def __init__(self, num_classes=61, mem_size=512, REGRESSOR=False):
super(SelfSupAttentionModel, self).__init__()
self.num_classes = num_classes
self.resNet = resnetMod.resnet34(True, True)
self.mem_size = mem_size
self.weight_softmax = self.resNet.fc.weight
self.lstm_cell = MyConvLSTMCell(512, mem_size)
self.avgpool = nn.AvgPool2d(7)
self.dropout = nn.Dropout(0.7)
self.fc = nn.Linear(mem_size, self.num_classes)
self.classifier = nn.Sequential(self.dropout, self.fc)
# Adding flag for the regression option
self.REGR = REGRESSOR
#Secondary, self-supervised, task branch
#Relu+conv+flatten+fullyconnected to get a 2*7*7 = 96 length
self.mmapPredictor = nn.Sequential()
self.mmapPredictor.add_module('mmap_relu',nn.ReLU(True))
self.mmapPredictor.add_module('convolution', nn.Conv2d(512, 100, kernel_size=1))
self.mmapPredictor.add_module('flatten',Flatten())
# Different dimensions for the standard selfSup and regSelfSul tasks
if self.REGR == True:
self.mmapPredictor.add_module('fc_2',nn.Linear(100*7*7,7*7))
else:
self.mmapPredictor.add_module('fc_2',nn.Linear(100*7*7,2*7*7))
def __init__(self, num_classes=51, mem_size=512, c_cam_classes=1000):
super(attentionModelLSTA, self).__init__()
self.num_classes = num_classes
self.resNet = resnetMod.resnet34(True, True)
self.mem_size = mem_size
self.lsta_cell = MyConvLSTACell(512, mem_size, c_cam_classes)
self.avgpool = nn.AvgPool2d(7)
self.dropout = nn.Dropout(0.7)
self.fc = nn.Linear(mem_size, self.num_classes)
self.classifier = nn.Sequential(self.dropout, self.fc)
def __init__(self, num_classes=61, mem_size=512):
super(clstm_Model, self).__init__()
self.num_classes = num_classes
self.resNet = resnetMod.resnet34(True, True)
self.mem_size = mem_size
self.lstm_cell = MyConvLSTMCell(512, mem_size)
self.avgpool = nn.AvgPool2d(7)
self.dropout = nn.Dropout(0.7)
self.fc = nn.Linear(mem_size, self.num_classes)
self.classifier = nn.Sequential(self.dropout, self.fc)
def __init__(self, num_classes=61, mem_size=512):
super(attentionModel, self).__init__()
self.num_classes = num_classes
# Initialize the ResNet
self.resNet = resnetMod.resnet34(True, True)
self.mem_size = mem_size
# Get the weighs of the last fc layer of the ResNet,
# we need this to comput the CAM and the attentionMAP
self.weight_softmax = self.resNet.fc.weight
# Initialize the convLSTM
self.lstm_cell = MyConvLSTMCell(512, mem_size)
# Here I initialize another avgpool needed after the convLSTM
self.avgpool = nn.AvgPool2d(7)
self.dropout = nn.Dropout(0.7)
# Here I initialize the last classifier
self.fc = nn.Linear(mem_size, self.num_classes)
self.classifier = nn.Sequential(self.dropout, self.fc)
def __init__(self, num_classes=61, mem_size=512):
super(RegSelfSupAttentionModel, self).__init__()
self.num_classes = num_classes
self.resNet = resnetMod.resnet34(True, True)
self.mem_size = mem_size
self.weight_softmax = self.resNet.fc.weight
self.lstm_cell = MyConvLSTMCell(512, mem_size)
self.avgpool = nn.AvgPool2d(7)
self.dropout = nn.Dropout(0.7)
self.fc = nn.Linear(mem_size, self.num_classes)
self.classifier = nn.Sequential(self.dropout, self.fc)
#Secondary task branch
self.mmapPredictor = nn.Sequential()
self.mmapPredictor.add_module('mmap_relu', nn.ReLU(True))
self.mmapPredictor.add_module('convolution',
nn.Conv2d(512, 100, kernel_size=1))
self.mmapPredictor.add_module('flatten', Flatten())
self.mmapPredictor.add_module('fc_2', nn.Linear(100 * 7 * 7, 7 * 7))