class Network(nn.Module):
def __init__(self, inputs,hidden_layers, outputs):
super(Network, self).__init__()
print(hidden_layers)
self.fc1 = nn.Linear(inputs, hidden_layers[0])
self.fc2 = nn.Linear(hidden_layers[0], hidden_layers[1])
self.fc3 = nn.Linear(hidden_layers[1], hidden_layers[2])
self.fc4 = nn.Linear(hidden_layers[2], hidden_layers[3])
self.fc5 = nn.Linear(hidden_layers[3], hidden_layers[4])
self.fc6 = nn.Linear(hidden_layers[4], hidden_layers[5])
self.fc9 = nn.Linear(hidden_layers[5], outputs)
def forward(self, x):
out = torch.relu(self.fc1(x))
out = torch.relu(self.fc2(out))
out = torch.sigmoid(self.fc3(out))
out = torch.sigmoid(self.fc4(out))
out = torch.sigmoid(self.fc5(out))
out = torch.sigmoid(self.fc6(out))
out = torch.sigmoid(self.fc9(out))
return out
inputs = SIZE*SIZE*2
print(inputs)
x = inputs
hidden_layers = []
while x>=2:
x/=4
if(x>=2):
hidden_layers.append(int(x))
print(hidden_layers)
outputs = 1
model = Network(inputs, hidden_layers, outputs)
criterion = nn.MSELoss()
optimizer = torch.optim.SGD(
model.parameters(),
lr=0.0001,
momentum=0.9)'
N_EPOCHS = 100
for epoch in range(N_EPOCHS):
optimizer.zero_grad()
model.train()
for i,(im1, im2, labels) in enumerate(train_dl):
i1 = torch.flatten(im1,1)
i2 = torch.flatten(im2,1)
inp = torch.cat([i1,i2],1)
b_x = Variable(inp) # batch x
b_y = Variable(labels) # batch y
y_ = model(b_x).squeeze()
y_ = (y_>0.5).float()
print(y_)
print(l)
loss = criterion(y_,b_y)
loss = Variable(loss, requires_grad = True)
print(loss.item())
loss.backward()
optimizer.step()You are rewrapping the loss into a new Variable in:
loss = Variable(loss, requires_grad = True)
and are thus breaking the computation graph.
Remove this line of code and all Variables as well, since they are deprecated since PyTorch 0.4.