def updates_wrt_cost(self, cost, learning_opts):
if self.using_shared_embeddings:
return []
learning_rate = learning_opts.learning_rate
gradient = util.clipped(T.grad(cost=cost, wrt=self.sequence_embeddings))
return [(self.Wx, T.inc_subtensor(self.sequence_embeddings,
-learning_rate * gradient))]
def updates_wrt_cost(self, cost, learning_opts):
if self.using_shared_embeddings:
return []
learning_rate = learning_opts.learning_rate
gradient = util.clipped(T.grad(cost=cost,
wrt=self.sequence_embeddings))
return [(self.Wx,
T.inc_subtensor(self.sequence_embeddings,
-learning_rate * gradient))]
def updates_wrt_cost(self, cost, learning_opts):
if not self.train_embeddings:
return []
# _one_ update for the embedding matrix; regardless of the number of rnns running
# over subslices
gradient = util.clipped(
T.grad(cost=cost, wrt=self.concatenated_sequence_embeddings))
learning_rate = learning_opts.learning_rate
return [(self.shared_embeddings,
T.inc_subtensor(self.concatenated_sequence_embeddings,
-learning_rate * gradient))]
def updates_wrt_cost(self, cost, learning_opts):
if not self.train_embeddings:
return []
# _one_ update for the embedding matrix; regardless of the number of rnns running
# over subslices
gradient = util.clipped(T.grad(cost=cost,
wrt=self.concatenated_sequence_embeddings))
learning_rate = learning_opts.learning_rate
return [(self.shared_embeddings,
T.inc_subtensor(self.concatenated_sequence_embeddings,
-learning_rate * gradient))]
def updates_wrt_cost(self, cost, learning_opts):
gradients = util.clipped(T.grad(cost=cost, wrt=self.dense_params()))
return self.update_fn(self.dense_params(), gradients, learning_opts)
def updates_wrt_cost(self, cost, learning_opts):
gradients = util.clipped(T.grad(cost=cost, wrt=self.dense_params()))
return self.update_fn(self.dense_params(), gradients, learning_opts)