from pdb import set_trace
import torch.nn.functional as F
import torch.nn as nn
import torch
import math
import torch.utils.model_zoo as model_zoo
__all__ = ['PResNet', 'presnet18', 'presnet34', 'presnet50', 'presnet101', 'presnet152']
act_fn = nn.ReLU
def init_cnn(m):
if getattr(m, 'bias', None) is not None: nn.init.constant_(m.bias, 0)
if isinstance(m, nn.Conv2d): nn.init.kaiming_normal_(m.weight)
elif isinstance(m, nn.Linear): m.weight.data.normal_(0, 0.01)
for l in m.children(): init_cnn(l)
def conv(ni, nf, ks=3, stride=1, bias=False):
return nn.Conv2d(ni, nf, kernel_size=ks, stride=stride, padding=ks//2, bias=bias)
def conv_layer(conv_1st, ni, nf, ks=3, stride=1, zero_bn=False, bias=False):
bn = nn.BatchNorm2d(nf if conv_1st else ni)
nn.init.constant_(bn.weight, 0. if zero_bn else 1.)
res = [act_fn(), bn]
cn = conv(ni, nf, ks, stride=stride, bias=bias)
res.insert(0 if conv_1st else 2, cn)
return nn.Sequential(*res)
def conv_act(*args, **kwargs): return conv_layer(True , *args, **kwargs)
def act_conv(*args, **kwargs): return conv_layer(False, *args, **kwargs)
class BasicBlock(Module):
expansion = 1
def __init__(self, ni, nf, stride=1, downsample=None):
super(BasicBlock, self).__init__()
self.conv1 = act_conv(ni, nf, stride=stride)
self.conv2 = act_conv(nf, nf, zero_bn=True)
self.downsample = downsample
self.stride = stride
def forward(self, x):
identity = x if self.downsample is None else self.downsample(x)
x = self.conv1(x)
x = self.conv2(x)
x += identity
return x
class Bottleneck(Module):
expansion = 4
def __init__(self, ni, nf, stride=1, downsample=None):
super(Bottleneck, self).__init__()
self.conv1 = act_conv(ni, nf, 1)
self.conv2 = act_conv(nf, nf, stride=stride)
self.conv3 = act_conv(nf, nf*self.expansion, 1)
self.downsample = downsample
self.stride = stride
def forward(self, x):
identity = x if self.downsample is None else self.downsample(x)
x = self.conv1(x)
x = self.conv2(x)
x = self.conv3(x)
x += identity
return x
class PResNet(Module):
def __init__(self, block, layers, num_classes=1000):
self.ni = 64
super().__init__()
self.conv1 = conv_act(3, 16, stride=2)
self.conv2 = conv_act(16, 32)
self.conv3 = conv_act(32, 64)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
self.layer1 = self._make_layer(block, 64, layers[0])
self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
self.layer4 = self._make_layer(block, 512, layers[3], stride=2)
ni = 512*block.expansion
self.avgpool = nn.Sequential(
act_fn(), nn.BatchNorm2d(ni), nn.AdaptiveAvgPool2d(1))
self.fc = nn.Linear(ni, num_classes)
init_cnn(self)
def _make_layer(self, block, nf, blocks, stride=1):
downsample = None
if stride != 1 or self.ni != nf*block.expansion:
layers = [act_fn(), nn.BatchNorm2d(self.ni),
nn.AvgPool2d(kernel_size=2)] if stride==2 else []
layers.append(conv(self.ni, nf*block.expansion))
downsample = nn.Sequential(*layers)
layers = [block(self.ni, nf, stride, downsample)]
self.ni = nf*block.expansion
for i in range(1, blocks): layers.append(block(self.ni, nf))
return nn.Sequential(*layers)
def forward(self, x):
x = self.conv1(x)
x = self.conv2(x)
x = self.conv3(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)
x = self.fc(x)
return x
model_urls = dict(presnet34='presnet34', presnet50='presnet50')
def presnet(block, n_layers, name, pre=False, **kwargs):
model = PResNet(block, n_layers, **kwargs)
if pre: model.load_state_dict(torch.load(model_urls[name]))
return model
def presnet18(pretrained=False, **kwargs):
return presnet(BasicBlock, [2, 2, 2, 2], 'presnet18', pre=pretrained, **kwargs)
def presnet34(pretrained=False, **kwargs):
return presnet(BasicBlock, [3, 4, 6, 3], 'presnet34', pre=pretrained, **kwargs)
def presnet50(pretrained=False, **kwargs):
return presnet(Bottleneck, [3, 4, 6, 3], 'presnet50', pre=pretrained, **kwargs)
def presnet101(pretrained=False, **kwargs):
return presnet(Bottleneck, [3, 4, 23, 3], 'presnet101', pre=pretrained, **kwargs)
def presnet152(pretrained=False, **kwargs):
return presnet(Bottleneck, [3, 8, 36, 3], 'presnet152', pre=pretrained, **kwargs)
