def forward(self, img1, img2):
#print("img1.shape", img1.shape)
img1 = img1.view(img1.size(0), -1)
#print("img1.shape", img1.shape)
img2 = img2.view(img2.size(0), -1)
out1 = self.policy_single(img1)
#print("out1.shape", out1.shape)
out2 = self.policy_single(img2)
combined = torch.cat(
(out1, out2), dim=1) # attaccate usando asse x (una sopra l'altra)
probs = self.policy_combined(combined)
# sampling
dist = Categorical(probs=probs)
if self.training:
actions = dist.sample()
else:
actions = dist.argmax(dim=1)
logprobs = dist.log_prob(actions)
entropy = dist.entropy()
return probs, actions, logprobs, entropy
def forward(self, img1, img2, mex):
mex = torch.nn.functional.one_hot(mex, num_classes=self.vocab_len)
img1 = img1.view(img1.size(0), -1)
img2 = img2.view(img2.size(0), -1)
out1 = self.policy_single_img(img1)
out2 = self.policy_single_img(img2)
symbol = mex.view(mex.size(0), -1).float()
symbol = self.policy_single_mex(symbol)
out1 = torch.bmm(symbol.view(symbol.size(0), 1, symbol.size(1)),
out1.view(out1.size(0), symbol.size(1),
1)) # un numero per ogni immagine
out2 = torch.bmm(symbol.view(symbol.size(0), 1, symbol.size(1)),
out2.view(out2.size(0), symbol.size(1), 1))
out1 = out1.view(out1.size(0), -1)
out2 = out2.view(out2.size(0), -1)
combined = torch.cat((out1, out2), dim=1)
probs = self.softmax(combined)
dist = Categorical(probs=probs)
if self.training:
actions = dist.sample()
else:
actions = dist.argmax(dim=1)
logprobs = dist.log_prob(actions)
entropy = dist.entropy()
return probs, actions, logprobs, entropy
