def test_google_weights(self): albert_model_name = "albert_base" albert_dir = bert.fetch_tfhub_albert_model(albert_model_name, ".models") albert_params = bert.albert_params(albert_model_name) l_bert = bert.BertModelLayer.from_params(albert_params, name="albert") l_input_ids = keras.layers.Input(shape=(128, ), dtype='int32', name="input_ids") l_token_type_ids = keras.layers.Input(shape=(128, ), dtype='int32', name="token_type_ids") output = l_bert([l_input_ids, l_token_type_ids]) output = keras.layers.Lambda(lambda x: x[:, 0, :])(output) output = keras.layers.Dense(2)(output) model = keras.Model(inputs=[l_input_ids, l_token_type_ids], outputs=output) model.build(input_shape=(None, 128)) model.compile( optimizer=keras.optimizers.Adam(), loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True), metrics=[keras.metrics.SparseCategoricalAccuracy(name="acc")]) for weight in l_bert.weights: print(weight.name) bert.load_albert_weights(l_bert, albert_dir) model.summary()
def test_albert_load_base_google_weights(self): # for coverage mainly albert_model_name = "albert_base" albert_dir = bert.fetch_tfhub_albert_model(albert_model_name, ".models") model_params = bert.albert_params(albert_model_name) l_bert = bert.BertModelLayer.from_params(model_params, name="albert") model = keras.models.Sequential([ keras.layers.InputLayer(input_shape=(8, ), dtype=tf.int32, name="input_ids"), l_bert, keras.layers.Lambda(lambda x: x[:, 0, :]), keras.layers.Dense(2), ]) model.build(input_shape=(None, 8)) model.compile( optimizer=keras.optimizers.Adam(), loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True), metrics=[keras.metrics.SparseCategoricalAccuracy(name="acc")]) bert.load_albert_weights(l_bert, albert_dir) model.summary()
def test_albert_google_weights(self): albert_model_name = "albert_base" albert_dir = bert.fetch_tfhub_albert_model(albert_model_name, ".models") albert_params = bert.albert_params(albert_model_name) model, l_bert = self.build_model(albert_params) skipped_weight_value_tuples = bert.load_albert_weights(l_bert, albert_dir) self.assertEqual(0, len(skipped_weight_value_tuples)) model.summary()
def test_albert_params(self): albert_model_name = "albert_base" albert_dir = bert.fetch_tfhub_albert_model(albert_model_name, ".models") dir_params = bert.albert_params(albert_dir) dir_params.attention_dropout = 0.1 # diff between README and assets/albert_config.json dir_params.hidden_dropout = 0.1 name_params = bert.albert_params(albert_model_name) self.assertEqual(name_params, dir_params) # coverage model_params = dir_params model_params.vocab_size = model_params.vocab_size + 2 model_params.adapter_size = 1 l_bert = bert.BertModelLayer.from_params(model_params, name="albert") l_bert(tf.zeros((1, 128))) bert.load_albert_weights(l_bert, albert_dir)
def Albert_model(max_seq_len): model_name = "albert_large" model_dir = bert.fetch_tfhub_albert_model(model_name, ".models") model_params = bert.albert_params(model_name) model_params.shared_layer = True model_params.embedding_size = 1024 l_bert = bert.BertModelLayer.from_params(model_params, name="albert") l_input_ids = keras.layers.Input(shape=(max_seq_len, ), dtype='int32') # using the default token_type/segment id 0 output = l_bert(l_input_ids) # output: [batch_size, max_seq_len, hidden_size] output = keras.layers.GlobalAveragePooling1D()(output) model = keras.Model(inputs=l_input_ids, outputs=output) model.build(input_shape=(None, max_seq_len)) # use in a Keras Model here, and call model.build() bert.load_albert_weights(l_bert, model_dir) # should be called after model.build() return model, model_dir
def load_bert_model(name_model, max_seq_len, trainable=False): """ models name supported, same as tf-2.0-bert """ model_name = name_model model_dir = bert.fetch_tfhub_albert_model(model_name, ".models") model_params = bert.albert_params(model_name) l_bert = bert.BertModelLayer.from_params(model_params, name=name_model) l_input_ids = tf.keras.layers.Input(shape=(max_seq_len, ), dtype='int32') output = l_bert( l_input_ids) # output: [batch_size, max_seq_len, hidden_size] model = tf.keras.Model(inputs=l_input_ids, outputs=output) model.build(input_shape=(None, max_seq_len)) # load google albert original weights after the build bert.load_albert_weights(l_bert, model_dir) model.trainable = trainable return model
def build_transformer(transformer, max_seq_length=None, num_labels=None, tagging=True, tokenizer_only=False): spm_model_file = None if transformer in zh_albert_models_google: from bert.tokenization.albert_tokenization import FullTokenizer model_url = zh_albert_models_google[transformer] albert = True elif transformer in albert_models_tfhub: from edparser.layers.transformers.albert_tokenization import FullTokenizer with stdout_redirected(to=os.devnull): model_url = fetch_tfhub_albert_model(transformer, os.path.join(hanlp_home(), 'thirdparty', 'tfhub.dev', 'google', transformer)) albert = True spm_model_file = glob.glob(os.path.join(model_url, 'assets', '*.model')) assert len(spm_model_file) == 1, 'No vocab found or unambiguous vocabs found' spm_model_file = spm_model_file[0] elif transformer in bert_models_google: from bert.tokenization.bert_tokenization import FullTokenizer model_url = bert_models_google[transformer] albert = False else: raise ValueError( f'Unknown model {transformer}, available ones: {list(bert_models_google.keys()) + list(zh_albert_models_google.keys()) + list(albert_models_tfhub.keys())}') bert_dir = get_resource(model_url) if spm_model_file: vocab = glob.glob(os.path.join(bert_dir, 'assets', '*.vocab')) else: vocab = glob.glob(os.path.join(bert_dir, '*vocab*.txt')) assert len(vocab) == 1, 'No vocab found or unambiguous vocabs found' vocab = vocab[0] lower_case = any(key in transformer for key in ['uncased', 'multilingual', 'chinese', 'albert']) if spm_model_file: # noinspection PyTypeChecker tokenizer = FullTokenizer(vocab_file=vocab, spm_model_file=spm_model_file, do_lower_case=lower_case) else: tokenizer = FullTokenizer(vocab_file=vocab, do_lower_case=lower_case) if tokenizer_only: return tokenizer if spm_model_file: bert_params = albert_params(bert_dir) else: bert_params = bert.params_from_pretrained_ckpt(bert_dir) l_bert = bert.BertModelLayer.from_params(bert_params, name='albert' if albert else "bert") if not max_seq_length: return l_bert, tokenizer, bert_dir l_input_ids = tf.keras.layers.Input(shape=(max_seq_length,), dtype='int32', name="input_ids") l_mask_ids = tf.keras.layers.Input(shape=(max_seq_length,), dtype='int32', name="mask_ids") l_token_type_ids = tf.keras.layers.Input(shape=(max_seq_length,), dtype='int32', name="token_type_ids") output = l_bert([l_input_ids, l_token_type_ids], mask=l_mask_ids) if not tagging: output = tf.keras.layers.Lambda(lambda seq: seq[:, 0, :])(output) if bert_params.hidden_dropout: output = tf.keras.layers.Dropout(bert_params.hidden_dropout, name='hidden_dropout')(output) logits = tf.keras.layers.Dense(num_labels, kernel_initializer=tf.keras.initializers.TruncatedNormal( bert_params.initializer_range))(output) model = tf.keras.Model(inputs=[l_input_ids, l_mask_ids, l_token_type_ids], outputs=logits) model.build(input_shape=(None, max_seq_length)) if not spm_model_file: ckpt = glob.glob(os.path.join(bert_dir, '*.index')) assert ckpt, f'No checkpoint found under {bert_dir}' ckpt, _ = os.path.splitext(ckpt[0]) with stdout_redirected(to=os.devnull): if albert: if spm_model_file: skipped_weight_value_tuples = bert.load_albert_weights(l_bert, bert_dir) else: # noinspection PyUnboundLocalVariable skipped_weight_value_tuples = load_stock_weights(l_bert, ckpt) else: # noinspection PyUnboundLocalVariable skipped_weight_value_tuples = bert.load_bert_weights(l_bert, ckpt) assert 0 == len(skipped_weight_value_tuples), f'failed to load pretrained {transformer}' return model, tokenizer
from datetime import datetime import bert from bert.tokenization import FullTokenizer from tensorflow import keras from helper import create_learn_rate_scheduler, f1_score MAX_SEQ_LEN = 128 ADAPTER_SIZE = None # Use None for Fine-Tuning MODEL_NAME = "albert_base" MODEL_URL = 'https://tfhub.dev/google/albert_base/2?tf-hub-format=compressed' CHECKPOINT_DIR = 'checkpoints' MODEL_DIR = bert.fetch_tfhub_albert_model(MODEL_NAME, CHECKPOINT_DIR) LOG_DIR = ".log/" + datetime.now().strftime("%Y%m%d-%H%M%s") tensorboard_callback = keras.callbacks.TensorBoard(log_dir=LOG_DIR) def flatten_layers(root_layer): if isinstance(root_layer, keras.layers.Layer): yield root_layer for layer in root_layer._layers: yield from flatten_layers(layer) def freeze_layers(root_layer, exclude=None): exclude = [] if exclude is None else exclude root_layer.trainable = False for layer in flatten_layers(root_layer): if layer.name in exclude: