def _concatenate(input, task):
flattened_input = flatten(input, batch=True)
flattened_task = flatten(task, batch=False)
concatenated = torch.cat([
flattened_input.unsqueeze(1),
batchexpand(flattened_task, flattened_input).unsqueeze(1)
], 1)
return concatenated
def compute(self, input, task):
flattened_input = flatten(input, batch=True)
flattened_task = flatten(task, batch=False)
flattened_task = batchexpand(flattened_task, flattened_input)
result = self.bilinear(flattened_input, flattened_task)
if self.batch_norm != None:
result = self.batch_norm(result)
return result
def forward(self, input, task):
concatenated = _concatenate(input, task)
combined = self.connected(concatenated)
combined = func.dropout(combined, training=True)
if self.batch_norm != None:
combined = self.batch_norm(flatten(combined, batch=True)).unsqueeze(1)
result = self.evaluator(combined)
return result
def forward(self, input, task=None):
if task != None and not self.frozen:
self.task_representation = self.task_embedding(task)
input_representation = self.embedding(input)
result = torch.zeros((input_representation.size()[0],
self.task_representation.size()[0]))
result = result.to(input.device)
for idx in range(self.task_representation.size()[0]):
subtask = self.task_representation[idx, :]
subresult = flatten(self.metric(input_representation, subtask))
result[:, idx] = func.sigmoid(subresult)
return result