def __init__(self, gate=True, size_arg="A", dropout=False, n_classes=2):
super(MIL_Attention_Softmax_fc, self).__init__()
self.size_dict = {"A": [1024, 512, 256], "B": [1024, 512, 384]}
size = self.size_dict[size_arg]
fc = [nn.Linear(size[0], size[1]), nn.ReLU()]
if dropout:
fc.append(nn.Dropout(0.25))
if gate:
attention_net = Attn_Net_Gated(L=size[1],
D=size[2],
dropout=dropout,
n_classes=1)
else:
attention_net = Attn_Net(L=size[1],
D=size[2],
dropout=dropout,
n_classes=1)
fc.append(attention_net)
self.attention_net = nn.Sequential(*fc)
self.classifier = nn.Linear(size[1], n_classes)
initialize_weights(self)
def __init__(self, gate=True, size_arg="A", dropout=False, n_classes=2):
super(MIL_Attention_Softmax, self).__init__()
self.size_dict = {"A": [1024, 256], "B": [1024, 512]}
size = self.size_dict[size_arg]
if gate:
self.attention_net = Attn_Net_Gated(L=size[0],
D=size[1],
dropout=dropout,
n_classes=1)
else:
self.attention_net = Attn_Net(L=size[0],
D=size[1],
dropout=dropout,
n_classes=1)
self.classifier = nn.Linear(size[0], n_classes)
initialize_weights(self)
def __init__(self,
gate=True,
size_arg="A",
dropout=False,
n_classes=2,
top_k=1):
super(MIL_fc_mc, self).__init__()
assert n_classes > 2
self.size_dict = {"A": [1024, 512], "B": [512, 512]}
size = self.size_dict[size_arg]
fc = [nn.Linear(size[0], size[1]), nn.ReLU()]
if dropout:
fc.append(nn.Dropout(0.25))
self.fc = nn.Sequential(*fc)
self.classifiers = nn.ModuleList(
[nn.Linear(size[1], 1) for i in range(n_classes)])
initialize_weights(self)
self.top_k = top_k
self.n_classes = n_classes
assert self.top_k == 1
def __init__(self, gate=True, size_arg="A", dropout=False):
super(CPC_MIL_Attention, self).__init__()
self.size_dict = {"A": [1024, 256], "B": [1024, 512]}
size = self.size_dict[size_arg]
if gate:
self.attention_net = Attn_Net_Gated(L=size[0],
D=size[1],
dropout=dropout,
n_classes=1)
else:
self.attention_net = Attn_Net(L=size[0],
D=size[1],
dropout=dropout,
n_classes=1)
self.classifier = nn.Sequential(nn.Linear(size[0], 1), nn.Sigmoid())
initialize_weights(
self
) # initialize weights before loading the weights for feature network
def __init__(self,
gate=True,
size_arg="A",
dropout=False,
n_classes=2,
top_k=1):
super(MIL_fc, self).__init__()
assert n_classes == 2
self.size_dict = {"A": [1000, 512], "B": [512, 512]}
size = self.size_dict[size_arg]
fc = [nn.Linear(size[0], size[0]), nn.ReLU()]
if dropout:
fc.append(nn.Dropout(0.25))
fc.extend([nn.Linear(size[0], size[1]), nn.ReLU()])
if dropout:
fc.append(nn.Dropout(0.25))
fc.append(nn.Linear(size[1], n_classes))
self.classifier = nn.Sequential(*fc)
initialize_weights(self)
self.top_k = top_k