def get_model(
n_vocab,
n_ctx=1024,
n_embd=768,
n_head=12,
n_layer=12,
fixed_input_shape=False # neededforTPU training
):
"""Get basic GPT-2 model.
:param n_vocab: Number of vocabulary tokens.
:param n_ctx: The length of each input.
:param n_embd: The dimension of embeddings.
:param n_head: Number of heads in transformer.
:param n_layer: Number of transformer blocks.
:return: The model.
"""
if fixed_input_shape:
input_layer_shape = (n_ctx, )
else:
input_layer_shape = (None, )
input_layer = keras.layers.Input(shape=input_layer_shape, name='Input')
embed_token, embeddings = EmbeddingRet(
input_dim=n_vocab,
output_dim=n_embd,
mask_zero=False,
name='Embed-Token',
)(input_layer)
embed_token_pos = PositionEmbedding(
input_dim=n_ctx,
output_dim=n_embd,
mode=PositionEmbedding.MODE_ADD,
name='Embed-Token-Pos',
)(embed_token)
last_layer = embed_token_pos
for i in range(n_layer):
last_layer = _get_encoder_component(
name='Encode-%d' % i,
input_layer=last_layer,
head_num=n_head,
hidden_dim=n_embd * 4,
attention_activation=None,
feed_forward_activation=gelu,
)
norm_layer = LayerNormalization(name='Norm', )(last_layer)
output_layer = EmbeddingSim(
use_bias=False,
name='Output',
)([norm_layer, embeddings])
model = keras.models.Model(inputs=input_layer, outputs=output_layer)
model.compile(
optimizer=keras.optimizers.Adam(),
loss=keras.losses.sparse_categorical_crossentropy,
)
return model
def test_no_mask(self):
input_layer = keras.layers.Input(shape=(None,), name='Input')
embed, embed_weights = EmbeddingRet(
input_dim=20,
output_dim=100,
name='Embedding',
)(input_layer)
output_layer = EmbeddingSim(
name='Embed-Sim',
)([embed, embed_weights])
model = keras.models.Model(inputs=input_layer, outputs=output_layer)
model.compile(optimizer='adam', loss='mse')
model_path = os.path.join(tempfile.gettempdir(), 'test_embed_sim_%f.h5' % np.random.random())
model.save(model_path)
model = keras.models.load_model(model_path, custom_objects=get_custom_objects())
model.summary(line_length=100)
batch_inputs = np.random.randint(low=0, high=19, size=(32, 100))
batch_outputs = model.predict(batch_inputs)
batch_outputs = np.argmax(batch_outputs, axis=-1)
self.assertEqual(batch_inputs.tolist(), batch_outputs.tolist())
def get_model(token_num,
embed_dim,
encoder_num,
decoder_num,
head_num,
hidden_dim,
attention_activation=None,
feed_forward_activation='relu',
dropout_rate=0.0,
use_same_embed=True,
embed_weights=None,
embed_trainable=None,
trainable=True):
"""Get full model without compilation.
:param token_num: Number of distinct tokens.
:param embed_dim: Dimension of token embedding.
:param encoder_num: Number of encoder components.
:param decoder_num: Number of decoder components.
:param head_num: Number of heads in multi-head self-attention.
:param hidden_dim: Hidden dimension of feed forward layer.
:param attention_activation: Activation for multi-head self-attention.
:param feed_forward_activation: Activation for feed-forward layer.
:param dropout_rate: Dropout rate.
:param use_same_embed: Whether to use the same token embedding layer. `token_num`, `embed_weights` and
`embed_trainable` should be lists of two elements if it is False.
:param embed_weights: Initial weights of token embedding.
:param embed_trainable: Whether the token embedding is trainable. It will automatically set to False if the given
value is None when embedding weights has been provided.
:param trainable: Whether the keras_layers are trainable.
:return: Keras model.
"""
if not isinstance(token_num, list):
token_num = [token_num, token_num]
encoder_token_num, decoder_token_num = token_num
if not isinstance(embed_weights, list):
embed_weights = [embed_weights, embed_weights]
encoder_embed_weights, decoder_embed_weights = embed_weights
if encoder_embed_weights is not None:
encoder_embed_weights = [encoder_embed_weights]
if decoder_embed_weights is not None:
decoder_embed_weights = [decoder_embed_weights]
if not isinstance(embed_trainable, list):
embed_trainable = [embed_trainable, embed_trainable]
encoder_embed_trainable, decoder_embed_trainable = embed_trainable
if encoder_embed_trainable is None:
encoder_embed_trainable = encoder_embed_weights is None
if decoder_embed_trainable is None:
decoder_embed_trainable = decoder_embed_weights is None
if use_same_embed:
encoder_embed_layer = decoder_embed_layer = EmbeddingRet(
input_dim=encoder_token_num,
output_dim=embed_dim,
mask_zero=True,
weights=encoder_embed_weights,
trainable=encoder_embed_trainable,
name='Token-Embedding',
)
else:
encoder_embed_layer = EmbeddingRet(
input_dim=encoder_token_num,
output_dim=embed_dim,
mask_zero=True,
weights=encoder_embed_weights,
trainable=encoder_embed_trainable,
name='Encoder-Token-Embedding',
)
decoder_embed_layer = EmbeddingRet(
input_dim=decoder_token_num,
output_dim=embed_dim,
mask_zero=True,
weights=decoder_embed_weights,
trainable=decoder_embed_trainable,
name='Decoder-Token-Embedding',
)
encoder_input = keras.layers.Input(shape=(None, ), name='Encoder-Input')
encoder_embed = TrigPosEmbedding(
mode=TrigPosEmbedding.MODE_ADD,
name='Encoder-Embedding',
)(encoder_embed_layer(encoder_input)[0])
encoded_layer = get_encoders(
encoder_num=encoder_num,
input_layer=encoder_embed,
head_num=head_num,
hidden_dim=hidden_dim,
attention_activation=attention_activation,
feed_forward_activation=feed_forward_activation,
dropout_rate=dropout_rate,
trainable=trainable,
)
decoder_input = keras.layers.Input(shape=(None, ), name='Decoder-Input')
decoder_embed, decoder_embed_weights = decoder_embed_layer(decoder_input)
decoder_embed = TrigPosEmbedding(
mode=TrigPosEmbedding.MODE_ADD,
name='Decoder-Embedding',
)(decoder_embed)
decoded_layer = get_decoders(
decoder_num=decoder_num,
input_layer=decoder_embed,
encoded_layer=encoded_layer,
head_num=head_num,
hidden_dim=hidden_dim,
attention_activation=attention_activation,
feed_forward_activation=feed_forward_activation,
dropout_rate=dropout_rate,
trainable=trainable,
)
dense_layer = EmbeddingSim(
trainable=trainable,
name='Output',
)([decoded_layer, decoder_embed_weights])
return keras.models.Model(inputs=[encoder_input, decoder_input],
outputs=dense_layer)
def build_xlnet(units,
training,
num_token,
num_block,
num_head,
hidden_dim,
batch_size,
memory_len,
target_len,
permute=None,
mask_index=Tokenizer.SYM_PAD,
dropout=0.0,
attention_dropout=0.0,
attention_type=ATTENTION_TYPE_BI,
clamp_len=None,
shared_biases=True):
"""Build XLNet.
:param units: Hidden dimensions throughout the model.
:param training: Whether in training mode.
:param num_token: Number of distinct tokens.
:param num_block: Number of basic encoder blocks.
:param num_head: Number of heads for attention.
:param hidden_dim: Dimension inside position-wise feed-forward layer.
:param batch_size: Maximum batch size.
:param memory_len: The maximum length of memories.
:param target_len: The length of prediction block.
:param permute: Whether to enable permutation.
:param mask_index: The index of padding.
:param dropout: General dropout rate.
:param attention_dropout: Dropout rate inside attention layer.
:param attention_type: 'uni' or 'bi'.
:param clamp_len: The maximum value of relative position.
:param shared_biases: Whether to use the same biases for all layers.
:return: The built model.
"""
if permute is None:
permute = training
token_input = keras.layers.Input(
shape=(target_len,),
name='Input-Token',
)
seg_input = keras.layers.Input(
shape=(target_len,),
name='Input-Segment',
)
memory_length_input = keras.layers.Input(
shape=(1,),
name='Input-Memory-Length',
)
inputs = [token_input, seg_input, memory_length_input]
if training:
query_input = keras.layers.Input(
shape=(target_len,),
name='Input-Mask',
)
inputs.append(query_input)
else:
query_input = None
token_embed, embed_weights = EmbeddingRet(
input_dim=num_token,
output_dim=units,
mask_zero=mask_index == 0,
name='Embed-Token',
)(token_input)
if mask_index is not None and mask_index != 0:
masking = CreateMask(
mask_value=mask_index,
name='Masking',
)(token_input)
token_embed = RestoreMask(name='Embed-Token-Masked')([token_embed, masking])
if training:
mask_embed = MaskEmbedding(
units=units,
name='Embed-Mask'
)([token_embed, query_input])
else:
mask_embed = None
if 0.0 < dropout < 1.0:
token_embed = keras.layers.Dropout(
rate=dropout,
name='Embed-Token-Dropout'
)(token_embed)
if training:
mask_embed = keras.layers.Dropout(
rate=dropout,
name='Embed-Mask-Dropout'
)(mask_embed)
memories = [Memory(
batch_size=batch_size,
memory_len=memory_len,
target_len=target_len,
output_dim=units,
name='Memory-0',
)([token_embed, memory_length_input])]
pos_embed = PositionalEmbedding(
output_dim=units,
clamp_len=clamp_len,
directional=attention_type == 'uni',
name='Embed-Pos',
)([token_embed, memories[0]])
content_mask, query_mask = PermutationMask(
enabled=permute,
directional=attention_type == 'uni',
name='Permutation',
)([token_embed, memories[0]])
context_bias, relative_bias, segment_bias = None, None, None
if shared_biases:
context_bias, relative_bias = RelativeBias(
units,
name='Relative-Bias',
)(memories[0])
segment_bias = SegmentBias(
units,
name='Segment-Bias',
)(memories[0])
content_output, query_output = token_embed, None
if training:
query_output = mask_embed
for i in range(num_block):
if not shared_biases:
context_bias, relative_bias = RelativeBias(
units,
name='Relative-Bias-{}'.format(i + 1),
)(memories[i])
segment_bias = SegmentBias(
units,
name='Segment-Bias-{}'.format(i + 1),
)(memories[i])
segment_mat, segment_embed = RelativeSegmentEmbedding(
units=units,
name='Embed-Segment-{}'.format(i + 1),
)([seg_input, memories[i]])
attention = Attention(
units=units,
num_head=num_head,
use_bias=False,
attention_dropout=attention_dropout,
name='Attention-{}'.format(i + 1),
)
if 0.0 < dropout < 1.0:
attention_dropout_layer = keras.layers.Dropout(
rate=dropout,
name='Attention-Dropout-{}'.format(i + 1),
)
else:
attention_dropout_layer = None
attention_add = keras.layers.Add(name='Attention-Residual-{}'.format(i + 1))
attention_layer_norm = LayerNormalization(name='Attention-Normal-{}'.format(i + 1))
feed_forward = FeedForward(
units=hidden_dim,
dropout_rate=dropout,
activation=gelu,
name='FeedForward-{}'.format(i + 1),
)
if 0.0 < dropout < 1.0:
feed_forward_dropout = keras.layers.Dropout(
rate=dropout,
name='FeedForward-Dropout-{}'.format(i + 1),
)
else:
feed_forward_dropout = None
feed_forward_add = keras.layers.Add(name='FeedForward-Residual-{}'.format(i + 1))
feed_forward_layer_norm = LayerNormalization(name='FeedForward-Normal-{}'.format(i + 1))
content = content_output
def _build_output(query, mask):
attention_input = query
_output = attention([
query, content, memories[i],
segment_mat, segment_embed, pos_embed,
context_bias, relative_bias, segment_bias,
mask,
])
if attention_dropout_layer is not None:
_output = attention_dropout_layer(_output)
_output = attention_add([attention_input, _output])
_output = attention_layer_norm(_output)
feed_forward_input = _output
_output = feed_forward(_output)
if feed_forward_dropout is not None:
_output = feed_forward_dropout(_output)
_output = feed_forward_add([feed_forward_input, _output])
_output = feed_forward_layer_norm(_output)
return _output
content_output = _build_output(content_output, content_mask)
if training:
query_output = _build_output(query_output, query_mask)
if i < num_block - 1:
memories.append(Memory(
batch_size=batch_size,
memory_len=memory_len,
target_len=target_len,
output_dim=units,
name='Memory-{}'.format(i + 1),
)([content_output, memory_length_input]))
if training:
output = EmbeddingSim(name='Softmax')([query_output, embed_weights])
else:
output = content_output
model = keras.models.Model(
inputs=inputs,
outputs=output
)
return model
def get_model(token_num,
embed_dim,
encoder_num,
decoder_num,
head_num,
hidden_dim,
num_classes,
add_new_node,
attention_activation=None,
feed_forward_activation='relu',
dropout_rate=0.0,
use_same_embed=True,
embed_weights=None,
embed_trainable=None,
trainable=True,
use_adapter=False,
adapter_units=None,
adapter_activation='relu'):
if not isinstance(token_num, list):
token_num = [token_num, token_num]
encoder_token_num, decoder_token_num = token_num
if not isinstance(embed_weights, list):
embed_weights = [embed_weights, embed_weights]
encoder_embed_weights, decoder_embed_weights = embed_weights
if encoder_embed_weights is not None:
encoder_embed_weights = [encoder_embed_weights]
if decoder_embed_weights is not None:
decoder_embed_weights = [decoder_embed_weights]
if not isinstance(embed_trainable, list):
embed_trainable = [embed_trainable, embed_trainable]
encoder_embed_trainable, decoder_embed_trainable = embed_trainable
if encoder_embed_trainable is None:
encoder_embed_trainable = encoder_embed_weights is None
if decoder_embed_trainable is None:
decoder_embed_trainable = decoder_embed_weights is None
if use_same_embed:
encoder_embed_layer = decoder_embed_layer = EmbeddingRet(
input_dim=encoder_token_num,
output_dim=embed_dim,
mask_zero=True,
weights=encoder_embed_weights,
trainable=encoder_embed_trainable,
name='Token-Embedding',
)
else:
encoder_embed_layer = EmbeddingRet(
input_dim=encoder_token_num,
output_dim=embed_dim,
mask_zero=True,
weights=encoder_embed_weights,
trainable=encoder_embed_trainable,
name='Encoder-Token-Embedding',
)
decoder_embed_layer = EmbeddingRet(
input_dim=decoder_token_num,
output_dim=embed_dim,
mask_zero=True,
weights=decoder_embed_weights,
trainable=decoder_embed_trainable,
name='Decoder-Token-Embedding',
)
encoder_input = keras.layers.Input(shape=(None,), name='Encoder-Input')
encoder_embed = TrigPosEmbedding(
mode=TrigPosEmbedding.MODE_ADD,
name='Encoder-Embedding',
)(encoder_embed_layer(encoder_input)[0])
encoded_layer = get_encoders(
encoder_num=encoder_num,
input_layer=encoder_embed,
head_num=head_num,
hidden_dim=hidden_dim,
attention_activation=attention_activation,
feed_forward_activation=feed_forward_activation,
dropout_rate=dropout_rate,
trainable=trainable,
use_adapter=use_adapter,
adapter_units=adapter_units,
adapter_activation=adapter_activation,
)
decoder_input = keras.layers.Input(shape=(None,), name='Decoder-Input')
decoder_embed, decoder_embed_weights = decoder_embed_layer(decoder_input)
decoder_embed = TrigPosEmbedding(
mode=TrigPosEmbedding.MODE_ADD,
name='Decoder-Embedding',
)(decoder_embed)
decoded_layer = get_decoders(
decoder_num=decoder_num,
input_layer=decoder_embed,
encoded_layer=encoded_layer,
head_num=head_num,
hidden_dim=hidden_dim,
attention_activation=attention_activation,
feed_forward_activation=feed_forward_activation,
dropout_rate=dropout_rate,
trainable=trainable,
use_adapter=use_adapter,
adapter_units=adapter_units,
adapter_activation=adapter_activation,
)
dense_layer = EmbeddingSim(
trainable=trainable,
name='normal_end',
)([decoded_layer, decoder_embed_weights])
if add_new_node == False:
dense = Dense(units=num_classes, activation="softmax")(decoded_layer)
elif add_new_node == True:
print("add new node")
dense = Dense(units=num_classes+1, activation="softmax")(decoded_layer)
#return keras.models.Model(inputs=[encoder_input], outputs=dense)
return keras.models.Model(inputs=[encoder_input, decoder_input], outputs=dense)
def get_model(n_vocab,
n_ctx=1024,
n_embd=768,
n_head=12,
n_layer=12,
batch_size=None,
fixed_input_shape=False):
"""Get basic GPT-2 model.
:param n_vocab: Number of vocabulary tokens.
:param n_ctx: The length of each input.
:param n_embd: The dimension of embeddings.
:param n_head: Number of heads in transformer.
:param n_layer: Number of transformer blocks.
:param batch_size: Batch size of the model.
:param fixed_input_shape: Whether the length of input is fixed. (Needed for TPU training)
:return: The model.
"""
if fixed_input_shape:
input_layer_shape = (batch_size, n_ctx)
else:
input_layer_shape = (batch_size, None)
lm_input_layer = tf.keras.layers.Input(
batch_shape=input_layer_shape,
name='LMInput',
)
mc_input_layer = tf.keras.layers.Input(
batch_shape=(batch_size, ),
name='MCInput',
)
embed_token, embeddings = EmbeddingRet(
input_dim=n_vocab,
output_dim=n_embd,
mask_zero=False,
name='Embed-Token',
)(lm_input_layer)
embed_token_pos = PositionEmbedding(
input_dim=n_ctx,
output_dim=n_embd,
mode=PositionEmbedding.MODE_ADD,
name='Embed-Token-Pos',
)(embed_token)
last_layer = embed_token_pos
for i in range(n_layer):
last_layer = _get_encoder_component(
name='Encode-%d' % i,
input_layer=last_layer,
head_num=n_head,
hidden_dim=n_embd * 4,
attention_activation=None,
feed_forward_activation=gelu,
)
norm_layer = LayerNormalization(name='Norm', )(last_layer)
lm_head = EmbeddingSim(
use_bias=False,
name='LMOutput',
)([norm_layer, embeddings])
mc_sequence_summary = SequenceSummary(name='MCSequenceSummary')(
[norm_layer, mc_input_layer])
mc_linear = Dense(units=1, input_shape=(n_embd, ),
name='MCDense')(mc_sequence_summary)
mc_head = Dropout(rate=0.1, name='MCOutput')(mc_linear)
losses = {
"LMOutput": lm_loss_function,
"MCOutput": mc_loss_function,
}
lossWeights = {"LMOutput": 2.0, "MCOutput": 1.0}
metrics = {"LMOutput": get_metrics(), 'MCOutput': get_metrics(is_mc=True)}
model = tf.keras.models.Model(inputs=[lm_input_layer, mc_input_layer],
outputs=[lm_head, mc_head])
model.compile(
optimizer=tf.keras.optimizers.Adam(clipnorm=1.),
loss=losses,
loss_weights=lossWeights,
#metrics=metrics
)
return model
