def build_transformer_model_for_pretraining(): """构建训练模型,通用于TPU/GPU 注意全程要用keras标准的层写法,一些比较灵活的“移花接木”式的 写法可能会在TPU上训练失败。此外,要注意的是TPU并非支持所有 tensorflow算子,尤其不支持动态(变长)算子,因此编写相应运算 时要格外留意。 """ bert, train_model, loss = build_transformer_model_with_mlm() # 优化器 optimizer = extend_with_weight_decay(Adam) if which_optimizer == 'lamb': optimizer = extend_with_layer_adaptation(optimizer) optimizer = extend_with_piecewise_linear_lr(optimizer) optimizer_params = { 'learning_rate': learning_rate, 'lr_schedule': lr_schedule, 'weight_decay_rate': weight_decay_rate, 'exclude_from_weight_decay': exclude_from_weight_decay, 'bias_correction': False, } if grad_accum_steps > 1: optimizer = extend_with_gradient_accumulation(optimizer) optimizer_params['grad_accum_steps'] = grad_accum_steps optimizer = optimizer(**optimizer_params) # 模型定型 train_model.compile(loss=loss, optimizer=optimizer) # 如果传入权重,则加载。注:须在此处加载,才保证不报错。 if checkpoint_path is not None: bert.load_weights_from_checkpoint(checkpoint_path) return train_model
def build(self): bert_model, _ = load_bert( config_path=os.path.join(self.config['pretrained_model_dir'], 'bert_config.json'), checkpoint_path=os.path.join(self.config['pretrained_model_dir'], 'bert_model.ckpt'), ) text_mask = L.Lambda( lambda x: K.cast(K.expand_dims(K.greater(x, 0), 2), K.floatx()))( bert_model.input[0]) # GI gi_in = L.Input(name="gi", shape=(self.config["max_len"], ), dtype="float32") gi = gi_in # AGN X = bert_model.output gi = L.Dense(self.config['max_len'], activation='tanh')(gi) # (B, L) gi = L.Lambda(lambda x: K.expand_dims(x, 2))(gi) # (B, L, 1) X, attn_weight = AGN(epsilon=self.config['epsilon'])([X, gi]) X = L.Lambda(lambda x: x[0] - 1e10 * (1.0 - x[1]))([X, text_mask]) output = L.Lambda(lambda x: K.max(x, 1))(X) #output = L.Dense(128, activation='relu')(output) output = L.Dropout(self.config.get('dropout', 0.2))(output) output = L.Dense(self.config['output_size'], activation='softmax')(output) self.model = keras.Model(inputs=(*bert_model.input, gi_in), outputs=output) self.attn_model = keras.Model(inputs=(*bert_model.input, gi_in), outputs=attn_weight) optimizer = extend_with_weight_decay(Adam) optimizer = extend_with_piecewise_linear_lr(optimizer) optimizer_params = { 'learning_rate': self.config['learning_rate'], 'lr_schedule': { self.config['steps_per_epoch'] * 2: 1, self.config['steps_per_epoch'] * 3: 0.2, self.config['steps_per_epoch'] * self.config['epochs']: 0.1 }, 'weight_decay_rate': 0.01, 'exclude_from_weight_decay': ['Norm', 'bias'], 'bias_correction': False, } self.model.compile( loss='sparse_categorical_crossentropy', optimizer=optimizer(**optimizer_params), ) self.model.summary() if self.config.get('apply_fgm', True): print('apply fgm') fgm(self.model, 'Embedding-Token', self.config.get('fgm_epsilon', 0.2))
def get_suggested_optimizer(init_lr=5e-5, total_steps=None): lr_schedule = {1000: 1, 10000: 0.01} if total_steps is not None: lr_schedule = {total_steps // 10: 1, total_steps: 0.1} optimizer = extend_with_weight_decay(Adam) optimizer = extend_with_piecewise_linear_lr(optimizer) optimizer_params = { 'learning_rate': init_lr, 'lr_schedule': lr_schedule, 'weight_decay_rate': 0.01, 'exclude_from_weight_decay': ['Norm', 'bias'], 'bias_correction': False, } optimizer = optimizer(**optimizer_params) return optimizer
def build_train_bert_model(): """构建训练模型,通用于TPU/GPU 注意全程要用keras标准的层写法,一些比较灵活的“移花接木”式的 写法可能会在TPU上训练失败。此外,要注意的是TPU并非支持所有 tensorflow算子,尤其不支持动态(变长)算子,因此编写相应运算 时要格外留意。 """ bert = build_bert_model(config_path, with_mlm='linear', application='lm', return_keras_model=False) token_ids = bert.model.input[0] proba = bert.model.output def lm_loss(inputs): """计算loss的函数,需要封装为一个层 """ y_true, y_pred, mask = inputs y_true = y_true[:, 1:] y_pred = y_pred[:, :-1] mask = mask[:, 1:] loss = K.sparse_categorical_crossentropy(y_true, y_pred, from_logits=True) loss = K.sum(loss * mask) / (K.sum(mask) + K.epsilon()) return loss def lm_acc(inputs): """计算准确率的函数,需要封装为一个层 """ y_true, y_pred, mask = inputs y_true = K.cast(y_true, K.floatx()) y_true = y_true[:, 1:] y_pred = y_pred[:, :-1] mask = mask[:, 1:] acc = keras.metrics.sparse_categorical_accuracy(y_true, y_pred) acc = K.sum(acc * mask) / (K.sum(mask) + K.epsilon()) return acc mask = bert.model.get_layer('Sequence-Mask').output loss = Lambda(lm_loss, name='lm_loss')([token_ids, proba, mask]) acc = Lambda(lm_acc, name='lm_acc')([token_ids, proba, mask]) train_model = Model(bert.model.inputs, [loss, acc]) # 优化器 optimizer = extend_with_weight_decay(Adam) if which_optimizer == 'lamb': optimizer = extend_with_layer_adaptation(optimizer) optimizer = extend_with_piecewise_linear_lr(optimizer) optimizer_params = { 'learning_rate': learning_rate, 'lr_schedule': lr_schedule, 'weight_decay_rate': weight_decay_rate, 'exclude_from_weight_decay': exclude_from_weight_decay, 'bias_correction': False, } if grad_accum_steps > 1: optimizer = extend_with_gradient_accumulation(optimizer) optimizer_params['grad_accum_steps'] = grad_accum_steps optimizer = optimizer(**optimizer_params) # 模型定型 train_model.compile( loss={ 'lm_loss': lambda y_true, y_pred: y_pred, 'lm_acc': lambda y_true, y_pred: K.stop_gradient(y_pred), }, optimizer=optimizer, ) # 如果传入权重,则加载。注:须在此处加载,才保证不报错。 if checkpoint_path is not None: bert.load_weights_from_checkpoint(checkpoint_path) return train_model
model = build_transformer_model( config_path, checkpoint_path, model='nezha', application='lm', keep_tokens=keep_tokens, # 只保留keep_tokens中的字,精简原字表 compound_tokens=compound_tokens, # 要扩充的词表 ) output = CrossEntropy(1)([model.inputs[0], model.outputs[0]]) model = Model(model.inputs, output) model.summary() AdamW = extend_with_weight_decay(Adam, 'AdamW') AdamWG = extend_with_gradient_accumulation(AdamW, 'AdamWG') optimizer = AdamWG(learning_rate=2e-5, weight_decay_rate=0.01, exclude_from_weight_decay=['Norm', 'bias'], grad_accum_steps=16) model.compile(optimizer=optimizer) class ChatBot(AutoRegressiveDecoder): """基于随机采样对话机器人 """ @AutoRegressiveDecoder.wraps(default_rtype='probas') def predict(self, inputs, output_ids, states): token_ids, segment_ids = inputs token_ids = np.concatenate([token_ids, output_ids], 1)
model = build_transformer_model( config_path, checkpoint_path, with_mlm='linear', keep_tokens=keep_tokens, # 只保留keep_tokens中的字,精简原字表 compound_tokens=compound_tokens, # 增加词,用字平均来初始化 ) # 训练用模型 y_in = keras.layers.Input(shape=(None, )) outputs = CrossEntropy(1)([y_in, model.output]) train_model = keras.models.Model(model.inputs + [y_in], outputs) AdamW = extend_with_weight_decay(Adam, name='AdamW') AdamWG = extend_with_gradient_accumulation(AdamW, name='AdamWG') optimizer = AdamWG( learning_rate=5e-6, weight_decay_rate=0.01, exclude_from_weight_decay=['Norm', 'bias'], grad_accum_steps=16, ) train_model.compile(optimizer=optimizer) train_model.summary() class Evaluator(keras.callbacks.Callback): """训练回调 """ def on_epoch_end(self, epoch, logs=None):
def build_train_bert_model(): """构建训练模型,通用于TPU/GPU 注意全程要用keras标准的层写法,一些比较灵活的“移花接木”式的 写法可能会在TPU上训练失败。此外,要注意的是TPU并非支持所有 tensorflow算子,尤其不支持动态(变长)算子,因此编写相应运算 时要格外留意。 """ bert = build_bert_model(config_path, with_mlm='linear', return_keras_model=False) bert_model = bert.model proba = bert_model.output # 辅助输入 token_ids = Input(shape=(None, ), dtype='int64', name='token_ids') # 目标id is_masked = Input(shape=(None, ), dtype='bool', name='is_masked') # mask标记 def mlm_loss(inputs): """计算loss的函数,需要封装为一个层 """ y_true, y_pred, is_masked = inputs is_masked = K.cast(is_masked, K.floatx()) loss = K.sparse_categorical_crossentropy(y_true, y_pred, from_logits=True) loss = K.sum(loss * is_masked) / (K.sum(is_masked) + K.epsilon()) return loss def mlm_acc(inputs): """计算准确率的函数,需要封装为一个层 """ y_true, y_pred, is_masked = inputs is_masked = K.cast(is_masked, K.floatx()) y_true = K.cast(y_true, K.floatx()) acc = keras.metrics.sparse_categorical_accuracy(y_true, y_pred) acc = K.sum(acc * is_masked) / (K.sum(is_masked) + K.epsilon()) return acc loss = Lambda(mlm_loss, name='mlm_loss')([token_ids, proba, is_masked]) acc = Lambda(mlm_acc, name='mlm_acc')([token_ids, proba, is_masked]) train_model = Model(bert_model.inputs + [token_ids, is_masked], [loss, acc]) # 优化器 optimizer = extend_with_weight_decay(Adam) if which_optimizer == 'lamb': optimizer = extend_with_layer_adaptation(optimizer) optimizer = extend_with_piecewise_linear_lr(optimizer) optimizer_params = { 'learning_rate': learning_rate, 'lr_schedule': lr_schedule, 'weight_decay_rate': weight_decay_rate, 'exclude_from_weight_decay': exclude_from_weight_decay, 'bias_correction': False, } if grad_accum_steps > 1: optimizer = extend_with_gradient_accumulation(optimizer) optimizer_params['grad_accum_steps'] = grad_accum_steps optimizer = optimizer(**optimizer_params) # 模型定型 train_model.compile( loss={ 'mlm_loss': lambda y_true, y_pred: y_pred, 'mlm_acc': lambda y_true, y_pred: K.stop_gradient(y_pred), }, optimizer=optimizer, ) # 如果传入权重,则加载。注:须在此处加载,才保证不报错。 if checkpoint_path is not None: bert.load_weights_from_checkpoint(checkpoint_path) return train_model