def get_match_scores(self, action, random_set):
embeddings = self.embeddings(random_set)
if not self.config.true_embeddings:
embeddings = F.tanh(embeddings)
# compute similarity probability based on L2 norm
diff = pairwise_distances(action, embeddings)
return diff
def get_match_scores(self, action):
# compute similarity probability based on L2 norm
embeddings = self.embeddings
if not self.config.true_embeddings: #TODO
embeddings = F.tanh(embeddings)
# compute similarity probability based on L2 norm
# a^2 + b^2 - 2ab
similarity = - pairwise_distances(action, embeddings) # Negate euclidean to convert diff into similarity score
# compute similarity probability based on dot product
# similarity = torch.mm(action, torch.transpose(embeddings, 0, 1)) # Dot product
return similarity
def get_match_scores(self, action):
#计算负的欧几里得距离,维度与action相同
#self.embeddings属于神经网络的一个参数，那么每一个动作都对应一个self.embeddings参数，还是所有的动作对应的 self.embeddings参数是一样的
# compute similarity probability based on L2 norm
embeddings = self.embeddings
if not self.true_embeddings:
embeddings = F.tanh(embeddings)
# compute similarity probability based on L2 norm
similarity = - pairwise_distances(action, embeddings) # Negate euclidean to convert diff into similarity score
# compute similarity probability based on dot product
# similarity = torch.mm(action, torch.transpose(embeddings, 0, 1)) # Dot product
return similarity
def get_match_scores(self, action):
# compute similarity probability based on L2 norm
embeddings = self.embeddings
if not self.config.true_embeddings:
embeddings = torch.tanh(embeddings)
# compute similarity probability based on L2 norm
similarity = -pairwise_distances(
action, embeddings
) # Negate euclidean to convert diff into similarity score
# compute similarity probability based on dot product
# similarity = torch.mm(action, torch.transpose(embeddings, 0, 1)) # Dot product
# Never choose the actions not in the active set
# Negative infinity ensures that these actions probability evaluates to 0 (e^-inf) during softmax as well
similarity[:, self.action_mask == False] = float(
'-inf') # Dimension = (bacth_size x unmasked number of actions)
return similarity