def r2plus1d_34(num_classes, pretrained=False, progress=False, arch=None):
model = VideoResNet(block=BasicBlock,
conv_makers=[Conv2Plus1D] * 4,
layers=[3, 4, 6, 3],
stem=R2Plus1dStem)
model.fc = nn.Linear(model.fc.in_features, out_features=num_classes)
# Fix difference in PyTorch vs Caffe2 architecture
# https://github.com/facebookresearch/VMZ/issues/89
# https://github.com/pytorch/vision/issues/1265
model.layer2[0].conv2[0] = Conv2Plus1D(128, 128, 288)
model.layer3[0].conv2[0] = Conv2Plus1D(256, 256, 576)
model.layer4[0].conv2[0] = Conv2Plus1D(512, 512, 1152)
# We need exact Caffe2 momentum for BatchNorm scaling
for m in model.modules():
if isinstance(m, nn.BatchNorm3d):
m.eps = 1e-3
m.momentum = 0.9
if pretrained:
state_dict = torch.hub.load_state_dict_from_url(model_urls[arch],
progress=progress)
model.load_state_dict(state_dict)
return model
def __init__(self, inplanes, planes, scale_factor=2): super(Refine2plus1d, self).__init__(inplanes, planes, scale_factor) self.convFS1 = Conv2Plus1D(inplanes, planes, planes*2 + 32) self.convFS2 = Conv2Plus1D(planes, planes, planes*2 + 32) self.convFS3 = Conv2Plus1D(planes, planes, planes*2 + 32) self.convMM1 = Conv2Plus1D(planes, planes, planes*2 + 32) self.convMM2 = Conv2Plus1D(planes, planes, planes*2 + 32)
def __init__(self, num_classes):
super().__init__(num_classes=2)
self.merge_conv: nn.Module = nn.Sequential(
Conv2Plus1D(128, 64, 144, 1), nn.BatchNorm3d(64),
nn.ReLU(inplace=True))
self.upsample = nn.Upsample(size=(16, 28, 28))
self.audio_conv = nn.Conv3d(256, 128, (4, 1, 1), (4, 1, 1), 0)
self.merge_conv: nn.Module = nn.Sequential(
Conv2Plus1D(256, 128, 230, 1), nn.BatchNorm3d(128),
nn.ReLU(inplace=True))
def r2plus1d_34(cfg):
model = VideoResNet(block=BasicBlock,
conv_makers=[Conv2Plus1D] * 4,
layers=[3, 4, 6, 3],
stem=R2Plus1dStem)
model.layer2[0].conv2[0] = Conv2Plus1D(128, 128, 288)
model.layer3[0].conv2[0] = Conv2Plus1D(256, 256, 576)
model.layer4[0].conv2[0] = Conv2Plus1D(512, 512, 1152)
model.fc = nn.Linear(model.fc.in_features, 359)
state_dict = torch.load('./logs/r2plus1d/pretrained.pth')
model.load_state_dict(state_dict)
model.fc = nn.Linear(model.fc.in_features, cfg.CONFIG.DATA.NUM_CLASSES)
print('Pretrained Model Weight Loaded')
return model
def r2plus1d_34(num_classes):
model = VideoResNet(block=BasicBlock,
conv_makers=[Conv2Plus1D] * 4,
layers=[3, 4, 6, 3],
stem=R2Plus1dStem)
model.fc = nn.Linear(model.fc.in_features, out_features=num_classes)
# Fix difference in PyTorch vs Caffe2 architecture
# https://github.com/facebookresearch/VMZ/issues/89
model.layer2[0].conv2[0] = Conv2Plus1D(128, 128, 288)
model.layer3[0].conv2[0] = Conv2Plus1D(256, 256, 576)
model.layer4[0].conv2[0] = Conv2Plus1D(512, 512, 1152)
# We need exact Caffe2 momentum for BatchNorm scaling
for m in model.modules():
if isinstance(m, nn.BatchNorm3d):
m.eps = 1e-3
m.momentum = 0.9
return model
def __init__(self, num_classes=5, pretrained=True):
super().__init__(num_classes=num_classes,
sequence_length=8,
contains_dropout=False)
self.r2plus1 = r2plus1d_18(pretrained=True)
self.r2plus1.layer3 = nn.Identity()
self.r2plus1.layer4 = nn.Identity()
self.r2plus1.fc = nn.Identity()
self.sync_net = PretrainedSyncNet()
self._set_requires_grad_for_module(self.sync_net, requires_grad=False)
self.relu = nn.ReLU()
self.padding = nn.ReflectionPad2d((0, 1, 0, 0))
self.upsample = nn.Upsample(size=(8, 56, 56))
self.merge_conv: nn.Module = nn.Sequential(
Conv2Plus1D(128, 64, 144, 1), nn.BatchNorm3d(64),
nn.ReLU(inplace=True))
self.out = nn.Sequential(nn.Linear(128, 50), nn.ReLU(),
nn.Linear(50, self.num_classes))