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_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
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