继BERT维度剪枝之后,尝试了BERT层数暴力裁剪,直接剪掉若干层。
试验结果:
结论:训练提升40%左右、效果下降1.2%,推断速度提升50%。
代码参考 : 最简单的模型轻量化方法:20行代码为BERT剪枝 最简单的模型轻量化方法:20行代码为BERT剪枝-腾讯云开发者社区-腾讯云, 但是进行了一些调整。
1)首先,将谷歌pretrain的模型直接裁剪另存:
=====L-2
import tensorflow as tf
import os
sess = tf.Session()
last_name = 'bert_model.ckpt'
model_path = './'
imported_meta = tf.train.import_meta_graph(os.path.join(model_path, last_name + '.meta'))
imported_meta.restore(sess, os.path.join(model_path, last_name))
init_op=tf.local_variables_initializer()
sess.run(init_op)
bert_dict = {}
for var in tf.global_variables():
if var.name.startswith('bert/encoder/layer_') and not var.name.startswith('bert/encoder/layer_0') and not var.name.startswith('bert/encoder/layer_11'):
pass
else:
bert_dict[var.name]=sess.run(var).tolist()
need_vars = []
for var in tf.global_variables():
if var.name.startswith('bert/encoder/layer_') and not var.name.startswith('bert/encoder/layer_0/') and not var.name.startswith('bert/encoder/layer_1/'):
pass
elif var.name.startswith('bert/encoder/layer_1/'):
new_name = var.name.replace("bert/encoder/layer_1","bert/encoder/layer_11")
op=tf.assign(var,bert_dict[new_name])
sess.run(op)
need_vars.append(var)
else:
need_vars.append(var)
saver = tf.train.Saver(need_vars)
saver.save(sess, os.path.join('../chinese_L-2_H-768_A-12/', 'bert_model.ckpt'))
2)参数修改,在预训练时修改bert的参数:
{
"attention_probs_dropout_prob": 0.1,
"directionality": "bidi",
"hidden_act": "gelu",
"hidden_dropout_prob": 0.1,
"hidden_size": 768,
"initializer_range": 0.02,
"intermediate_size": 3072,
"max_position_embeddings": 512,
"num_attention_heads": 12,
"num_hidden_layers": 2,
"pooler_fc_size": 768,
"pooler_num_attention_heads": 12,
"pooler_num_fc_layers": 3,
"pooler_size_per_head": 128,
"pooler_type": "first_token_transform",
"type_vocab_size": 2,
"vocab_size": 21128
}