def test_cutoffs_no_projection_bind(self):
input_layer = keras.layers.Input(shape=(None, ))
embed_layer = AdaptiveEmbedding(
input_dim=30,
output_dim=8,
cutoffs=[10, 20, 25],
div_val=2,
mask_zero=True,
force_projection=False,
return_embeddings=True,
return_projections=True,
)(input_layer)
softmax_layer = AdaptiveSoftmax(
input_dim=8,
output_dim=30,
cutoffs=[10, 20, 25],
div_val=2,
force_projection=False,
bind_embeddings=True,
bind_projections=True,
)(embed_layer)
model = keras.models.Model(input_layer, softmax_layer)
model_path = os.path.join(
tempfile.gettempdir(),
'test_ada_softmax_%f.h5' % np.random.random())
model.save(model_path)
model = keras.models.load_model(model_path,
custom_objects={
'AdaptiveEmbedding':
AdaptiveEmbedding,
'AdaptiveSoftmax': AdaptiveSoftmax,
})
model.summary()
def test_force_projection_no_binding(self):
input_layer = keras.layers.Input(shape=(None, ))
embed_layer = AdaptiveEmbedding(
input_dim=3,
output_dim=16,
force_projection=True,
return_embeddings=True,
return_projections=True,
)(input_layer)
softmax_layer = AdaptiveSoftmax(
input_dim=16,
output_dim=3,
force_projection=True,
)(embed_layer)
model = keras.models.Model(input_layer, softmax_layer)
model_path = os.path.join(
tempfile.gettempdir(),
'test_ada_softmax_%f.h5' % np.random.random())
model.save(model_path)
model = keras.models.load_model(model_path,
custom_objects={
'AdaptiveEmbedding':
AdaptiveEmbedding,
'AdaptiveSoftmax': AdaptiveSoftmax,
})
model.summary()
def test_fit(self):
input_layer = keras.layers.Input(shape=(None, ))
embed_layer = AdaptiveEmbedding(
input_dim=30,
output_dim=32,
cutoffs=[5, 15, 25],
div_val=2,
return_embeddings=True,
return_projections=True,
mask_zero=True,
)(input_layer)
dense_layer = keras.layers.Dense(
units=32,
activation='tanh',
)(embed_layer[0])
softmax_layer = AdaptiveSoftmax(
input_dim=32,
output_dim=30,
cutoffs=[5, 15, 25],
div_val=2,
bind_embeddings=True,
bind_projections=True,
)([dense_layer] + embed_layer[1:])
model = keras.models.Model(inputs=input_layer, outputs=softmax_layer)
model.compile('adam', 'sparse_categorical_crossentropy')
model.summary()
inputs = np.random.randint(0, 30, (4096, 10))
outputs = np.expand_dims(inputs, axis=-1)
model.fit(
inputs,
outputs,
epochs=100,
callbacks=[
keras.callbacks.EarlyStopping(monitor='loss',
min_delta=1e-4,
patience=2),
],
)
model = keras.models.Model(input_layer, softmax_layer)
model_path = os.path.join(
tempfile.gettempdir(),
'test_ada_softmax_%f.h5' % np.random.random())
model.save(model_path)
model = keras.models.load_model(model_path,
custom_objects={
'AdaptiveEmbedding':
AdaptiveEmbedding,
'AdaptiveSoftmax': AdaptiveSoftmax,
})
inputs = np.random.randint(0, 30, (128, 10))
outputs = model.predict(inputs).argmax(axis=-1)
outputs *= np.not_equal(inputs, 0).astype('int32')
diff = np.sum(np.not_equal(inputs, outputs))
self.assertLess(diff, 5)
def test_sample_default(self):
input_layer = keras.layers.Input(shape=(None, ))
embed_layer = AdaptiveEmbedding(
input_dim=3,
output_dim=16,
return_embeddings=True,
return_projections=True,
)(input_layer)
func = K.function([input_layer], embed_layer)
outputs = func([np.array([[0, 1, 2]])])
self.assertTrue(np.allclose(outputs[0], outputs[1]))
def test_single_projection(self):
input_layer = keras.layers.Input(shape=(None, ))
embed_layer = AdaptiveEmbedding(
input_dim=3,
output_dim=16,
embed_dim=5,
return_embeddings=True,
return_projections=True,
)(input_layer)
model = keras.models.Model(input_layer, embed_layer)
model_path = os.path.join(tempfile.gettempdir(),
'test_ada_embed_%f.h5' % np.random.random())
model.save(model_path)
model = keras.models.load_model(
model_path,
custom_objects={'AdaptiveEmbedding': AdaptiveEmbedding})
model.summary()
def build_albert(token_num,
pos_num=512,
seq_len=512,
embed_dim=128,
hidden_dim=768,
transformer_num=12,
head_num=12,
feed_forward_dim=3072,
dropout_rate=0.1,
attention_activation=None,
feed_forward_activation='gelu',
training=True,
trainable=None,
output_layers=None):
"""Get ALBERT model.
See: https://arxiv.org/pdf/1909.11942.pdf
:param token_num: Number of tokens.
:param pos_num: Maximum position.
:param seq_len: Maximum length of the input sequence or None.
:param embed_dim: Dimensions of embeddings.
:param hidden_dim: Dimensions of hidden layers.
:param transformer_num: Number of transformers.
:param head_num: Number of heads in multi-head attention
in each transformer.
:param feed_forward_dim: Dimension of the feed forward layer
in each transformer.
:param dropout_rate: Dropout rate.
:param attention_activation: Activation for attention layers.
:param feed_forward_activation: Activation for feed-forward layers.
:param training: A built model with MLM and NSP outputs will be returned
if it is `True`, otherwise the input layers and the last
feature extraction layer will be returned.
:param trainable: Whether the model is trainable.
:param output_layers: A list of indices of output layers.
"""
if attention_activation == 'gelu':
attention_activation = gelu
if feed_forward_activation == 'gelu':
feed_forward_activation = gelu
if trainable is None:
trainable = training
def _trainable(_layer):
if isinstance(trainable, (list, tuple, set)):
for prefix in trainable:
if _layer.name.startswith(prefix):
return True
return False
return trainable
# Build inputs
input_token = keras.layers.Input(shape=(seq_len, ), name='Input-Token')
input_segment = keras.layers.Input(shape=(seq_len, ), name='Input-Segment')
inputs = [input_token, input_segment]
# Build embeddings
embed_token, embed_weights, embed_projection = AdaptiveEmbedding(
input_dim=token_num,
output_dim=hidden_dim,
embed_dim=embed_dim,
mask_zero=True,
trainable=trainable,
return_embeddings=True,
return_projections=True,
name='Embed-Token',
)(input_token)
embed_segment = keras.layers.Embedding(
input_dim=2,
output_dim=hidden_dim,
trainable=trainable,
name='Embed-Segment',
)(input_segment)
embed_layer = keras.layers.Add(name='Embed-Token-Segment')(
[embed_token, embed_segment])
embed_layer = PositionEmbedding(
input_dim=pos_num,
output_dim=hidden_dim,
mode=PositionEmbedding.MODE_ADD,
trainable=trainable,
name='Embedding-Position',
)(embed_layer)
if dropout_rate > 0.0:
dropout_layer = keras.layers.Dropout(
rate=dropout_rate,
name='Embedding-Dropout',
)(embed_layer)
else:
dropout_layer = embed_layer
embed_layer = LayerNormalization(
trainable=trainable,
name='Embedding-Norm',
)(dropout_layer)
# Build shared transformer
attention_layer = MultiHeadAttention(
head_num=head_num,
activation=attention_activation,
name='Attention',
)
attention_normal = LayerNormalization(name='Attention-Normal')
feed_forward_layer = FeedForward(units=feed_forward_dim,
activation=feed_forward_activation,
name='Feed-Forward')
feed_forward_normal = LayerNormalization(name='Feed-Forward-Normal')
transformed = embed_layer
transformed_layers = []
for i in range(transformer_num):
attention_input = transformed
transformed = attention_layer(transformed)
if dropout_rate > 0.0:
transformed = keras.layers.Dropout(
rate=dropout_rate,
name='Attention-Dropout-{}'.format(i + 1),
)(transformed)
transformed = keras.layers.Add(
name='Attention-Add-{}'.format(i + 1), )(
[attention_input, transformed])
transformed = attention_normal(transformed)
feed_forward_input = transformed
transformed = feed_forward_layer(transformed)
if dropout_rate > 0.0:
transformed = keras.layers.Dropout(
rate=dropout_rate,
name='Feed-Forward-Dropout-{}'.format(i + 1),
)(transformed)
transformed = keras.layers.Add(
name='Feed-Forward-Add-{}'.format(i + 1), )(
[feed_forward_input, transformed])
transformed = feed_forward_normal(transformed)
transformed_layers.append(transformed)
if training:
# Build tasks
mlm_dense_layer = keras.layers.Dense(
units=hidden_dim,
activation=feed_forward_activation,
name='MLM-Dense',
)(transformed)
mlm_norm_layer = LayerNormalization(name='MLM-Norm')(mlm_dense_layer)
mlm_pred_layer = AdaptiveSoftmax(
input_dim=hidden_dim,
output_dim=token_num,
embed_dim=embed_dim,
bind_embeddings=True,
bind_projections=True,
name='MLM-Sim',
)([mlm_norm_layer, embed_weights, embed_projection])
masked_layer = Masked(name='MLM')([mlm_pred_layer, inputs[-1]])
extract_layer = Extract(index=0, name='Extract')(transformed)
nsp_dense_layer = keras.layers.Dense(
units=hidden_dim,
activation='tanh',
name='SOP-Dense',
)(extract_layer)
nsp_pred_layer = keras.layers.Dense(
units=2,
activation='softmax',
name='SOP',
)(nsp_dense_layer)
model = keras.models.Model(inputs=inputs,
outputs=[masked_layer, nsp_pred_layer])
for layer in model.layers:
layer.trainable = _trainable(layer)
return model
if output_layers is not None:
if isinstance(output_layers, list):
output_layers = [
transformed_layers[index] for index in output_layers
]
output = keras.layers.Concatenate(name='Output', )(output_layers)
else:
output = transformed_layers[output_layers]
model = keras.models.Model(inputs=inputs, outputs=output)
return model
model = keras.models.Model(inputs=inputs, outputs=transformed)
for layer in model.layers:
layer.trainable = _trainable(layer)
return inputs, transformed
def build_transformer_xl(units,
embed_dim,
hidden_dim,
num_token,
num_block,
num_head,
batch_size,
memory_len,
target_len,
dropout=0.0,
attention_dropout=0.0,
cutoffs=None,
div_val=1,
force_projection=None,
bind_embeddings=True,
bind_projections=True,
clamp_len=None,
share_biases=True):
"""Build transformer-XL model.
:param units: Units inside the transformer.
:param embed_dim: Dimension of embeddings.
:param hidden_dim: Dimension inside position-wise feed-forward layer.
:param num_token: Number of distinct input tokens.
:param num_block: Number of basic encoder blocks.
:param num_head: Number of heads for attention.
:param batch_size: Maximum batch size.
:param memory_len: The maximum length of memories.
:param target_len: The length of prediction block.
:param dropout: General dropout rate.
:param attention_dropout: Dropout rate inside attention layer.
:param cutoffs: Cutoffs of adaptive embedding.
:param div_val: Scale factor of adaptive embedding.
:param force_projection: Add projection when the dimensions are equal.
:param bind_embeddings: Whether to bind embeddings to adaptive softmax.
:param bind_projections: Whether to bind projections to adaptive softmax.
:param clamp_len: The maximum value of relative position.
:param share_biases: Whether to use the same biases for all layers.
:return: The built model.
"""
token_input = keras.layers.Input(shape=(target_len,), name='Input-Token')
memory_length_input = keras.layers.Input(shape=(1,), name='Input-Memory-Length')
inputs = [token_input, memory_length_input]
results = AdaptiveEmbedding(
input_dim=num_token,
output_dim=units,
embed_dim=embed_dim,
cutoffs=cutoffs,
div_val=div_val,
mask_zero=True,
force_projection=force_projection,
return_embeddings=True,
return_projections=True,
name='Embed-Token',
)(token_input)
token_embed, embedding_weights = results[0], results[1:]
token_embed = Scale(scale=np.sqrt(units), name='Embed-Token-Scaled')(token_embed)
last_memory = Memory(
batch_size=batch_size,
memory_len=memory_len,
target_len=target_len,
output_dim=units,
name='Memory-0',
)([token_embed, memory_length_input])
position_embed = PositionalEmbedding(
output_dim=units,
clamp_len=clamp_len,
name='Embed-Position',
)([token_input, last_memory])
if 0.0 < dropout < 1.0:
token_embed = keras.layers.Dropout(rate=dropout, name='Embed-Token-Dropped')(token_embed)
position_embed = keras.layers.Dropout(rate=dropout, name='Embed-Position-Dropped')(position_embed)
context_bias, relative_bias = None, None
if share_biases:
context_bias, relative_bias = RelativeBias(units=units, name='Biases')(last_memory)
outputs = [token_embed]
for i in range(num_block):
block_input, block_output = outputs[-1], outputs[-1]
if not share_biases:
context_bias, relative_bias = RelativeBias(units=units, name='Biases-{}'.format(i + 1))(last_memory)
block_output = RelativePartialMultiHeadSelfAttention(
units=units,
num_head=num_head,
use_bias=False,
attention_dropout=attention_dropout,
name='Attention-{}'.format(i + 1),
)([block_output, position_embed, last_memory, context_bias, relative_bias])
if 0.0 < dropout < 1.0:
block_output = keras.layers.Dropout(rate=dropout, name='Attention-Dropped-{}'.format(i + 1))(block_output)
block_output = keras.layers.Add(name='Attention-Res-{}'.format(i + 1))([block_input, block_output])
block_output = LayerNormalization(name='Attention-Norm-{}'.format(i + 1))(block_output)
block_input = block_output
block_output = FeedForward(
units=hidden_dim,
dropout_rate=dropout,
name='FeedForward-{}'.format(i + 1),
)(block_output)
if 0.0 < dropout < 1.0:
block_output = keras.layers.Dropout(rate=dropout, name='FeedForward-Dropped-{}'.format(i + 1))(block_output)
block_output = keras.layers.Add(name='FeedForward-Res-{}'.format(i + 1))([block_input, block_output])
block_output = LayerNormalization(name='FeedForward-Norm-{}'.format(i + 1))(block_output)
if i < num_block - 1:
last_memory = Memory(
batch_size=batch_size,
memory_len=memory_len,
target_len=target_len,
output_dim=units,
name='Memory-{}'.format(i + 1),
)([block_output, memory_length_input])
outputs.append(block_output)
if 0.0 < dropout < 1.0:
outputs[-1] = keras.layers.Dropout(rate=dropout, name='Output-Dropped')(outputs[-1])
softmax = AdaptiveSoftmax(
input_dim=units,
output_dim=num_token,
embed_dim=embed_dim,
cutoffs=cutoffs,
div_val=div_val,
force_projection=force_projection,
bind_embeddings=bind_embeddings,
bind_projections=bind_projections,
name='Softmax',
)(outputs[-1:] + embedding_weights)
model = keras.models.Model(inputs=inputs, outputs=softmax)
return model
def test_sample_cutoffs(self):
embed_0 = np.array([
[
0.7562694862279867,
-0.7532437781410828,
-0.2882295795429552,
-1.6990371818805843,
-0.09864164817566004,
-0.5235034477186453,
-1.600153091413999,
0.03441732751250957,
],
[
-0.3680529905261407,
1.1673600332887637,
-0.6914459306809843,
-0.7645030146906124,
2.0434827620248606,
-0.2743642839675437,
0.04834288951969495,
-1.0368596183756285,
],
[
-0.8440324158987662,
0.05585795322288273,
-0.5827731797867599,
1.502853709909658,
-0.09311037618863122,
1.366316512453695,
-0.3834091917878978,
-1.2647642860801802,
],
[
1.5212768184170435,
-0.7854311748221854,
-0.4674213048014483,
-1.0460200278367862,
0.3705555995848165,
-0.12273261562651422,
1.8138708310050653,
-0.26957084415202764,
],
[
-0.15162771245260723,
-0.19654664890167275,
-1.77930041719533,
-0.6987101769248606,
0.32681036318004547,
0.19156716698736181,
0.8386004334587568,
-1.8390076172747616,
],
[
-1.1363779747587972,
-0.15233131547247872,
0.158423477487577,
-0.6984487776859649,
1.2424950830966563,
-0.16130616338419873,
-1.6298737099566283,
1.7229575808498785,
],
[
0.613169803410901,
-1.5391239758406403,
-1.2476893436624792,
-0.05514513857644962,
-0.5537408608863357,
-0.9965187549427492,
-0.6842234254089083,
-1.2420165307152238,
],
[
-0.4086071455923046,
-0.7286151488450243,
1.2938629380821804,
0.7450912596769113,
-0.13042129128885002,
-1.4269400640112133,
-0.713571658756806,
-0.5036154349645774,
],
[
0.7326026846217363,
0.12752591749386982,
0.7387086112901271,
-1.4161019970745967,
-0.6396944907214142,
-2.0010110577965783,
0.5843029435066284,
-0.4033331631189724,
],
[
1.22301664512685,
-0.024541032664251092,
-0.27128167541306714,
1.910258142043872,
-0.9673069099782774,
0.6614265651081772,
-1.165650716838653,
-0.5085143504562967,
],
])
embed_1 = np.array([
[
0.6593494357199338, -0.06233478795012013, 0.3394579881849406,
0.05894554241531747
],
[
1.0015451559801243, 0.7487130375684998, -0.4244371286817957,
-0.45182923128222996
],
[
-0.41965070720383035, -0.2875756074838825, 1.8712603426351773,
2.531083895835167
],
[
-0.6800689195006436, -0.39454047242128376, 0.5442439581019961,
-0.21672610899025968
],
[
-1.3119449289237803, 1.5645034642903253, 1.3203132828621442,
1.7673879116655695
],
[
-0.8817194029613362, -0.6655645822150862, 0.2341787847442309,
-0.7641095447924122
],
[
-0.47497798682688624, 1.0109350638555383, -0.5514102704837403,
-0.1450007600387442
],
[
-0.531267085230172, 0.12862169808408846, 0.18339345878624577,
1.5279135983387981
],
[
0.43338928943049837, 0.2660771849859784, 1.4227633495535283,
-0.5072818940455809
],
[
0.8704222505796531, 0.9361117741463981, 0.7442665348863866,
0.91392694614948
],
])
embed_2 = np.array([
[1.2712292341556446, 1.009655780936284],
[0.4420362222435132, 1.5186087787070979],
[-0.10018465175352317, -0.09182475290216006],
[-1.246047485363712, 1.6404603895987184],
[1.4427767754835976, 1.2102150762070925],
])
embed_3 = np.array([
[0.8285545743394414],
[0.7111875779008273],
[0.35799413043562894],
[-0.15005629449852656],
[0.6263946579941496],
])
proj_0 = np.array([
[0.3409731658714878, 0.032745006392315756, 0.668797744010083],
[-0.3082491589087075, -1.0028023345331745, 0.2122102239605163],
[-0.3751562822576601, -0.5825445529201775, 0.43389258576225614],
[0.26067868083146517, 0.8192897299406429, 0.073726048897453],
[1.1346146882950412, -2.456072992985481, -0.054474463562940736],
[-1.0283521269636255, -0.1983876737118115, 1.0132159972212373],
[2.72334361610427, 0.5683724225575054, 2.403638230905517],
[-0.2137114185905606, 0.3048293347650425, 1.510425235737199],
])
proj_1 = np.array([
[0.42186259731067743, 0.6034344571434473, 2.362015513199549],
[-0.9313583984951119, -0.8242699945665621, 0.2596454482698166],
[0.8871149648450185, -0.663397984939589, -1.195129355668761],
[0.8016784490871957, 0.13830808473255815, -0.6580242457235711],
])
proj_2 = np.array([
[1.4802477891158519, 0.12638370704617574, -0.18503256737397666],
[-0.3900434531439191, 0.14771223879593204, -0.8863321455068343],
])
proj_3 = np.array(
[[-0.589729339138385, 2.018799784975004, -0.08431336326635828]])
input_layer = keras.layers.Input(shape=(None, ))
embed_layer = AdaptiveEmbedding(
input_dim=30,
output_dim=3,
embed_dim=8,
cutoffs=[10, 20, 25],
div_val=2,
mask_zero=True,
return_embeddings=True,
return_projections=True,
)
func = K.function([input_layer], embed_layer(input_layer))
embed_layer.set_weights([
embed_0,
proj_0,
embed_1,
proj_1,
embed_2,
proj_2,
embed_3,
proj_3,
])
outputs = func([np.array([list(range(30))])])
expected = np.array([
[-3.783413887023926, -0.9968423843383789, -4.223631381988525],
[2.528728485107422, -6.659335613250732, -2.194012403488159],
[-1.9791769981384277, 0.8412808179855347, -2.137157917022705],
[6.2075581550598145, 0.31576472520828247, 4.379002094268799],
[3.3448808193206787, -0.268412709236145, -1.552351474761963],
[-3.813311815261841, -3.9697980880737305, -2.3214385509490967],
[-0.06424117088317871, 3.0353987216949463, -4.962082862854004],
[-0.7221541404724121, 0.6183103322982788, -3.726100444793701],
[2.573601245880127, 0.48284363746643066, -0.4642190933227539],
[-3.8191750049591064, 3.2147698402404785, -2.0111422538757324],
[0.6846045255661011, 0.23221178352832794, 1.0967247486114502],
[-1.013551950454712, 0.20630428194999695, 3.3646368980407715],
[3.7799394130706787, -0.9075126051902771, -4.967802047729492],
[0.3896251916885376, -0.4761944115161896, -2.216604709625244],
[0.5775725841522217, -2.712695360183716, -5.433547496795654],
[-0.1569119393825531, -0.24449113011360168, -2.0325169563293457],
[-1.7473266124725342, -0.7741519212722778, -0.10500013828277588],
[1.04367196559906, -0.33694392442703247, -2.4460482597351074],
[0.7904950380325317, -0.971816897392273, -0.2738245725631714],
[0.8882685303688049, -0.6137074828147888, 0.8081271648406982],
[1.487924575805664, 0.3098011910915375, -1.130109190940857],
[0.06199967861175537, 0.2801832854747772, -1.4277828931808472],
[-0.1124824583530426, -0.026225347071886063, 0.09992465376853943],
[-2.484309673309326, 0.0848359763622284, -1.2234333753585815],
[1.6636306047439575, 0.3611070513725281, -1.3396131992340088],
[-0.4886229634284973, 1.6726857423782349, -0.06985822319984436],
[-0.4194082021713257, 1.435745358467102, -0.059962619096040726],
[-0.2111196517944336, 0.7227184772491455, -0.030183689668774605],
[0.08849260210990906, -0.30293360352516174, 0.012651749886572361],
[-0.36940330266952515, 1.264565348625183, -0.05281343683600426],
])
self.assertTrue(np.allclose(expected, outputs[0][0]))
