def _make_stack(self, seq_length=4):
self.embedding_dim = embedding_dim = 3
self.vocab_size = vocab_size = 10
self.seq_length = seq_length
def compose_network(inp, inp_dim, outp_dim, vs, name="compose"):
# Just add the two embeddings!
W = T.concatenate([T.eye(outp_dim), T.eye(outp_dim)], axis=0)
return inp.dot(W)
X = T.imatrix("X")
transitions = T.imatrix("transitions")
apply_dropout = T.scalar("apply_dropout")
vs = VariableStore()
self.stack = HardStack(embedding_dim,
embedding_dim,
vocab_size,
seq_length,
compose_network,
IdentityLayer,
apply_dropout,
vs,
X=X,
transitions=transitions,
make_test_fn=True)
# Swap in our own dummy embeddings and weights.
embeddings = np.arange(vocab_size).reshape(
(vocab_size, 1)).repeat(embedding_dim, axis=1)
self.stack.embeddings.set_value(embeddings)
def _make_stack(self, seq_length=4):
self.embedding_dim = embedding_dim = 3
self.vocab_size = vocab_size = 10
self.seq_length = seq_length
def compose_network(inp, inp_dim, outp_dim, vs, name="compose"):
# Just add the two embeddings!
W = T.concatenate([T.eye(outp_dim), T.eye(outp_dim)], axis=0)
return inp.dot(W)
X = T.imatrix("X")
transitions = T.imatrix("transitions")
training_mode = T.scalar("training_mode")
ground_truth_transitions_visible = T.scalar(
"ground_truth_transitions_visible", dtype="int32")
vs = VariableStore()
# Swap in our own dummy embeddings and weights.
initial_embeddings = np.arange(vocab_size).reshape(
(vocab_size, 1)).repeat(embedding_dim, axis=1)
self.stack = HardStack(embedding_dim,
embedding_dim,
vocab_size,
seq_length,
compose_network,
IdentityLayer,
training_mode,
ground_truth_transitions_visible,
vs,
X=X,
transitions=transitions,
make_test_fn=True,
initial_embeddings=initial_embeddings,
use_input_batch_norm=False,
use_input_dropout=False)
