def test_get_config(self):
num_attention_heads = 2
decoder_block = transformer.TransformerDecoderBlock(
num_attention_heads=num_attention_heads,
intermediate_size=32,
intermediate_activation='relu',
dropout_rate=0.1,
attention_dropout_rate=0.1,
use_bias=False,
norm_first=True,
norm_epsilon=1e-6,
intermediate_dropout=0.1,
attention_initializer=tf.keras.initializers.RandomUniform(
minval=0., maxval=1.))
decoder_block_config = decoder_block.get_config()
new_decoder_block = transformer.TransformerDecoderBlock.from_config(
decoder_block_config)
self.assertEqual(decoder_block_config, new_decoder_block.get_config())
def test_decoder_block_with_cache(self):
num_attention_heads = 2
hidden_size = 16
decoder_block = transformer.TransformerDecoderBlock(
num_attention_heads=num_attention_heads,
intermediate_size=32,
intermediate_activation='relu',
dropout_rate=0.1,
attention_dropout_rate=0.1)
# Forward path.
dummy_tensor = tf.zeros([2, 4, 16], dtype=tf.float32)
dummy_mask = tf.zeros([2, 4, 4], dtype=tf.float32)
inputs = [dummy_tensor, dummy_tensor, dummy_mask, dummy_mask]
cache = _create_cache(2, 0, num_attention_heads,
hidden_size // num_attention_heads)
output, cache = decoder_block(inputs, cache)
self.assertEqual(output.shape, (2, 4, hidden_size))
self.assertEqual(cache['value'].shape, (2, 4, 2, 8))
def test_use_bias_norm_first(self):
num_attention_heads = 2
hidden_size = 16
decoder_block = transformer.TransformerDecoderBlock(
num_attention_heads=num_attention_heads,
intermediate_size=32,
intermediate_activation='relu',
dropout_rate=0.1,
attention_dropout_rate=0.1,
use_bias=False,
norm_first=True,
norm_epsilon=1e-6,
intermediate_dropout=0.1,
attention_initializer=tf.keras.initializers.RandomUniform(
minval=0., maxval=1.))
# Forward path.
dummy_tensor = tf.zeros([2, 4, 16], dtype=tf.float32)
dummy_mask = tf.zeros([2, 4, 4], dtype=tf.float32)
inputs = [dummy_tensor, dummy_tensor, dummy_mask, dummy_mask]
output, _ = decoder_block(inputs)
self.assertEqual(output.shape, (2, 4, hidden_size))