tf2pb: tensorflow 模型 ckpt ,h5 转换为 pb 或服务 pb

- - 编码：utf-8 - -

1. tf ckpt 转 pb , tf h5 转 pb

    支持普通pb和fastertransformer pb转换
    1. fastertransformer pb 可提高1.9x - 3.x加速, fastertransformer 目前只支持官方bert transformer系列
    2. keras h5py模型转换pb

    建议pb模型均可以通过nn-sdk推理
    fastertransformer 4.0
         #cuda 11.3 pip install fastertransformer==4.0.0.113
         #cuda 11.6 pip install fastertransformer==4.0.0.116
    fastertransformer 5.0
         #cuda 11.3 pip install fastertransformer==5.0.0.113
         #cuda 11.6 pip install fastertransformer==5.0.0.116
    推荐 tensorflow 链接如下,建议使用cuda11.3.1 环境tensorflow 1.15
    tensorflow链接: https://pan.baidu.com/s/1PXelYOJ2yqWfWfY7qAL4wA 提取码: rpxv 复制这段内容后打开百度网盘手机App，操作更方便哦
    tf经过测试 ， bert 加速3.x

2.ckpt转换pb

# -*- coding: utf-8 -*-
import os
import tensorflow as tf
import shutil
import tf2pb

#if not fastertransformer , don't advice change

try:
    #cuda 11.3 pip install fastertransformer==4.0.0.113
    #cuda 11.6 pip install fastertransformer==4.0.0.116
    import fastertransformer
    convert_config = {
        "fastertransformer": {
            "floatx": "float32",
            "remove_padding": False,
            "int8_mode": 0,  # need nvidia card supoort,do not suggest
        }
    }
except:
    convert_config = {}


# BertModel_module  加载官方bert模型和fastertransformer模型
#如果是正常pb, 可以直接使用官方modeling 模块 import modeling

def load_model_tensor(bert_config_file,max_seq_len,num_labels):
    BertModel_module = tf2pb.get_modeling(convert_config)
    if BertModel_module is None:
        raise Exception('tf2pb get_modeling failed')
    bert_config = BertModel_module.BertConfig.from_json_file(bert_config_file)

    def create_model(bert_config, is_training, input_ids, input_mask, segment_ids, num_labels, use_one_hot_embeddings):
        """Creates a classification model."""
        model = BertModel_module.BertModel(
            config=bert_config,
            is_training=is_training,
            input_ids=input_ids,
            input_mask=input_mask,
            token_type_ids=segment_ids,
            use_one_hot_embeddings=use_one_hot_embeddings)

        output_layer = model.get_pooled_output()
        hidden_size = output_layer.shape[-1].value
        output_weights = tf.get_variable(
            "output_weights", [num_labels, hidden_size],
            dtype="float32",
            initializer=tf.truncated_normal_initializer(stddev=0.02))
        output_bias = tf.get_variable(
            "output_bias", [num_labels],
            dtype="float32",
            initializer=tf.zeros_initializer())
        logits = tf.matmul(output_layer, output_weights, transpose_b=True)
        logits = tf.nn.bias_add(logits, output_bias)
        probabilities = tf.nn.softmax(logits, axis=-1)
        return probabilities

    input_ids = tf.placeholder(tf.int32, (None, max_seq_len), 'input_ids')
    input_mask = tf.placeholder(tf.int32, (None, max_seq_len), 'input_mask')
    segment_ids = None
    # 这里简单使用分类，具体根据自己需求修改
    probabilities = create_model(bert_config, False, input_ids, input_mask, segment_ids, num_labels, False)
    save_config = {
        "input_tensor": {
            'input_ids': input_ids,
            'input_mask': input_mask
        },
        "output_tensor": {
            "pred_ids": probabilities
        },
    }
    return save_config

if __name__ == '__main__':

    # 训练ckpt权重
    weight_file = r'/home/tk/tk_nlp/script/ner/ner_output/bert/model.ckpt-2704'
    output_dir = r'/home/tk/tk_nlp/script/ner/ner_output/bert'

    bert_config_file = r'/data/nlp/pre_models/tf/bert/chinese_L-12_H-768_A-12/bert_config.json'
    if not os.path.exists(bert_config_file):
        raise Exception("bert_config does not exist")

    max_seq_len = 340
    num_labels = 16 * 4 + 1

    #normal pb
    pb_config = {
        "ckpt_filename": weight_file,  # 训练ckpt权重
        "save_pb_file": os.path.join(output_dir,'bert_inf.pb'),
    }
    #serving pb
    pb_serving_config = {
        'use':False,#默认注释掉保存serving模型
        "ckpt_filename": weight_file,  # 训练ckpt权重
        "save_pb_path_serving": os.path.join(output_dir,'serving'),  # tf_serving 保存模型路径
        'serve_option': {
            'method_name': 'tensorflow/serving/predict',
            'tags': ['serve'],
        }
    }

    if pb_config['save_pb_file'] and os.path.exists(pb_config['save_pb_file']):
        os.remove(pb_config['save_pb_file'])

    if pb_serving_config['use'] and pb_serving_config['save_pb_path_serving'] and os.path.exists(pb_serving_config['save_pb_path_serving']):
        shutil.rmtree(pb_serving_config['save_pb_path_serving'])


    def convert2pb(is_save_serving):
        def create_network_fn():
            save_config = load_model_tensor(bert_config_file=bert_config_file,max_seq_len=max_seq_len,num_labels=num_labels)
            save_config.update(pb_serving_config if is_save_serving else pb_config)
            return save_config

        if not is_save_serving:
            ret = tf2pb.freeze_pb(create_network_fn)
            if ret ==0:
                tf2pb.pb_show(pb_config['save_pb_file'])
            else:
                print('tf2pb.freeze_pb failed ',ret)
        else:
            ret = tf2pb.freeze_pb_serving(create_network_fn)
            if ret ==0:
                tf2pb.pb_serving_show(pb_serving_config['save_pb_path_serving'],pb_serving_config['serve_option']['tags'])  # 查看
            else:
                print('tf2pb.freeze_pb_serving failed ',ret)

    convert2pb(is_save_serving = False)
    if pb_serving_config['use']:
        convert2pb(is_save_serving = True)

3. h5转换pb

import sys
import tensorflow as tf
import tf2pb
import os
from keras.models import Model,load_model

# bert_model is construct by your src code
weight_file = os.path.join(output_dir, 'best_model.h5')
bert_model.load_weights(weight_file , by_name=False)
# or bert_model = load_model(weight_file)

print(bert_model.inputs)
#modify output name
pred_ids = tf.identity(bert_model.output, "pred_ids")



config = {
    'model': bert_model,# the model your trained
    'input_tensor' : {
        "Input-Token": bert_model.inputs[0], # Tensor such as  bert.Input[0]
        "Input-Segment": bert_model.inputs[1], # Tensor such as  bert.Input[0]
    },
    'output_tensor' : {
        "pred_ids": pred_ids, # Tensor output tensor
    },
    'save_pb_file': r'/root/save_pb_file.pb', # pb filename
}

if os.path.exists(config['save_pb_file']):
    os.remove(config['save_pb_file'])
#直接转换
tf2pb.freeze_keras_pb(config)