def fc_orth(input_shape, num_classes, dense_classifier=False, pretrained=False, L=6, N=100, nonlinearity=nn.ReLU()): def _orthogonal_init(m): if (type(m) == Layers.layers.Linear): torch.nn.init.orthogonal_(m.weight) size = np.prod(input_shape) # Linear feature extractor modules = [nn.Flatten()] modules.append(layers.Linear(size, N)) modules.append(nonlinearity) for i in range(L - 2): modules.append(layers.Linear(N, N)) modules.append(nonlinearity) # Linear classifier if dense_classifier: modules.append(nn.Linear(N, num_classes)) else: modules.append(layers.Linear(N, num_classes)) model = nn.Sequential(*modules) model.apply(_orthogonal_init) # Pretrained model if pretrained: print("WARNING: this model does not have pretrained weights.") return model
def fc(input_shape, num_classes, dense_classifier=False, pretrained=False, L=5, N=256, nonlinearity=nn.ReLU()): size = np.prod(input_shape) # Linear feature extractor modules = [nn.Flatten()] modules.append(layers.Linear(size, N)) modules.append(nonlinearity) for i in range(L - 2): modules.append(layers.Linear(N, N)) modules.append(nonlinearity) # Linear classifier if dense_classifier: modules.append(nn.Linear(N, num_classes)) else: modules.append(layers.Linear(N, num_classes)) model = nn.Sequential(*modules) # Pretrained model if pretrained: print("WARNING: this model does not have pretrained weights.") return model
def __init__(self, features, num_classes=1000, init_weights=True): super(VGG, self).__init__() self.features = features self.avgpool = nn.AdaptiveAvgPool2d((7, 7)) self.classifier = nn.Sequential( layers.Linear(512 * 7 * 7, 4096), nn.ReLU(True), nn.Dropout(), layers.Linear(4096, 4096), nn.ReLU(True), nn.Dropout(), layers.Linear(4096, num_classes), ) if init_weights: self._initialize_weights()
def conv(input_shape, num_classes, dense_classifier=False, pretrained=False, L=3, N=32, nonlinearity=nn.ReLU()): channels, width, height = input_shape # Convolutional feature extractor modules = [] modules.append(layers.Conv2d(channels, N, kernel_size=3, padding=3 // 2)) modules.append(nonlinearity) for i in range(L - 2): modules.append(layers.Conv2d(N, N, kernel_size=3, padding=3 // 2)) modules.append(nonlinearity) # Linear classifier modules.append(nn.Flatten()) if dense_classifier: modules.append(nn.Linear(N * width * height, num_classes)) else: modules.append(layers.Linear(N * width * height, num_classes)) model = nn.Sequential(*modules) # Pretrained model if pretrained: print("WARNING: this model does not have pretrained weights.") return model
def __init__(self, plan, num_classes, dense_classifier): super(ResNet, self).__init__() # Initial convolution. current_filters = plan[0][0] self.conv = layers.Conv2d( 3, current_filters, kernel_size=3, stride=1, padding=1, bias=False ) self.bn = layers.BatchNorm2d(current_filters) # The subsequent blocks of the ResNet. blocks = [] for segment_index, (filters, num_blocks) in enumerate(plan): for block_index in range(num_blocks): downsample = segment_index > 0 and block_index == 0 blocks.append(Block(current_filters, filters, downsample)) current_filters = filters self.blocks = nn.Sequential(*blocks) self.fc = layers.Linear(plan[-1][0], num_classes) if dense_classifier: self.fc = nn.Linear(plan[-1][0], num_classes) self._initialize_weights()
def __init__(self, plan, conv, num_classes=10, dense_classifier=False, path=False): super(VGG, self).__init__() layer_list = [] filters = 3 self.path = path for spec in plan: if spec == 'M': if self.path: layer_list.append(nn.LPPool2d(1.0, kernel_size=2, stride=2)) else: layer_list.append(nn.MaxPool2d(kernel_size=2, stride=2)) else: layer_list.append(conv(filters, spec)) filters = spec self.layers = nn.Sequential(*layer_list) self.fc = layers.Linear(512, num_classes) if dense_classifier: self.fc = nn.Linear(512, num_classes) self._initialize_weights()
def __init__(self, block, num_block, base_width, num_classes=200, dense_classifier=False): super().__init__() self.in_channels = 64 self.conv1 = nn.Sequential( layers.Conv2d(3, 64, kernel_size=3, padding=1, bias=False), layers.BatchNorm2d(64), nn.ReLU(inplace=True)) #we use a different inputsize than the original paper #so conv2_x's stride is 1 self.conv2_x = self._make_layer(block, 64, num_block[0], 1, base_width) self.conv3_x = self._make_layer(block, 128, num_block[1], 2, base_width) self.conv4_x = self._make_layer(block, 256, num_block[2], 2, base_width) self.conv5_x = self._make_layer(block, 512, num_block[3], 2, base_width) self.avg_pool = nn.AdaptiveAvgPool2d((1, 1)) self.fc = layers.Linear(512 * block.expansion, num_classes) if dense_classifier: self.fc = nn.Linear(512 * block.expansion, num_classes) self._initialize_weights()
def __init__(self, block, layer_list, num_classes=1000, zero_init_residual=False, groups=1, width_per_group=64, replace_stride_with_dilation=None, norm_layer=None): super(ResNet, self).__init__() if norm_layer is None: norm_layer = layers.BatchNorm2d self._norm_layer = norm_layer self.inplanes = 64 self.dilation = 1 if replace_stride_with_dilation is None: # each element in the tuple indicates if we should replace # the 2x2 stride with a dilated convolution instead replace_stride_with_dilation = [False, False, False] if len(replace_stride_with_dilation) != 3: raise ValueError("replace_stride_with_dilation should be None " "or a 3-element tuple, got {}".format( replace_stride_with_dilation)) self.groups = groups self.base_width = width_per_group self.conv1 = layers.Conv2d(3, self.inplanes, kernel_size=7, stride=2, padding=3, bias=False) self.bn1 = norm_layer(self.inplanes) self.relu = nn.ReLU(inplace=True) self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) self.layer1 = self._make_layer(block, 64, layer_list[0]) self.layer2 = self._make_layer(block, 128, layer_list[1], stride=2, dilate=replace_stride_with_dilation[0]) self.layer3 = self._make_layer(block, 256, layer_list[2], stride=2, dilate=replace_stride_with_dilation[1]) self.layer4 = self._make_layer(block, 512, layer_list[3], stride=2, dilate=replace_stride_with_dilation[2]) self.avgpool = nn.AdaptiveAvgPool2d((1, 1)) self.fc = layers.Linear(512 * block.expansion, num_classes) for m in self.modules(): if isinstance(m, layers.Conv2d): nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') elif isinstance(m, (layers.BatchNorm2d, nn.GroupNorm)): nn.init.constant_(m.weight, 1) nn.init.constant_(m.bias, 0) # Zero-initialize the last BN in each residual branch, # so that the residual branch starts with zeros, and each residual block behaves like an identity. # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677 if zero_init_residual: for m in self.modules(): if isinstance(m, Bottleneck): nn.init.constant_(m.bn3.weight, 0) elif isinstance(m, BasicBlock): nn.init.constant_(m.bn2.weight, 0)