def _make_rnn(self, seq_length=4):
self.embedding_dim = embedding_dim = 3
self.vocab_size = vocab_size = 10
self.seq_length = seq_length
def compose_network(h_prev,
inp,
embedding_dim,
model_dim,
vs,
name="compose"):
# Just add the two embeddings!
W = T.concatenate([T.eye(model_dim), T.eye(model_dim)], axis=0)
i = T.concatenate([h_prev, inp], axis=1)
return i.dot(W)
X = T.imatrix("X")
training_mode = T.scalar("training_mode")
vs = VariableStore()
embeddings = np.arange(vocab_size).reshape(
(vocab_size, 1)).repeat(embedding_dim, axis=1)
self.model = RNN(embedding_dim,
embedding_dim,
vocab_size,
seq_length,
compose_network,
IdentityLayer,
training_mode,
None,
vs,
X=X,
make_test_fn=True,
initial_embeddings=embeddings)
class RNNTestCase(unittest.TestCase):
"""Basic functional tests for RNN with dummy data."""
def _make_rnn(self, seq_length=4):
self.embedding_dim = embedding_dim = 3
self.vocab_size = vocab_size = 10
self.seq_length = seq_length
def compose_network(h_prev,
inp,
embedding_dim,
model_dim,
vs,
name="compose"):
# Just add the two embeddings!
W = T.concatenate([T.eye(model_dim), T.eye(model_dim)], axis=0)
i = T.concatenate([h_prev, inp], axis=1)
return i.dot(W)
X = T.imatrix("X")
training_mode = T.scalar("training_mode")
vs = VariableStore()
embeddings = np.arange(vocab_size).reshape(
(vocab_size, 1)).repeat(embedding_dim, axis=1)
self.model = RNN(embedding_dim,
embedding_dim,
vocab_size,
seq_length,
compose_network,
IdentityLayer,
training_mode,
None,
vs,
X=X,
make_test_fn=True,
initial_embeddings=embeddings)
def test_basic_ff(self):
self._make_rnn(4)
X = np.array([[3, 1, 2, 0], [3, 2, 4, 5]], dtype=np.int32)
expected = np.array([[6, 6, 6], [14, 14, 14]])
ret = self.model.scan_fn(X, 1.0)
np.testing.assert_almost_equal(ret, expected)
def _make_rnn(self, seq_length=4):
self.embedding_dim = embedding_dim = 3
self.vocab_size = vocab_size = 10
self.seq_length = seq_length
def compose_network(h_prev, inp, embedding_dim, model_dim, vs, name="compose"):
# Just add the two embeddings!
W = T.concatenate([T.eye(model_dim), T.eye(model_dim)], axis=0)
i = T.concatenate([h_prev, inp], axis=1)
return i.dot(W)
X = T.imatrix("X")
training_mode = T.scalar("training_mode")
vs = VariableStore()
embeddings = np.arange(vocab_size).reshape(
(vocab_size, 1)).repeat(embedding_dim, axis=1)
self.model = RNN(
embedding_dim, embedding_dim, vocab_size, seq_length, compose_network,
IdentityLayer, training_mode, None, vs,
X=X, make_test_fn=True, initial_embeddings=embeddings)
class RNNTestCase(unittest.TestCase):
"""Basic functional tests for RNN with dummy data."""
def _make_rnn(self, seq_length=4):
self.embedding_dim = embedding_dim = 3
self.vocab_size = vocab_size = 10
self.seq_length = seq_length
def compose_network(h_prev, inp, embedding_dim, model_dim, vs, name="compose"):
# Just add the two embeddings!
W = T.concatenate([T.eye(model_dim), T.eye(model_dim)], axis=0)
i = T.concatenate([h_prev, inp], axis=1)
return i.dot(W)
X = T.imatrix("X")
training_mode = T.scalar("training_mode")
vs = VariableStore()
embeddings = np.arange(vocab_size).reshape(
(vocab_size, 1)).repeat(embedding_dim, axis=1)
self.model = RNN(
embedding_dim, embedding_dim, vocab_size, seq_length, compose_network,
IdentityLayer, training_mode, None, vs,
X=X, make_test_fn=True, initial_embeddings=embeddings)
def test_basic_ff(self):
self._make_rnn(4)
X = np.array([
[3, 1, 2, 0],
[3, 2, 4, 5]
], dtype=np.int32)
expected = np.array([[6, 6, 6],
[14, 14, 14]])
ret = self.model.scan_fn(X, 1.0)
np.testing.assert_almost_equal(ret, expected)