"""SlowFast_Network model for Pytorch.

Reference:

SlowFast Networks for Video Recognition Adapted code from: @inproceedings{hara3dcnns, author={Kensho Hara and Hirokatsu Kataoka and Yutaka Satoh}, title={Can Spatiotemporal 3D CNNs Retrace the History of 2D CNNs and ImageNet?}, booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)}, pages={6546--6555}, year={2018}, }. """

import torch import torch.nn as nn import torch.nn.functional as F from torch.autograd import Variable from functools import partial

all = ['resnet50', 'resnet101', 'resnet152', 'resnet200']

def conv3x3x3(in_planes, out_planes, stride=1): # 3x3x3 convolution with padding return nn.Conv3d( in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False)

def downsample_basic_block(x, planes, stride): out = F.avg_pool3d(x, kernel_size=1, stride=stride) zero_pads = torch.Tensor( out.size(0), planes - out.size(1), out.size(2), out.size(3), out.size(4)).zero_() if isinstance(out.data, torch.cuda.FloatTensor): zero_pads = zero_pads.cuda()

  out = Variable(torch.cat([out.data, zero_pads], dim=1))

  return out

class Bottleneck(nn.Module): expansion = 4

  def __init__(self, inplanes, planes, stride=1, downsample=None, head_conv=1):
      super(Bottleneck, self).__init__()
      if head_conv == 1:
          self.conv1 = nn.Conv3d(inplanes, planes, kernel_size=1, bias=False)
          self.bn1 = nn.BatchNorm3d(planes)
      elif head_conv == 3:
          self.conv1 = nn.Conv3d(inplanes, planes, kernel_size=(3, 1, 1), bias=False, padding=(1, 0, 0))
          self.bn1 = nn.BatchNorm3d(planes)
      else:
          raise ValueError("Unsupported head_conv!")
      self.conv2 = nn.Conv3d(
          planes, planes, kernel_size=(1, 3, 3), stride=(1, stride, stride), padding=(0, 1, 1), bias=False)
      self.bn2 = nn.BatchNorm3d(planes)
      self.conv3 = nn.Conv3d(planes, planes * 4, kernel_size=1, bias=False)
      self.bn3 = nn.BatchNorm3d(planes * 4)
      self.relu = nn.ReLU(inplace=True)
      self.downsample = downsample
      self.stride = stride

  def forward(self, x):
      residual = x

      out = self.conv1(x)
      out = self.bn1(out)
      out = self.relu(out)

      out = self.conv2(out)
      out = self.bn2(out)
      out = self.relu(out)

      out = self.conv3(out)
      out = self.bn3(out)

      if self.downsample is not None:
          residual = self.downsample(x)

      out += residual
      out = self.relu(out)

      return out

def get_fine_tuning_parameters(model, ft_begin_index): if ft_begin_index == 0: return model.parameters()

  ft_module_names = []
  for i in range(ft_begin_index, 5):
      ft_module_names.append('layer{}'.format(i))
  ft_module_names.append('fc')

  parameters = []
  for k, v in model.named_parameters():
      for ft_module in ft_module_names:
          if ft_module in k:
              parameters.append({'params': v})
              break
      else:
          parameters.append({'params': v, 'lr': 0.0})

  return parameters

class SlowFast(nn.Module): def init(self, block=Bottleneck, layers=[3, 4, 6, 3], class_num=27, shortcut_type='B', dropout=0.5, alpha=8, beta=0.125): super(SlowFast, self).init() self.alpha = alpha self.beta = beta

      self.fast_inplanes = int(64 * beta)
      fast_inplanes = self.fast_inplanes
      self.fast_conv1 = nn.Conv3d(3, fast_inplanes, kernel_size=(5, 7, 7), stride=(1, 2, 2), padding=(2, 3, 3),
                                  bias=False)
      self.fast_bn1 = nn.BatchNorm3d(8)
      self.fast_relu = nn.ReLU(inplace=True)
      self.fast_maxpool = nn.MaxPool3d(kernel_size=(1, 3, 3), stride=(1, 2, 2), padding=(0, 1, 1))
      self.fast_res1 = self._make_layer_fast(block, 8, layers[0], shortcut_type, head_conv=3)
      self.fast_res2 = self._make_layer_fast(
          block, 16, layers[1], shortcut_type, stride=2, head_conv=3)
      self.fast_res3 = self._make_layer_fast(
          block, 32, layers[2], shortcut_type, stride=2, head_conv=3)
      self.fast_res4 = self._make_layer_fast(
          block, 64, layers[3], shortcut_type, stride=2, head_conv=3)

      self.slow_inplanes = 64
      slow_inplanes = self.slow_inplanes
      self.slow_conv1 = nn.Conv3d(3, slow_inplanes, kernel_size=(1, 7, 7), stride=(1, 2, 2), padding=(0, 3, 3),
                                  bias=False)
      self.slow_bn1 = nn.BatchNorm3d(64)
      self.slow_relu = nn.ReLU(inplace=True)
      self.slow_maxpool = nn.MaxPool3d(kernel_size=(1, 3, 3), stride=(1, 2, 2), padding=(0, 1, 1))
      self.slow_res1 = self._make_layer_slow(block, 64, layers[0], shortcut_type, head_conv=1)
      self.slow_res2 = self._make_layer_slow(
          block, 128, layers[1], shortcut_type, stride=2, head_conv=1)
      self.slow_res3 = self._make_layer_slow(
          block, 256, layers[2], shortcut_type, stride=2, head_conv=1)
      self.slow_res4 = self._make_layer_slow(
          block, 512, layers[3], shortcut_type, stride=2, head_conv=1)

      self.Tconv1 = nn.Conv3d(8, 16, kernel_size=(5, 1, 1), stride=(alpha, 1, 1), padding=(2, 0, 0), bias=False)
      self.Tconv2 = nn.Conv3d(32, 64, kernel_size=(5, 1, 1), stride=(alpha, 1, 1), padding=(2, 0, 0), bias=False)
      self.Tconv3 = nn.Conv3d(64, 128, kernel_size=(5, 1, 1), stride=(alpha, 1, 1), padding=(2, 0, 0), bias=False)
      self.Tconv4 = nn.Conv3d(128, 256, kernel_size=(5, 1, 1), stride=(alpha, 1, 1), padding=(2, 0, 0), bias=False)

      self.dp = nn.Dropout(dropout)
      self.fc = nn.Linear(self.fast_inplanes + self.slow_inplanes, class_num, bias=False)

  def forward(self, input):

      fast, Tc = self.FastPath(input[:, :, ::2, :, :])
      slow = self.SlowPath(input[:, :, ::8, :, :], Tc)
      x = torch.cat([slow, fast], dim=1)
      x = self.dp(x)
      x = self.fc(x)
      return x

  def SlowPath(self, input, Tc):
      x = self.slow_conv1(input)
      x = self.slow_bn1(x)
      x = self.slow_relu(x)
      x = self.slow_maxpool(x)
      x = torch.cat([x, Tc[0]], dim=1)
      x = self.slow_res1(x)
      x = torch.cat([x, Tc[1]], dim=1)
      x = self.slow_res2(x)
      x = torch.cat([x, Tc[2]], dim=1)
      x = self.slow_res3(x)
      x = torch.cat([x, Tc[3]], dim=1)
      x = self.slow_res4(x)
      x = nn.AdaptiveAvgPool3d(1)(x)
      x = x.view(-1, x.size(1))
      return x

  def FastPath(self, input):
      x = self.fast_conv1(input)
      x = self.fast_bn1(x)
      x = self.fast_relu(x)
      x = self.fast_maxpool(x)
      Tc1 = self.Tconv1(x)
      x = self.fast_res1(x)
      Tc2 = self.Tconv2(x)
      x = self.fast_res2(x)
      Tc3 = self.Tconv3(x)
      x = self.fast_res3(x)
      Tc4 = self.Tconv4(x)
      x = self.fast_res4(x)
      x = nn.AdaptiveAvgPool3d(1)(x)
      x = x.view(-1, x.size(1))
      return x, [Tc1, Tc2, Tc3, Tc4]

  def _make_layer_fast(self, block, planes, blocks, shortcut_type, stride=1, head_conv=1):
      downsample = None
      if stride != 1 or self.fast_inplanes != planes * block.expansion:
          if shortcut_type == 'A':
              downsample = partial(
                  downsample_basic_block,
                  planes=planes * block.expansion,
                  stride=stride)
          else:
              downsample = nn.Sequential(
                  nn.Conv3d(
                      self.fast_inplanes,
                      planes * block.expansion,
                      kernel_size=1,
                      stride=(1, stride, stride),
                      bias=False), nn.BatchNorm3d(planes * block.expansion))

      layers = []
      layers.append(block(self.fast_inplanes, planes, stride, downsample, head_conv=head_conv))
      self.fast_inplanes = planes * block.expansion
      for i in range(1, blocks):
          layers.append(block(self.fast_inplanes, planes, head_conv=head_conv))

      return nn.Sequential(*layers)

  def _make_layer_slow(self, block, planes, blocks, shortcut_type, stride=1, head_conv=1):
      downsample = None
      if stride != 1 or self.slow_inplanes != planes * block.expansion:
          if shortcut_type == 'A':
              downsample = partial(
                  downsample_basic_block,
                  planes=planes * block.expansion,
                  stride=stride)
          else:
              downsample = nn.Sequential(
                  nn.Conv3d(
                      self.slow_inplanes + self.slow_inplanes // self.alpha * 2,
                      planes * block.expansion,
                      kernel_size=1,
                      stride=(1, stride, stride),
                      bias=False), nn.BatchNorm3d(planes * block.expansion))

      layers = []
      layers.append(block(self.slow_inplanes + self.slow_inplanes // self.alpha * 2, planes, stride, downsample,
                          head_conv=head_conv))
      self.slow_inplanes = planes * block.expansion
      for i in range(1, blocks):
          layers.append(block(self.slow_inplanes, planes, head_conv=head_conv))

      return nn.Sequential(*layers)

def resnet50(**kwargs): """Constructs a ResNet-50 model. """ model = SlowFast(Bottleneck, [3, 4, 6, 3], **kwargs) return model

def resnet101(**kwargs): """Constructs a ResNet-101 model. """ model = SlowFast(Bottleneck, [3, 4, 23, 3], **kwargs) return model

def resnet152(**kwargs): """Constructs a ResNet-101 model. """ model = SlowFast(Bottleneck, [3, 8, 36, 3], **kwargs) return model

def resnet200(**kwargs): """Constructs a ResNet-101 model. """ model = SlowFast(Bottleneck, [3, 24, 36, 3], **kwargs) return model

if name == "main": num_classes = 174 input_tensor = torch.autograd.Variable(torch.rand(1, 3, 64, 224, 224)) model = resnet152(class_num=num_classes) output = model(input_tensor) print(output.size())

kwangjinoh/SlowFast-Networks

Reference: