def forward(self, inputs):
if self.demo_mode:
inputs = reshape_to_indeces(inputs, ((1, 4), (1, 4)),
80).contiguous()
inputs = inputs.reshape((-1, 1, 10, 8))
x = F.relu(self.ct1(inputs))
x = F.relu(self.ct3(x))
x = F.relu(self.ct5(x))
x = F.relu(self.ct6(x))
x = F.relu(self.ctr(x))
x = F.relu(self.c1(x))
x = F.relu(self.cu(x))
x = F.relu(self.ck(x))
x = F.relu(self.cj(x))
###
if self.round_at is not None:
x = x.masked_fill((x >= self.round_at), 1.)
x = x.masked_fill((x < self.round_at), 0.)
x = F.relu(self.maxpool(self.cc2(x)))
x = F.relu(self.maxpool(self.cc3(x)))
x = F.relu(self.maxpool(self.cc4(x)))
x = F.relu(self.maxpool(self.cc5(x)))
x = F.relu(self.cc6(x))
x = x.view((x.shape[0], 2 * 1024, -1)).contiguous()
x = x.mean(-1).contiguous()
x = self.dropout(x)
x = F.relu(self.lin1(x))
x = self.dropout(x)
x = torch.sigmoid(self.lin3(x))
return x
def forward(self, _input):
if self.demo_mode:
_input = reshape_to_indeces(_input, ((1, 8), (1, 8)), 20)
_input = _input.reshape(-1, 20)
out = F.leaky_relu(self.fc(_input))
out = F.relu(self.fc2(out))
out = torch.sigmoid(self.fc3(out))
return out
def forward(self, _input):
# With the following hack, we can use the full data here too -> saving online storage
if self.demo_mode:
_input = reshape_to_indeces(_input, ((1, 4), (1, 4)), 80)
_input = _input.reshape(-1, 80)
out = F.leaky_relu(self.fc(_input))
out = F.relu(self.fc2(out))
out = torch.sigmoid(self.fc3(out))
return out
def forward(self, inputs):
if self.demo_mode:
inputs = reshape_to_indeces(inputs, ((1, 8), (1, 8)),
20).contiguous()
inp = inputs.reshape((-1, 1, 5, 4))
k = F.relu(self.ct1(inp))
k = F.relu(self.ct2(k))
k = F.relu(self.ct3(k))
k = F.relu(self.ct4(k))
k = F.relu(self.details(k))
k = torch.sigmoid(self.details2(k))
return torch.squeeze(k, dim=1)
def forward(self, inputs):
if self.demo_mode:
inputs = reshape_to_indeces(inputs, ((1, 4), (1, 4)),
80).contiguous()
inp = inputs.reshape((-1, 1, 10, 8)).contiguous()
x = F.relu(self.ct1(inp)).contiguous()
x = F.relu(self.ct3(x)).contiguous()
x = F.relu(self.ct5(x)).contiguous()
x = F.relu(self.ct6(x)).contiguous()
x = F.relu(self.ctr(x)).contiguous()
x = F.relu(self.c1(x))
x = F.relu(self.cu(x))
x = F.relu(self.ck(x))
x = torch.sigmoid(self.cj(x))
if self.round_at is not None:
x = x.masked_fill((x >= self.round_at), 1.)
x = x.masked_fill((x < self.round_at), 0.)
return torch.squeeze(x, dim=1)
def forward(self, inputs):
if self.demo_mode:
inputs = reshape_to_indeces(inputs, ((1, 4), (1, 4)),
80).contiguous()
inputs = inputs.reshape((-1, 1, 10, 8))
x = F.relu(self.ct1(inputs))
x = F.relu(self.ct3(x))
x = F.relu(self.ct5(x))
x = F.relu(self.ct6(x))
x = F.relu(self.ctr(x))
x = F.relu(self.c1(x))
x = F.relu(self.cu(x))
x = F.relu(self.ck(x))
x = F.relu(self.cj(x))
###
x = x.repeat(1, 3, 1, 1)
x = self.upsample(x)
x = self.transfer_model(x)
x = F.sigmoid(x)
return x
def forward(self, inputs):
if self.demo_mode:
inputs = reshape_to_indeces(inputs, ((1, 8), (1, 8)),
20).contiguous()
frs = inputs.reshape((-1, 1, 5, 4))
x = F.relu(self.ct1(frs))
x = F.relu(self.ct2(x))
x = F.relu(self.ct3(x))
x = F.relu(self.ct4(x))
x = F.relu(self.details(x))
if self.round_at is not None:
x = x.masked_fill((x >= self.round_at), 1.)
x = x.masked_fill((x < self.round_at), 0.)
# Shape: [1, 8, 127, 111]
# x = self.bn8(x)
x = F.relu(self.c2(x))
x = self.maxpool(x)
x = F.relu(self.c3(x))
x = self.maxpool(x)
x = F.relu(self.c4(x))
x = self.maxpool(x)
x = F.relu(self.c5(x))
x = self.maxpool(x)
x = F.relu(self.c6(x))
x = F.relu(self.c7(x))
# x = self.bn512(x)
x = x.view((x.shape[0], 1024, -1)).contiguous()
x = x.mean(-1).contiguous()
x = F.relu(self.lin1(x))
x = self.dropout(x)
x = torch.sigmoid(self.lin2(x))
return x