def test_transformer_decoder_block3():
""" Typical use case: encoder feed into decoder with multi layer """
a = C.sequence.input_variable(10)
b = C.sequence.input_variable(10)
encoder_block = TransformerEncoderBlock(num_heads=2, model_dim=10, intermediate_dim=30, dropout_rate=0.1)
decoder_block1 = TransformerDecoderBlock(num_heads=2, model_dim=10, intermediate_dim=30, dropout_rate=0.1,
obey_sequence_order=True, max_seq_len=100)
decoder_block2 = TransformerDecoderBlock(num_heads=2, model_dim=10, intermediate_dim=30, dropout_rate=0.1,
obey_sequence_order=True, max_seq_len=100)
encoded = encoder_block(a)
decoded = decoder_block1(encoded, b)
decoded = decoder_block2(encoded, decoded)
assert decoded.shape == (10, )
n = [np.random.random((2, 10)).astype(np.float32),
np.random.random((4, 10)).astype(np.float32),
np.random.random((6, 10)).astype(np.float32)]
m = [np.random.random((2, 10)).astype(np.float32),
np.random.random((4, 10)).astype(np.float32),
np.random.random((6, 10)).astype(np.float32)]
results = decoded.eval({a: n, b: m})
def test_transformer_decoder_block2():
""" default settings: encoder block output feed into decoder block """
a = C.sequence.input_variable(10)
b = C.sequence.input_variable(10)
encoder_block = TransformerEncoderBlock(num_heads=2,
model_dim=10,
intermediate_dim=30,
dropout_rate=0.1)
decoder_block = TransformerDecoderBlock(num_heads=2,
model_dim=10,
intermediate_dim=30,
dropout_rate=0.1,
obey_sequence_order=True,
max_seq_len=100)
encoded = encoder_block(a)
decoded = decoder_block(encoded, b)
assert decoded.shape == (10, )
n = [
np.random.random((2, 10)).astype(np.float32),
np.random.random((4, 10)).astype(np.float32),
np.random.random((6, 10)).astype(np.float32)
]
m = [
np.random.random((2, 10)).astype(np.float32),
np.random.random((4, 10)).astype(np.float32),
np.random.random((6, 10)).astype(np.float32)
]
results = decoded.eval({a: n, b: m})