes, out_planes, stride=1):
"""
Create a 3x3 2d convolution block
"""
return nn.Conv2d(
in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False
)
def conv5x5(in_planes, out_planes, stride=2):
"""
Create a 5x5 2d convolution block
"""
return nn.Conv2d(
in_planes, out_planes, kernel_size=5, stride=stride, padding=2, bias=False
)
def none_norm(*_, **__):
return nn.Sequential()
def cfg(depth, norm="none"):
depth_lst = [18, 34, 50, 101, 152]
if depth not in depth_lst:
raise ValueError(
"Error : Resnet depth should be either " "18, 34, 50, 101, 152"
)
if norm == "batch":
cfg_dict = {
"18": (BasicBlock, [2, 2, 2, 2], {}),
"34": (BasicBlock, [3, 4, 6, 3], {}),
"50": (Bottleneck, [3, 4, 6, 3], {}),
"101": (Bottleneck, [3, 4, 23, 3], {}),
"152": (Bottleneck, [3, 8, 36, 3], {}),
}
elif norm == "weight":
cfg_dict = {
"18": (BasicBlockWN, [2, 2, 2, 2], {}),
"34": (BasicBlockWN, [3, 4, 6, 3], {}),
"50": (BottleneckWN, [3, 4, 6, 3], {}),
"101": (BottleneckWN, [3, 4, 23, 3], {}),
"152": (BottleneckWN, [3, 8, 36, 3], {}),
}
elif norm == "none":
cfg_dict = {
"18": (BasicBlock, [2, 2, 2, 2], {"norm_layer": none_norm}),
"34": (BasicBlock, [3, 4, 6, 3], {"norm_layer": none_norm}),
"50": (Bottleneck, [3, 4, 6, 3], {"norm_layer": none_norm}),
"101": (Bottleneck, [3, 4, 23, 3], {"norm_layer": none_norm}),
"152": (Bottleneck, [3, 8, 36, 3], {"norm_layer": none_norm}),
}
else:
raise ValueError(f'Invalid normalization method: "{norm}"')
return cfg_dict[str(depth)]
class BasicBlock(NeuralNetworkModule):
# Expansion parameter, output will have "expansion * in_planes" depth.
expansion = 1
def __init__(self, in_planes, out_planes, stride=1, norm_layer=None, **__):
"""
Create a basic block of resnet.
Args:
in_planes: Number of input planes.
out_planes: Number of output planes.
stride: Stride of convolution.
"""
if norm_layer is None:
norm_layer = nn.BatchNorm2d
super().__init__()
self.conv1 = conv3x3(in_planes, out_planes, stride)
self.bn1 = norm_layer(out_planes)
self.conv2 = conv3x3(out_planes, self.expansion * out_planes)
self.bn2 = norm_layer(self.expansion * out_planes)
# Create a shortcut from input to output.
# An empty sequential structure means no transformation
# is made on input X.
self.shortcut = nn.Sequential()
self.set_input_module(self.conv1)
# A convolution is needed if we cannot directly add input X to output
# BatchNorm2d produces NaN gradient?
if stride != 1 or in_planes != self.expansion * out_planes:
self.shortcut = nn.Sequential(
conv1x1(in_planes, self.expansion * out_planes, stride),
norm_layer(self.expansion * out_planes),
)
def forward(self, x):
out = t.relu(self.bn1(self.conv1(x)))
out = self.bn2(self.conv2(out))
out += self.shortcut(x)
out = t.relu(out)
return out
class Bottleneck(NeuralNetworkModule):
# Expansion parameter, output will have "expansion * in_planes" depth.
expansion = 4
def __init__(self, in_planes, out_planes, stride=1, norm_layer=None, **__):
"""
Create a bottleneck block of resnet.
Args:
in_planes: Number of input planes.
out_planes: Number of output planes.
stride: Stride of convolution.
"""
if norm_layer is None:
norm_layer = nn.BatchNorm2d
super().__init__()
self.conv1 = conv1x1(in_planes, out_planes)
self.conv2 = conv3x3(out_planes, out_planes, stride=stride)
self.conv3 = conv1x1(out_planes, self.expansion * out_planes)
self.bn1 = norm_layer(out_planes)
self.bn2 = norm_layer(out_planes)
self.bn3 = norm_layer(self.expansion * out_planes)
self.shortcut = nn.Sequential()
self.set_input_module(self.conv1)
if stride != 1 or in_planes != self.expansion * out_planes:
self.shortcut = nn.Sequential(
conv1x1(in_planes, self.expansion * out_planes, stride),
norm_layer(self.expansion * out_planes),
)
def forward(self, x):
out = t.relu(self.bn1(self.conv1(x)))
out = t.relu(self.bn2(self.conv2(out)))
out = self.bn3(self.conv3(out))
out += self.shortcut(x)
out = t.relu(out)
return out
class BasicBlockWN(NeuralNetworkModule):
"""
Basic block with weight normalization
"""
# Expansion parameter, output will have "expansion * in_planes" depth.
expansion = 1
def __init__(self, in_planes, out_planes, stride=1, **__):
"""
Create a basic block of resnet.
Args:
in_planes: Number of input planes.
out_planes: Number of output planes.
stride: Stride of convolution.
"""
super().__init__()
self.conv1 = weight_norm(conv3x3(in_planes, out_planes, stride))
self.conv2 = weight_norm(conv3x3(out_planes, self.expansion * out_planes))
# Create a shortcut from input to output.
# An empty sequential structure means no transformation
# is made on input X.
self.shortcut = nn.Sequential()
self.set_input_module(self.conv1)
# A convolution is needed if we cannot directly add input X to output.
if stride != 1 or in_planes != self.expansion * out_planes:
self.shortcut = nn.Sequential(
weight_norm(conv1x1(in_planes, self.expansion * out_planes, stride)),
)
def forward(self, x):
out = t.relu(self.conv1(x))
out = self.conv2(out)
out += self.shortcut(x)
out = t.relu(out)
return out
class BottleneckWN(NeuralNetworkModule):
"""
Bottleneck block with weight normalization
"""
# expansion parameter, output will have "expansion * in_planes" depth
expansion = 4
def __init__(self, in_planes, out_planes, stride=1, **__):
"""
Create a bottleneck block of resnet.
Args:
in_planes: Number of input planes.
out_planes: Number of output planes.
stride: Stride of convolution.
"""
super().__init__()
self.conv1 = weight_norm(
nn.Conv2d(in_planes, out_planes, kernel_size=1, bias=False)
)
self.conv2 = weight_norm(
nn.Conv2d(
out_planes,
out_planes,
kernel_size=3,
stride=stride,
padding=1,
bias=False,
)
)
self.conv3 = weight_norm(
nn.Conv2d(
out_planes, self.expansion * out_planes, kernel_size=1, bias=False
)
)
self.shortcut = nn.Sequential()
self.set_input_module(self.conv1)
if stride != 1 or in_planes != self.expansion * out_planes:
self.shortcut = nn.Sequential(
weight_norm(conv1x1(in_planes, self.expansion * out_planes, stride)),
)
def forward(self, x):
out = t.relu(self.conv1(x))
out = t.relu(self.conv2(out))
out = self.conv3(out)
out += self.shortcut(x)
out = t.relu(out)
return out
[docs]class ResNet(NeuralNetworkModule):
def __init__(
self,
in_planes: int,
depth: int,
out_planes: int,
out_pool_size=(1, 1),
norm="none",
):
"""
Create a resnet of specified depth.
Args:
in_planes: Number of input planes.
depth: Depth of resnet. Could be one of ``18, 34, 50, 101, 152``.
out_planes: Number of output planes.
out_pool_size: Size of pooling output
norm: Normalization method, could be one of "none", "batch" or
"weight".
"""
super().__init__()
self.norm = norm
self.depth = depth
self.in_planes = in_planes
self.out_planes = out_planes
self.out_pool_size = out_pool_size
self._cur_in_planes = 64
block, num_blocks, kw = cfg(depth, norm)
self.conv1 = nn.Conv2d(
in_planes, 64, kernel_size=7, stride=2, padding=3, bias=False
)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
if norm == "batch":
self.bn1 = nn.BatchNorm2d(64)
self.layer1 = self._make_layer(block, 64, num_blocks[0], 1, kw)
self.layer2 = self._make_layer(block, 128, num_blocks[1], 2, kw)
self.layer3 = self._make_layer(block, 256, num_blocks[2], 2, kw)
self.layer4 = self._make_layer(block, 512, num_blocks[3], 2, kw)
self.base = nn.Sequential(
self.conv1,
self.bn1,
nn.ReLU(),
self.maxpool,
self.layer1,
self.layer2,
self.layer3,
self.layer4,
)
else:
self.layer1 = self._make_layer(block, 64, num_blocks[0], 1, kw)
self.layer2 = self._make_layer(block, 128, num_blocks[1], 2, kw)
self.layer3 = self._make_layer(block, 256, num_blocks[2], 2, kw)
self.layer4 = self._make_layer(block, 512, num_blocks[3], 2, kw)
self.base = nn.Sequential(
self.conv1,
nn.ReLU(),
self.maxpool,
self.layer1,
self.layer2,
self.layer3,
self.layer4,
)
self.fc = nn.Linear(512 * out_pool_size[0] * out_pool_size[1], out_planes)
self.set_input_module(self.conv1)
def _make_layer(self, block, planes, num_blocks, stride, kwargs):
strides = [stride] + [1] * (num_blocks - 1)
layers = []
for stride in strides:
layers.append(block(self._cur_in_planes, planes, stride, **kwargs))
self._cur_in_planes = planes * block.expansion
return nn.Sequential(*layers)
[docs] def forward(self, x):
assert x.shape[2] == 224 and x.shape[3] == 224
x = self.base(x)
kernel_size = (
np.int(np.floor(x.size(2) / self.out_pool_size[0])),
np.int(np.floor(x.size(3) / self.out_pool_size[1])),
)
x = nn.functional.avg_pool2d(x, kernel_size)
x = x.view(x.size(0), -1)
x = self.fc(x)
return x