def tl_disbert_model(param={}):
trainable = param['Trainable']
max_seq_len = param['Max_length']
inputs = Input(shape= (max_seq_len,), dtype ='int64', name='inputs')
masks = Input(shape = (max_seq_len,), dtype='int64', name='masks')
disBert_model = TFDistilBertModel.from_pretrained('distilbert-base-uncased')
disBert_model.trainable = param['Trainable']
disBert_output = disBert_model(inputs, attention_mask = masks)
disBert_last_hidden = disBert_output.last_hidden_state
disBert_CLS_output = disBert_last_hidden [:,0,:]
x = Flatten()(disBert_CLS_output)
x = LayerNormalization()(x)
x = Dense(param['first_layer'], activation='relu')(x)
x = Dropout(param['dropout'])(x)
x = LayerNormalization()(x)
x = Dense(param['second_layer'], activation='relu')(x)
x = Dropout(param['dropout'])(x)
probs = Dense(3, activation='softmax')(x)
model = keras.Model(inputs = [inputs, masks], outputs=probs)
model.summary()
return model
def _create_sentence_transformer(self, input_shape): input_ids = tf.keras.Input(shape=input_shape, name='input_ids', dtype=tf.int32) attention_mask = tf.keras.Input(shape=input_shape, name='attention_mask', dtype=tf.int32) transformer_model = TFDistilBertModel.from_pretrained(self.model_name, config = self.model_config) word_embedding_layer = transformer_model([input_ids, attention_mask])[0] sentence_embedding_layer = PoolingLayer(pooling_type="mean")([word_embedding_layer, attention_mask]) return tf.keras.Model([input_ids, attention_mask], sentence_embedding_layer)
def build_model():
"""
This model is build based upon the DistilBert model taken from the Huggingface's Transformer library
The model has to be compiled before weight loading.
"""
pretrained_model = TFDistilBertModel.from_pretrained(
'distilbert-base-uncased', output_attentions=False
)
input_ids = tf.keras.layers.Input(
shape=(None,), dtype=tf.int32, name='input_ids_pl'
)
attention_mask = tf.keras.layers.Input(
shape=(None,), dtype=tf.int32, name='attention_mask_pl'
)
# get the output of the '[CLS'] token on the last layer
bert_output = pretrained_model(
{'input_ids': input_ids, 'attention_mask': attention_mask},
return_dict=True
)['last_hidden_state'][:, 0]
pre_classification = tf.keras.layers.Dense(128, activation='tanh')(bert_output)
dropout_1 = tf.keras.layers.Dropout(0.3)(pre_classification)
classification_output = tf.keras.layers.Dense(2, activation='softmax')(dropout_1)
model = tf.keras.models.Model(
inputs=[input_ids, attention_mask], outputs=classification_output
)
return model
def extract_embeddings_for_other_clf():
distil_bert = "distilbert-base-uncased"
config = DistilBertConfig(dropout=0.2, attention_dropout=0.2)
config.output_hidden_states = False
transformer_model = TFDistilBertModel.from_pretrained(distil_bert,
config=config)
input_ids_in = tf.keras.layers.Input(shape=(25, ),
name="input_token",
dtype="int32")
input_masks_in = tf.keras.layers.Input(shape=(25, ),
name="masked_token",
dtype="int32")
embedding_layer = transformer_model(input_ids_in,
attention_mask=input_masks_in)[0]
cls_token = embedding_layer[:, 0, :]
X = tf.keras.layers.BatchNormalization()(cls_token)
X = tf.keras.layers.Dense(192, activation="relu")(X)
X = tf.keras.layers.Dropout(0.2)(X)
X = tf.keras.layers.Dense(3, activation="softmax")(X)
model = tf.keras.Model(inputs=[input_ids_in, input_masks_in], outputs=X)
for layer in model.layers[:3]:
layer.trainable = False
return model
def __init__(
self,
pretrained_model_name_or_path='distilbert-base-uncased',
reduce_output='sum',
trainable=True,
num_tokens=None,
**kwargs
):
super(DistilBERTEncoder, self).__init__()
try:
from transformers import TFDistilBertModel
except ModuleNotFoundError:
logger.error(
' transformers is not installed. '
'In order to install all text feature dependencies run '
'pip install ludwig[text]'
)
sys.exit(-1)
self.transformer = TFDistilBertModel.from_pretrained(
pretrained_model_name_or_path
)
self.reduce_output = reduce_output
self.reduce_sequence = SequenceReducer(reduce_mode=reduce_output)
self.transformer.trainable = trainable
self.transformer.resize_token_embeddings(num_tokens)
def run_distilibert(strategy: tf.distribute.TPUStrategy, x_train: np.array,
x_valid: np.array, _y_train: np.array, y_valid: np.array,
train_dataset: tf.data.Dataset,
valid_dataset: tf.data.Dataset,
test_dataset: tf.data.Dataset, max_len: int, epochs: int,
batch_size: int) -> tf.keras.models.Model:
"""
create and run distilbert on training and testing data
"""
logger.info('build distilbert')
with strategy.scope():
transformer_layer = TFDistilBertModel.from_pretrained(MODEL)
model = build_model(transformer_layer, max_len=max_len)
model.summary()
# train given model
n_steps = x_train.shape[0] // batch_size
history = model.fit(train_dataset,
steps_per_epoch=n_steps,
validation_data=valid_dataset,
epochs=epochs)
plot_train_val_loss(history, 'distilbert')
n_steps = x_valid.shape[0] // batch_size
_train_history_2 = model.fit(valid_dataset.repeat(),
steps_per_epoch=n_steps,
epochs=epochs * 2)
scores = model.predict(test_dataset, verbose=1)
logger.info(f"AUC: {roc_auc(scores, y_valid):.4f}")
return model
def distilbert_model(input_shape,
transformer_model,
output_shape=96,
output_activation='softmax',
optimizer='Adam',
optimizer_params={'lr': 1e-5},
loss='categorical_crossentropy',
metrics=None):
input_ids = Input((input_shape, ), dtype=tf.int32)
input_mask = Input((input_shape, ), dtype=tf.int32)
transformer_encoder = TFDistilBertModel.from_pretrained(
transformer_model, from_pt=True, output_hidden_states=True)
outputs = transformer_encoder.distilbert(input_ids,
attention_mask=input_mask)
x = outputs[0]
x = GlobalAveragePooling1D()(x)
output = Dense(output_shape, activation=output_activation)(x)
model = Model(inputs=[input_ids, input_mask], outputs=output)
model.compile(loss=loss,
metrics=metrics,
optimizer=getattr(optimizers, optimizer)(**optimizer_params))
return model
def create_model(model_config: CommentClassifierConfig,
saved_weights_path: str = None,
max_seq_length: int = MAX_SEQ_LENGTH) -> tf.keras.Model:
"""
:param model_config: CommentClassifierConfig
:param saved_weights_path: If defined, model weights will be loaded
from the provided checkpoint path
:param max_seq_length: Maximum length of the tokenized input to BERT
:return:
Model for text classification using DistilBert transformers
"""
# Load pre-trained DistilBERT
bert_config = DistilBertConfig(
dropout=model_config.bert_dropout,
attention_dropout=model_config.bert_attention_dropout,
num_labels=NUM_CLASSES)
bert_config.output_hidden_states = False
transformer_model = TFDistilBertModel.from_pretrained(MODEL_NAME,
config=bert_config)
input_ids_in = tf.keras.layers.Input(shape=(max_seq_length, ),
name='input_token',
dtype='int32')
input_masks_in = tf.keras.layers.Input(shape=(max_seq_length, ),
name='masked_token',
dtype='int32')
embedding_layer = transformer_model(input_ids_in,
attention_mask=input_masks_in)[0]
x = tf.keras.layers.Bidirectional(
tf.keras.layers.LSTM(
model_config.lstm_units,
return_sequences=True,
dropout=model_config.lstm_dropout,
recurrent_dropout=model_config.lstm_recurrent_dropout))(
embedding_layer)
x = tf.keras.layers.GlobalMaxPool1D()(x)
x = tf.keras.layers.Dense(
model_config.hidden_layer_dim,
activation=model_config.hidden_layer_activation)(x)
x = tf.keras.layers.Dropout(model_config.final_layer_dropout)(x)
x = tf.keras.layers.Dense(
NUM_CLASSES, activation=model_config.final_layer_activation)(x)
model = tf.keras.Model(inputs=[input_ids_in, input_masks_in], outputs=x)
# Use transfer learning only - do not train BERT again
for layer in model.layers[:3]:
layer.trainable = False
# Load weights from a checkpoint, but allow partial matching
# (e.g. due to a change in the optimizer)
if saved_weights_path is not None:
model.load_weights(saved_weights_path).expect_partial()
return model
def run_bert_meta_regression_tfmodel():
""" Run self defined combined model."""
timestamp = datetime.now().strftime('%Y%m%d-%H%M%S')
log_dir = os.path.join(os.getenv('OUTPUT_DIR'), timestamp)
model_plot = f'regression_model_{timestamp}.png'
tokenizer = AutoTokenizer.from_pretrained(os.getenv('MODEL_NAME'))
config = AutoConfig.from_pretrained(os.getenv('MODEL_NAME'), num_labels=1)
distilebert_model = TFDistilBertModel.from_pretrained(
os.getenv('MODEL_NAME'), config=config)
print(config, tokenizer, sep='\n')
# tf.keras.utils.plot_model(distilebert_model, to_file=model_plot, show_shapes=True)
tc = TopCoder()
encoded_text = tc.get_bert_encoded_txt_features(tokenizer)
metadata = tc.get_meta_data_features(encoded_tech=True, softmax_tech=True)
target = tc.get_target()
split = int((4 / 5) * len(target))
dataset = tf.data.Dataset.from_tensor_slices(
(dict(**encoded_text, meta_input=metadata), target))
dataset = dataset.shuffle(len(target))
train_ds, test_ds = dataset.take(split).batch(16), dataset.skip(
split).batch(8)
print(train_ds, test_ds, sep='\n')
# for i in train_ds.take(2):
# pprint(i)
# print()
# for i in test_ds.take(2):
# pprint(i)
# model = TCPMDistilBertRegression.from_pretrained(os.getenv('MODEL_NAME'), config=config)
tensorboard_cb = tf.keras.callbacks.TensorBoard(log_dir=log_dir,
histogram_freq=1)
model = build_tcpm_model_distilbert_regression(distilebert_model)
model.summary()
model.compile(optimizer=tf.keras.optimizers.Adam(2e-6),
loss='mse',
metrics=['mae', 'mse', mre])
history = model.fit(
train_ds,
epochs=12,
)
result = model.evaluate(
test_ds,
return_dict=True,
)
pprint(result)
history_df = pd.DataFrame(history.history)
history_df.to_json(os.path.join(log_dir, 'train_history.json'),
orient='index',
indent=4)
with open(os.path.join(log_dir, 'result.json'), 'w') as f:
json.dump(result, f, indent=4)
def test_TFDistilBertModel(self):
from transformers import DistilBertTokenizer, TFDistilBertModel
pretrained_weights = 'distilbert-base-uncased'
tokenizer = DistilBertTokenizer.from_pretrained(pretrained_weights)
text, inputs, inputs_onnx = self._prepare_inputs(tokenizer)
model = TFDistilBertModel.from_pretrained(pretrained_weights)
predictions = model.predict(inputs)
onnx_model = keras2onnx.convert_keras(model, model.name)
self.assertTrue(
run_onnx_runtime(onnx_model.graph.name, onnx_model, inputs_onnx,
predictions, self.model_files))
def get_transformer(bert_model_type, output_hidden_states=False):
config = get_bert_config(bert_model_type, output_hidden_states)
if bert_model_type in [
'bert-base-uncased', 'bert-base-cased', 'bert-large-uncased',
'bert-large-uncased-whole-word-masking',
'bert-large-uncased-whole-word-masking-finetuned-squad'
]:
return TFBertModel.from_pretrained(BERT_MODEL_FILE[bert_model_type],
config=config)
elif bert_model_type in [
'prod-bert-base-uncased', 'tune_bert-base-uncased_nsp'
]:
return TFBertModel.from_pretrained(BERT_MODEL_FILE[bert_model_type],
config=config,
from_pt=True)
elif bert_model_type in [
'roberta-base', 'roberta-large', 'roberta-large-mnli',
'distilroberta-base'
]:
return TFRobertaModel.from_pretrained(BERT_MODEL_FILE[bert_model_type],
config=config)
elif bert_model_type in ['prod-roberta-base-cased']:
return TFRobertaModel.from_pretrained(BERT_MODEL_FILE[bert_model_type],
config=config,
from_pt=True)
elif bert_model_type in ['xlnet-base-cased']:
return TFXLNetModel.from_pretrained(BERT_MODEL_FILE[bert_model_type],
config=config)
elif bert_model_type in [
'albert-base-v1', 'albert-large-v1', 'albert-xlarge-v1',
'albert-xxlarge-v1'
]:
return TFAlbertModel.from_pretrained(BERT_MODEL_FILE[bert_model_type],
config=config)
elif bert_model_type in ['gpt2', 'gpt2-medium']:
return TFGPT2Model.from_pretrained(BERT_MODEL_FILE[bert_model_type],
config=config)
elif bert_model_type in ['transfo-xl']:
return TFTransfoXLModel.from_pretrained(
BERT_MODEL_FILE[bert_model_type], config=config)
elif bert_model_type in [
'distilbert-base-uncased',
'distilbert-base-uncased-distilled-squad'
]:
return TFDistilBertModel.from_pretrained(
BERT_MODEL_FILE[bert_model_type], config=config)
else:
raise ValueError(
f'`bert_model_type` not understood: {bert_model_type}')
def model_arch_multitask():
num_labels = 2
bert = TFDistilBertModel.from_pretrained("distilbert-base-cased")
dropout = tf.keras.layers.Dropout(0.4)
answer_logits = tf.keras.layers.Dense(10, kernel_initializer=tf.keras.initializers.TruncatedNormal(stddev=0.02), name="answer_logits", activation = "softmax")
classifier = tf.keras.layers.Dense(2, kernel_initializer=tf.keras.initializers.TruncatedNormal(stddev=0.02), name="seq_logits")
input_ids = tf.keras.layers.Input(shape = (None,), dtype=tf.int32)
attention_mask = tf.keras.layers.Input(shape = (None,), dtype=tf.int32)
outputs = bert([input_ids, attention_mask])
answer_output = answer_logits(outputs[0][:, 0, :])
sequence_output = outputs[0]
sequence_output = dropout(sequence_output)
logits = classifier(sequence_output)
model = tf.keras.models.Model(inputs = [input_ids, attention_mask], outputs = [logits, answer_output])
model.compile(loss={'seq_logits': custom_loss_logits, 'answer_logits': "categorical_crossentropy"}, optimizer=optimizer, loss_weights = {"answer_logits": 1.0, "seq_logits": 1.0}, metrics=["accuracy"])
return model
def get_transformer(LM: bool):
if LM:
if model_name == 'distilbert-base-cased':
model = TFDistilBertForMaskedLM.from_pretrained(
'distilbert-base-cased')
elif model_name == 'huggingface/CodeBERTa-small-v1':
model = AutoModelWithLMHead.from_pretrained(
'huggingface/CodeBERTa-small-v1')
model = pt_to_tf(model, TFRobertaForMaskedLM)
else:
if model_name == 'distilbert-base-cased':
model = TFDistilBertModel.from_pretrained(
'distilbert-base-cased', )
elif model_name == 'huggingface/CodeBERTa-small-v1':
model = AutoModel.from_pretrained('huggingface/CodeBERTa-small-v1')
model = pt_to_tf(model, TFRobertaModel)
return model
def create_model(max_seq_len, classes):
config = DistilBertConfig(dropout=0.2, attention_dropout=0.2)
config.output_hidden_states = False
tfm = TFDistilBertModel.from_pretrained('./MODEL/uncased/', config=config)
input_ids = keras.layers.Input(shape=(max_seq_len,), dtype='int32', name="input_ids")
bert_output = tfm(input_ids)[0]
cls_out = keras.layers.Lambda(lambda seq: seq[:, 0, :])(bert_output)
cls_out = keras.layers.Dropout(0.5)(cls_out)
logits = keras.layers.Dense(units=768, activation="tanh")(cls_out)
logits = keras.layers.Dropout(0.5)(logits)
logits = keras.layers.Dense(units=512, activation="tanh")(cls_out)
logits = keras.layers.Dropout(0.5)(logits)
logits = keras.layers.Dense(units=256, activation="tanh")(logits)
logits = keras.layers.Dropout(0.5)(logits)
logits = keras.layers.Dense(units=len(classes), activation="softmax")(logits)
model = keras.Model(inputs=input_ids, outputs=logits)
model.build(input_shape=(None, max_seq_len))
return model
def model_arch_tok_classification():
num_labels = 2
max_len = 128
bert = TFDistilBertModel.from_pretrained("distilbert-base-cased")
dropout = tf.keras.layers.Dropout(0.4)
classifier = tf.keras.layers.Dense(2, kernel_initializer=tf.keras.initializers.TruncatedNormal(stddev=0.02), name="seq_logits")
question_input_ids = tf.keras.layers.Input(shape = (max_len,), dtype=tf.int32)
question_attention_mask = tf.keras.layers.Input(shape = (max_len,), dtype=tf.int32)
question_output = bert([question_input_ids, question_attention_mask])
question_output = question_output[0][:, 0, :]
question_output = tf.keras.layers.RepeatVector(max_len)(question_output)
context_input_ids = tf.keras.layers.Input(shape = (max_len,), dtype=tf.int32)
context_attention_mask = tf.keras.layers.Input(shape = (max_len,), dtype=tf.int32)
outputs = bert([context_input_ids, context_attention_mask])
sequence_output = outputs[0]
sequence_output = tf.keras.layers.concatenate([sequence_output, question_output], axis = -1)
sequence_output = dropout(sequence_output)
logits = classifier(sequence_output)
model = tf.keras.models.Model(inputs = [question_input_ids, question_attention_mask, context_input_ids, context_attention_mask], outputs = logits)
model.compile(loss=custom_loss_logits, optimizer=optimizer)
return model
def initialize_hugface_model(hugging_face_model):
# if hugging_face_model == "xlnet":
# tokenizer = XLNetTokenizer.from_pretrained('xlnet-base-cased')
# model = TFXLNetModel.from_pretrained('xlnet-base-cased')
# elif hugging_face_model == "roberta":
# tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
# model = TFRobertaModel.from_pretrained('roberta-base')
# elif hugging_face_model == "ernie":
# tokenizer = AutoTokenizer.from_pretrained("nghuyong/ernie-2.0-en")
# model = TFAutoModel.from_pretrained("nghuyong/ernie-2.0-en")
#FAST TOKENIZERS
if hugging_face_model == "distilbert":
tokenizer = DistilBertTokenizerFast.from_pretrained("distilbert-base-uncased")
model = TFDistilBertModel.from_pretrained("distilbert-base-uncased")
elif hugging_face_model == "bert":
tokenizer = BertTokenizerFast.from_pretrained('bert-base-cased')
model = TFBertModel.from_pretrained('bert-base-cased')
else:
raise ValueError('Invalid embedding type')
return tokenizer, model
(x_valid, y_valid)).batch(BATCH_SIZE).cache().prefetch(AUTO))
# test_dataset = (
# tf.data.Dataset
# .from_tensor_slices(x_test)
# .batch(BATCH_SIZE)
# )
###################################################################
# LOAD MODEL
###################################################################
print("loading model ...")
with strategy.scope():
#transformer_layer = TFAutoModel.from_pretrained(MODEL)
transformer_layer = TFDistilBertModel.from_pretrained(MODEL)
model = HelperFns.build_model(transformer_layer, max_len=MAX_LEN)
model.summary()
###################################################################
# TRAINING
###################################################################
print("run training ...")
n_steps = x_train.shape[0] // BATCH_SIZE
train_history = model.fit(train_dataset,
steps_per_epoch=n_steps,
validation_data=valid_dataset,
epochs=EPOCHS)
def build_custom_model(self, validation_data, validation_label):
# --------------------------------------------------------------------------------
# Input layer (token indices and attention masks)
# --------------------------------------------------------------------------------
input_ids = tf.keras.layers.Input(shape=(self.max_sequence_length, ),
dtype=tf.int32,
name='input_ids')
attention_mask = tf.keras.layers.Input((self.max_sequence_length, ),
dtype=tf.int32,
name='attention_mask')
# --------------------------------------------------------------------------------
# Base layer
# --------------------------------------------------------------------------------
# TFBaseModelOutput.last_hidden_state has shape (batch_size, max_sequence_length, 768)
# Each sequence has [CLS]...[SEP] structure of shape (max_sequence_length, 768)
# Extract [CLS] embeddings of shape (batch_size, 768) as last_hidden_state[:, 0, :]
# --------------------------------------------------------------------------------
base = TFDistilBertModel.from_pretrained(self.model_name, )
# Freeze the base model weights.
if self.freeze_pretrained_base_model:
for layer in base.layers:
layer.trainable = False
base.summary()
output = base([input_ids, attention_mask]).last_hidden_state[:, 0, :]
# --------------------------------------------------------------------------------
# TODO:
# Need to verify the effect of regularizers.
#
# [bias regularizer]
# It looks bias_regularizer adjusts the ROC threshold towards 0.5.
# Without it, the threshold of the ROC with BinaryCrossEntropy loss was approx 0.02.
# With it, the threshold of the ROC with BinaryCrossEntropy loss was approx 0.6.
# --------------------------------------------------------------------------------
activation = "sigmoid" if self.num_labels == 1 else "softmax"
output = tf.keras.layers.Dense(
units=self.num_labels,
kernel_initializer='glorot_uniform',
# https://huggingface.co/transformers/v4.3.3/main_classes/optimizer_schedules.html#adamweightdecay-tensorflow
# kernel_regularizer=tf.keras.regularizers.l2(l2=self.l2),
# bias_regularizer=tf.keras.regularizers.l2(l2=self.l2),
# activity_regularizer=tf.keras.regularizers.l2(l2=self.l2/10.0),
activation=activation,
name=activation)(output)
# --------------------------------------------------------------------------------
# Loss layer
# --------------------------------------------------------------------------------
if self.num_labels == 1: # Binary classification
loss_fn = tf.keras.losses.BinaryCrossentropy(from_logits=False)
else: # Categorical classification
loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(
from_logits=False)
# --------------------------------------------------------------------------------
# Model Metrics
# --------------------------------------------------------------------------------
if self.num_labels == 1:
if self.USE_METRIC_AUC: # ROC/AUC
# AUC is for Binary Classification. Error if used for categorical"
# "alueError: Shapes (None, <num_classes>) and (None, 1) are incompatible"
# Because AUC is expecting shape(None, 1) as binary input into the loss fn.
self._metric_name = "auc"
self._monitor_metric = f"val_{self._metric_name}"
self._monitor_mode = 'max'
self._metrics = [
tf.keras.metrics.AUC(from_logits=False,
name=self._metric_name),
tf.keras.metrics.Recall(name="recall"), "accuracy"
]
self._callbacks = self.build_custom_model_auc_callbacks(
validation_data, validation_label)
else:
self._metric_name = "recall" # Recall
self._monitor_metric = f"val_{self._metric_name}"
self._monitor_mode = 'max'
self._metrics = [
tf.keras.metrics.Recall(name=self._metric_name), "accuracy"
]
self._callbacks = self.build_custom_model_acc_callbacks()
else: # Validation loss
self._metric_name = "accuracy"
self._monitor_metric = "val_loss"
self._monitor_mode = 'min'
# metrics=[tf.keras.metrics.Accuracy(name=metric_name)]
self._metrics = [self._metric_name]
self._callbacks = self.build_custom_model_acc_callbacks()
# --------------------------------------------------------------------------------
# Build model
# --------------------------------------------------------------------------------
# TODO: Replace TIMESTAMP with instance variable
name = f"{TIMESTAMP}_{self.model_name.upper()}"
self._model = tf.keras.models.Model(inputs=[input_ids, attention_mask],
outputs=output,
name=name)
self.model.compile(
# https://huggingface.co/transformers/v4.3.3/main_classes/optimizer_schedules.html#adamweightdecay-tensorflow
# optimizer=tf.keras.optimizers.Adam(learning_rate=self.learning_rate),
optimizer=transformers.AdamWeightDecay(
learning_rate=self.learning_rate),
loss=loss_fn,
metrics=self._metrics)
# --------------------------------------------------------------------------------
# Load model parameters if the saved weight file exits
# --------------------------------------------------------------------------------
path_to_h5 = self.model_directory + os.path.sep + "model.h5"
if os.path.isfile(path_to_h5) and os.access(path_to_h5, os.R_OK):
print(
f"\nloading the saved model parameters from {path_to_h5}...\n")
self.model.load_weights(path_to_h5)
dataset_test=tf.data.Dataset.from_tensor_slices((Xids_test,Xmask_test))
def map_func(input_ids,mask):
return {'input_ids':input_ids,'attention_mask':mask}
dataset_test=dataset_test.map(map_func)
dataset_test=dataset_test.batch(32).prefetch(1000)
#Build the model
from transformers import TFDistilBertModel, DistilBertConfig
distil_bert = 'distilbert-base-uncased'
config = DistilBertConfig(dropout=0.2, attention_dropout=0.2)
config.output_hidden_states = False
transformer_model = TFDistilBertModel.from_pretrained(distil_bert, config = config)
input_ids_in = tf.keras.layers.Input(shape=(SEQ_length,), name='input_ids', dtype='int32')
input_masks_in = tf.keras.layers.Input(shape=(SEQ_length,), name='attention_mask', dtype='int32')
embedding_layer = transformer_model(input_ids_in, attention_mask=input_masks_in)[0]
X = tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(50, return_sequences=True, dropout=0.1, recurrent_dropout=0.1))(embedding_layer)
X = tf.keras.layers.GlobalMaxPool1D()(X)
X = tf.keras.layers.Dense(50, activation='relu')(X)
X = tf.keras.layers.Dropout(0.2)(X)
X = tf.keras.layers.Dense(1, activation='sigmoid')(X)
model = tf.keras.Model(inputs=[input_ids_in, input_masks_in], outputs = X)
for layer in model.layers[:3]:
layer.trainable = False
data = pd.read_csv('/content/drive/My Drive/data/train_E6oV3lV.csv').sample(frac=0.3)
X_train, X_test, y_train, y_test = train_test_split(data.tweet, data.label)
model = Pipeline([
('vect', TfidfVectorizer(
stop_words='english',
# ngram_range=(1, 3)
)),
('clf', SGDClassifier()),
])
model.fit(X_train, y_train)
print('tfidf f1:', f1_score(y_test, model.predict(X_test), average='binary'))
model = TFDistilBertModel.from_pretrained('distilbert-base-uncased')
tokenizer = AutoTokenizer.from_pretrained('distilbert-base-uncased',
# pad_to_max_length=True,
# max_length=100000
)
pipe = pipeline('feature-extraction', model=model,
tokenizer=tokenizer)
features = pipe(X_train.to_list(),
# pad_to_max_length=True
)
print()
# model = TFDistilBertModel.from_pretrained('distilbert-base-uncased')
# tokenizer = AutoTokenizer.from_pretrained('distilbert-base-uncased')
# pipe = pipeline('feature-extraction', model=model,
0: 1,
1: 2,
2: 3,
3: 4,
4: 5
},
label2id={
1: 0,
2: 1,
3: 2,
4: 3,
5: 4
},
)
transformer_model = TFDistilBertModel.from_pretrained(
"distilbert-base-uncased", config=config)
input_ids = tf.keras.layers.Input(shape=(max_seq_length, ),
name="input_ids",
dtype="int32")
input_mask = tf.keras.layers.Input(shape=(max_seq_length, ),
name="input_mask",
dtype="int32")
embedding_layer = transformer_model.distilbert(
input_ids, attention_mask=input_mask)[0]
X = tf.keras.layers.Bidirectional(
tf.keras.layers.LSTM(
50,
return_sequences=True,
dropout=0.1,
from train import train, trainLiar, trainPoliti
from tensorflow.keras import regularizers, initializers, optimizers, callbacks
from tensorflow.keras.layers import *
import pandas as pd
import numpy as np
import keras
import tensorflow as tf
from transformers import TFDistilBertModel, DistilBertConfig
# DistilBERT
config = DistilBertConfig(dropout=0.2,
attention_dropout=0.2,
output_hidden_states=True)
dbert_model = TFDistilBertModel.from_pretrained(
'distilbert-base-uncased', config=config)
# Define model using Keras functional API
def buildModel(seq_length, md_length, sco_length, his_length, n_output1, n_output2):
input_ids1 = Input(shape=(seq_length, ),
dtype=tf.int32, name="input_ids1")
attention_mask1 = Input(shape=(seq_length, ),
dtype=tf.int32, name="attention_mask1")
input_ids2 = Input(shape=(md_length, ),
dtype=tf.int32, name="input_ids2")
attention_mask2 = Input(shape=(md_length, ),
dtype=tf.int32, name="attention_mask2")
score = Input(shape=(sco_length,), name="score")
history = Input(shape=(his_length,), name="history")
import numpy as np
import pandas as pd
import tensorflow as tf
import transformers
from sklearn.model_selection import train_test_split
from sklearn.metrics import roc_auc_score
import re
from transformers import TFDistilBertModel, DistilBertTokenizer
model_name = 'distilbert-base-uncased'
pretrained_model = TFDistilBertModel.from_pretrained(model_name)
tokenizer = DistilBertTokenizer.from_pretrained(model_name)
train_path = 'data\\train.csv'
test_path = 'data\\test.csv'
input_column = 'tweet'
label_column = 'label'
# ============= load dataset ===============
train_df = pd.read_csv(train_path)
test_df = pd.read_csv(test_path)
# --- get classes and convert to categorical ---
label_cols = train_df[label_column].unique()
if type(label_cols) == np.ndarray:
train_df["target"] = train_df[label_column]
# train_df.pop(label_column)
num_classes = label_cols.shape[0]
else:
num_classes = len(label_column)
def feature_extracter_from_texts(self, mashup_api=None):
"""
对mashup，service的description均需要提取特征，右路的文本的整个特征提取过程
公用的话应该封装成新的model！
:param mashup_api: 默认是None，只有'HDP'/'Bert'时为非空
:return: 输出的是一个封装好的model，所以可以被mashup和api公用
"""
if self.args.text_extracter_mode in fixed_vector_modes and mashup_api is not None:
if self.args.text_extracter_mode == 'Bert':
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
bertModel = BertModel.from_pretrained("bert-base-uncased")
if mashup_api == 'mashup':
if self.mashup_text_feature_extracter is None: # 没求过
mashup_texts = get_iterable_values(
data_repository.get_md().mashup_df,
'final_description',
return_ele_type='str')
dense_mashup_features = bertModel(
tokenizer(mashup_texts, return_tensors='tf'))
self.mashup_text_feature_extracter = vector_feature_extracter_from_texts(
'mashup', dense_mashup_features)
return self.mashup_text_feature_extracter
elif mashup_api == 'api':
if self.api_text_feature_extracter is None:
api_texts = get_iterable_values(
data_repository.get_md().api_df,
'final_description',
return_ele_type='str')
dense_api_features = bertModel(
tokenizer(api_texts, return_tensors='tf'))
self.api_text_feature_extracter = vector_feature_extracter_from_texts(
'api', dense_api_features)
return self.api_text_feature_extracter
else:
raise TypeError('wrong mashup_api mode!')
else:
if self.gd is None:
self.gd = get_default_gd(
tag_times=0, mashup_only=False,
strict_train=True) # 用gensim处理文本,文本中不加tag
self.gd.model_pcs(self.args.text_extracter_mode) #
if mashup_api == 'mashup':
if self.mashup_text_feature_extracter is None: # 没求过
self.mashup_text_feature_extracter = vector_feature_extracter_from_texts(
'mashup', self.gd.dense_mashup_features)
return self.mashup_text_feature_extracter
elif mashup_api == 'api':
if self.api_text_feature_extracter is None:
self.api_text_feature_extracter = vector_feature_extracter_from_texts(
'api', self.gd.dense_api_features)
return self.api_text_feature_extracter
else:
raise TypeError('wrong mashup_api mode!')
elif self.text_feature_extracter is None: # 没求过
if 'trainable_bert' in self.args.text_extracter_mode.lower():
self.text_feature_extracter = TFDistilBertModel.from_pretrained(
"distilbert-base-uncased") # layer
if self.args.frozen_bert:
self.text_feature_extracter.trainable = False
else:
text_input = Input(shape=(self.args.MAX_SEQUENCE_LENGTH, ),
dtype='int32')
text_embedding_layer = self.get_text_embedding_layer(
) # 参数还需设为外部输入！
text_embedded_sequences = text_embedding_layer(
text_input) # 转化为2D
if self.args.text_extracter_mode in (
'inception', 'textCNN'): # 2D转3D,第三维是channel
# print(text_embedded_sequences.shape)
text_embedded_sequences = Lambda(
lambda x: tf.expand_dims(x, axis=3))(
text_embedded_sequences) # tf 和 keras的tensor 不同！！！
print(text_embedded_sequences.shape)
if self.args.text_extracter_mode == 'inception':
x = inception_layer(
text_embedded_sequences, self.args.embedding_dim,
self.args.inception_channels,
self.args.inception_pooling) # inception处理
print('built inception layer, done!')
elif self.args.text_extracter_mode == 'textCNN':
x = textCNN_feature_extracter_from_texts(
text_embedded_sequences, self.args)
elif self.args.text_extracter_mode == 'LSTM':
x = LSTM_feature_extracter_from_texts(
text_embedded_sequences, self.args)
else:
raise TypeError('wrong extracter!')
print('text feature after inception/textCNN/LSTM whole_model,',
x) # 观察MLP转化前，模块输出的特征
for FC_unit_num in self.args.inception_fc_unit_nums:
x = Dense(FC_unit_num,
kernel_regularizer=l2(self.args.l2_reg))(
x) # , activation='relu'
if self.args.inception_MLP_BN:
x = BatchNormalization(scale=False)(x)
x = PReLU()(x) #
if self.args.inception_MLP_dropout:
x = tf.keras.layers.Dropout(0.5)(x)
self.text_feature_extracter = Model(
text_input, x, name='text_feature_extracter')
return self.text_feature_extracter