def __init__(self,
num_classes=2,
numDiPerVideos=1,
model_name='resnet50',
joinType='maxTempPool',
freezeConvLayers=True):
super(ResNet_ROI_Pool, self).__init__()
self.numDiPerVideos = numDiPerVideos
self.num_classes = num_classes
self.joinType = joinType
self.model_ft = None
self.model_ft = models.resnet50(pretrained=True)
self.num_ftrs = self.model_ft.fc.in_features
self.model_ft.fc = Identity()
self.bn = nn.BatchNorm2d(2048)
self.model_ft = nn.Sequential(*list(
self.model_ft.children())[:-2]) # to tempooling
self.roi_pool = RoIPool(3, 1)
# model_ft = None
self.AdaptiveAvgPool2d = nn.AdaptiveAvgPool2d((1, 1))
# set_parameter_requires_grad(self.model_ft, feature_extract)
set_parameter_requires_grad(self.model_ft, freezeConvLayers)
if self.joinType == constants.TEMP_MAX_POOL:
self.linear = nn.Linear(2048, self.num_classes)
def __init__(self, num_classes, numDiPerVideos, model_name, joinType,
freezeConvLayers):
super(ResNet, self).__init__()
self.numDiPerVideos = numDiPerVideos
self.num_classes = num_classes
self.joinType = joinType
self.model_ft = None
if model_name == 'resnet18':
self.model_ft = models.resnet18(pretrained=True)
elif model_name == 'resnet34':
self.model_ft = models.resnet34(pretrained=True)
elif model_name == 'resnet50':
self.model_ft = models.resnet50(pretrained=True)
self.num_ftrs = self.model_ft.fc.in_features
self.model_ft.fc = Identity()
if model_name == 'resnet18' or model_name == 'resnet34':
self.bn = nn.BatchNorm2d(512)
elif model_name == 'resnet50':
self.bn = nn.BatchNorm2d(2048)
self.convLayers = nn.Sequential(*list(
self.model_ft.children())[:-2]) # to tempooling
model_ft = None
self.AdaptiveAvgPool2d = nn.AdaptiveAvgPool2d((1, 1))
# set_parameter_requires_grad(self.model_ft, feature_extract)
set_parameter_requires_grad(self.convLayers, freezeConvLayers)
if self.joinType == constants.TEMP_MAX_POOL or self.joinType == constants.MULT_TEMP_POOL or self.joinType == constants.TEMP_AVG_POOL or self.joinType == constants.TEMP_STD_POOL:
if model_name == 'resnet18' or model_name == 'resnet34':
self.linear = nn.Linear(512, self.num_classes)
elif model_name == 'resnet50':
self.linear = nn.Linear(2048, self.num_classes)
def __init__(self, num_classes, numDiPerVideos, joinType, feature_extract):
super(Densenet, self).__init__()
self.numDiPerVideos = numDiPerVideos
self.joinType = joinType
self.num_classes = num_classes
self.model = models.densenet121(pretrained=True)
set_parameter_requires_grad(self.model, feature_extract)
self.num_ftrs = self.model.classifier.in_features
self.model.classifier = Identity()
self.linear = nn.Linear(self.num_ftrs, num_classes)
def __init__(self, num_classes, model_name, feature_extract):
super(ResNetRGB, self).__init__()
self.num_classes = num_classes
if model_name == 'resnet18':
self.model_ft = models.resnet18(pretrained=True)
elif model_name == 'resnet34':
self.model_ft = models.resnet34(pretrained=True)
elif model_name == 'resnet50':
self.model_ft = models.resnet50(pretrained=True)
self.num_ftrs = self.model_ft.fc.in_features
set_parameter_requires_grad(self.model_ft, feature_extract)
self.model_ft.fc = nn.Linear(self.num_ftrs, self.num_classes)
def __init__(self, num_classes, numDiPerVideos, joinType, feature_extract):
super(AlexNet, self).__init__()
self.numDiPerVideos = numDiPerVideos
self.joinType = joinType
self.num_classes = num_classes
self.model = models.alexnet(pretrained=True)
set_parameter_requires_grad(self.model, feature_extract)
self.linear = None
self.model = nn.Sequential(*list(
self.model.children())[:-2]) # to tempooling
self.tmpPooling = nn.MaxPool2d((numDiPerVideos, 1))
self.linear = nn.Linear(256 * 6 * 6, self.num_classes)
def __init__(self, num_classes, numDiPerVideos, joinType, feature_extract):
super(Inception, self).__init__()
self.numDiPerVideos = numDiPerVideos
self.joinType = joinType
self.num_classes = num_classes
self.model = models.inception_v3(pretrained=True)
set_parameter_requires_grad(self.model, feature_extract)
# Handle the auxilary net
num_ftrs = self.model.AuxLogits.fc.in_features
self.model.AuxLogits.fc = nn.Linear(num_ftrs, num_classes)
# Handle the primary net
num_ftrs = self.model.fc.in_features
self.model.fc = nn.Linear(num_ftrs, num_classes)