tensorflow的模型保存文件.ckpt包含什么以及如何读取、并将其转换为Keras的.h5文件

【时间】2019.12.03

【题目】tensorflow的模型保存文件.ckpt包含什么以及如何读取、并将其转换为Keras的.h5文件

参考：tensorflow中读取模型中保存的值, tf.train.NewCheckpointReader

一、tensorflow的模型保存文件.ckpt

.ckpt即checkpoint文件。

一个标准的.ckpt模型文件以下文件：

model.ckpt文件组成：
model.ckpt.meta
model.ckpt.data-00000-of-00001
model.ckpt.index

二、读取.ckpt文件中的内容

使用 NewCheck =tf.train.NewCheckpointReader(ckpt_file)读取：

import tensorflow as tf
import pprint # 使用pprint 提高打印的可读性
NewCheck =tf.train.NewCheckpointReader("model.ckpt")

2.1.打印模型中的所有变量

print("debug_string:\n")
pprint.pprint(NewCheck.debug_string().decode("utf-8")) #类型是str

其中有3个字段, 分别是名字name, 数据类型type, 尺寸shape

2.2 获取变量中的值--get_tensor

print("get_tensor:\n")
pprint.pprint(NewCheck.get_tensor("D/conv2d/bias"))

2.3 、

2.2相当于：NewCheck.get_variable_to_dtype_map()+get_variable_to_shape_map()

不过得到的是字典类型的数据。

print("get_variable_to_dtype_map\n")
pprint.pprint(NewCheck.get_variable_to_dtype_map())
print("get_variable_to_shape_map\n")
pprint.pprint(NewCheck.get_variable_to_shape_map())

三、读取ckpt文件的权重并转换为keras的.h5文件

代码：

代码来自：https://github.com/johannesu/NASNet-keras

def load_weights_from_tf_checkpoint(model, checkpoint_file, background_label):
    print('Load weights from tensorflow checkpoint')
    progbar = Progbar(target=len(model.layers))

    reader = tf.train.NewCheckpointReader(checkpoint_file)
    for index, layer in enumerate(model.layers):
        progbar.update(current=index)

        if isinstance(layer, layers.convolutional.SeparableConv2D):
            depthwise = reader.get_tensor('{}/depthwise_weights'.format(layer.name))
            pointwise = reader.get_tensor('{}/pointwise_weights'.format(layer.name))

            if K.image_data_format() == 'channels_first':
                depthwise = convert_kernel(depthwise)
                pointwise = convert_kernel(pointwise)

            layer.set_weights([depthwise, pointwise])
        elif isinstance(layer, layers.convolutional.Convolution2D):
            weights = reader.get_tensor('{}/weights'.format(layer.name))

            if K.image_data_format() == 'channels_first':
                weights = convert_kernel(weights)

            layer.set_weights([weights])
        elif isinstance(layer, layers.BatchNormalization):
            beta = reader.get_tensor('{}/beta'.format(layer.name))
            gamma = reader.get_tensor('{}/gamma'.format(layer.name))
            moving_mean = reader.get_tensor('{}/moving_mean'.format(layer.name))
            moving_variance = reader.get_tensor('{}/moving_variance'.format(layer.name))

            layer.set_weights([gamma, beta, moving_mean, moving_variance])
        elif isinstance(layer, layers.Dense):
            weights = reader.get_tensor('{}/weights'.format(layer.name))
            biases = reader.get_tensor('{}/biases'.format(layer.name))

            if background_label:
                layer.set_weights([weights, biases])
            else:
                layer.set_weights([weights[:, 1:], biases[1:]])


def load_pretrained_weights(model, fname, origin, md5_hash, background_label=False, cache_dir=None):
    """Download and convert tensorflow checkpoints"""
    # origin = 'https://storage.googleapis.com/download.tensorflow.org/models/nasnet-a_large_04_10_2017.tar.gz'

    if cache_dir is None:
        cache_dir = os.path.expanduser(os.path.join('~', '.keras', 'models'))

    weight_path = os.path.join(cache_dir, '{}_{}_{}.h5'.format(model.name, md5_hash, K.image_data_format()))

    if os.path.exists(weight_path):
        model.load_weights(weight_path)
    else:
        if not os.path.exists(cache_dir):
            os.makedirs(cache_dir)
            print(cache_dir)
        path = get_file(fname, origin=origin, extract=True, md5_hash=md5_hash, cache_dir=cache_dir,cache_subdir='.')
        checkpoint_file = os.path.join(path, '..', 'model.ckpt')
        load_weights_from_tf_checkpoint(model, checkpoint_file, background_label)
        model.save_weights(weight_path)

分析：

1、用Keras代码构建model,并且每一层layer的name与.ckpt文件中的层名字相同。

2、实例化reader=tf.train.NewCheckpointReader()

3、对于model的每一层，根据layer的类别，reader.get_file(layer_name)获取权值（array数组），并使用layer.set_weights()进行赋值。

4、最后使用model.save_weights()保存.h5权重。

有没有更简单的方法？？