模型轉換[yolov3模型在keras與darknet之間轉換]

時間  2020-05-05
標籤 模型 轉換 yolov3 yolov keras darknet 之間 简体版
原文   原文鏈接

首先借助qqwweee/keras-yolo3中的convert.py和tensorrt例子yolov3_onnx,並從新編寫了代碼，實現將darknet格式的yolov3的yolov3.cfg和yolov3.weights轉換成keras（tensorflow）的h5格式html

1 將darknet格式的yolov3.cfg和yolov3.weights轉換成kears(tf)的h5格式

# -*- coding: utf-8 -*-

import os
import io
import argparse
import configparser


import numpy as np
from keras import backend as K
from keras.layers import (Conv2D, Input, ZeroPadding2D, Add,
                          UpSampling2D, MaxPooling2D, Concatenate)
from keras.layers.advanced_activations import LeakyReLU
from keras.layers.normalization import BatchNormalization
from keras.models import Model
from keras.regularizers import l2
from keras.utils.vis_utils import plot_model as plot



def parser():
    parser = argparse.ArgumentParser(description="Darknet\'s yolov3.cfg and yolov3.weights \
                                      converted into Keras\'s yolov3.h5!")
    parser.add_argument('-cfg_path', help='yolov3.cfg')
    parser.add_argument('-weights_path', help='yolov3.weights')
    parser.add_argument('-output_path', help='yolov3.h5')
    parser.add_argument('-weights_only', action='store_true',help='only save weights in yolov3.h5')
    return parser.parse_args()


class WeightLoader(object):

    def __init__(self,weight_path):
        self.fhandle = open(weight_path,'rb')
        self.read_bytes = 0

    def parser_buffer(self,shape,dtype='int32',buffer_size=None):
        self.read_bytes += buffer_size
        return np.ndarray(shape=shape,
                          dtype=dtype,
                          buffer=self.fhandle.read(buffer_size) )

    def head(self):

        major, minor, revision = self.parser_buffer(
                                   shape=(3,),
                                   dtype='int32',
                                   buffer_size=12)

        if major*10+minor >= 2 and major < 1000 and minor < 1000:
            seen = self.parser_buffer(
                             shape=(1,),
                             dtype='int64',
                             buffer_size=8)
        else:
            seen = self.parser_buffer(
                             shape=(1,),
                             dtype='int32',
                             buffer_size=4)

        return major, minor, revision, seen

    def close(self):
        self.fhandle.close()

class DarkNetParser(object):
    def __init__(self, cfg_path, weights_path):

        self.block_gen = self._get_block(cfg_path)
        self.weight_loader = WeightLoader(weights_path)
        
        major, minor, revision, seen = self.weight_loader.head()
        print('weights header: ',major, minor, revision, seen)

        self.input_layer = Input(shape=(None, None, 3))
        self.out_index = []
        self.prev_layer = self.input_layer
        self.all_layers = []
        self.count = [0,0]

    
    def _get_block(self,cfg_path):

        block = {}
        with open(cfg_path,'r', encoding='utf-8') as fr:
            for line in fr:
                line = line.strip()
                if '[' in line and ']' in line:        
                    if block:
                        yield block
                    block = {}
                    block['type'] = line.strip(' []')
                elif not line or '#' in line:
                    continue
                else:
                    key,val = line.strip().replace(' ','').split('=')
                    key,val = key.strip(), val.strip()
                    block[key] = val

            yield block


    def conv(self, block):
        '''在讀取darknet的yolov3.weights文件時，順序是
          1 - bias；
          2 - 若是有bn，則接着讀取三個scale，mean，var
          3 - 讀取權重
        '''
        # Darknet serializes convolutional weights as:
        # [bias/beta, [gamma, mean, variance], conv_weights]
        self.count[0] += 1
        # read conv block
        filters = int(block['filters'])
        size = int(block['size'])
        stride = int(block['stride'])
        pad = int(block['pad'])
        activation = block['activation']
        
        padding = 'same' if pad == 1 and stride == 1 else 'valid'
        batch_normalize = 'batch_normalize' in block
        
        prev_layer_shape = K.int_shape(self.prev_layer)
        weights_shape = (size, size, prev_layer_shape[-1], filters)
        darknet_w_shape = (filters, weights_shape[2], size, size)
        weights_size = np.product(weights_shape)

        print('+',self.count[0],'conv2d', 
              'bn' if batch_normalize else ' ',
              activation,
              weights_shape)

        # 讀取濾波器個偏置
        conv_bias = self.weight_loader.parser_buffer(
                                 shape=(filters,),
                                 dtype='float32',
                                 buffer_size=filters*4)
 
        # 若是有bn，則接着讀取濾波器個scale，mean，var
        if batch_normalize:
            bn_weight_list = self.bn(filters, conv_bias)

        # 讀取權重
        conv_weights = self.weight_loader.parser_buffer(
                              shape=darknet_w_shape,
                              dtype='float32',
                              buffer_size=weights_size*4)
        # DarkNet conv_weights are serialized Caffe-style:
        # (out_dim, in_dim, height, width)
        # We would like to set these to Tensorflow order:
        # (height, width, in_dim, out_dim)

        conv_weights = np.transpose(conv_weights, [2, 3, 1, 0])
        conv_weights = [conv_weights] if batch_normalize else \
                              [conv_weights, conv_bias]

        act_fn = None
        if activation == 'leaky':
            pass
        elif activation != 'linear':
            raise

        if stride > 1:
            self.prev_layer = ZeroPadding2D(((1,0),(1,0)))(self.prev_layer)

        conv_layer = (Conv2D(
                filters, (size, size),
                strides=(stride, stride),
                kernel_regularizer=l2(self.weight_decay),
                use_bias=not batch_normalize,
                weights=conv_weights,
                activation=act_fn,
                padding=padding))(self.prev_layer)

        if batch_normalize:
             conv_layer = BatchNormalization(weights=bn_weight_list)(conv_layer)
        self.prev_layer = conv_layer

        if activation == 'linear':
            self.all_layers.append(self.prev_layer)
        elif activation == 'leaky':
            act_layer = LeakyReLU(alpha=0.1)(self.prev_layer)
            self.prev_layer = act_layer
            self.all_layers.append(act_layer)


    def bn(self,filters,conv_bias):
        '''bn有4個參數，分別是bias，scale，mean，var，
          其中bias已經讀取完畢，這裏讀取剩下三個，scale,mean,var '''
        bn_weights = self.weight_loader.parser_buffer(
                              shape=(3,filters),
                              dtype='float32',
                              buffer_size=(filters*3)*4)
        # scale, bias, mean,var
        bn_weight_list = [bn_weights[0],
                          conv_bias,
                          bn_weights[1],
                          bn_weights[2] ]
        return bn_weight_list
       
    def maxpool(self,block):
        size = int(block['size'])
        stride = int(block['stride'])
        maxpool_layer = MaxPooling2D(pool_size=(size,size),
                        strides=(stride,stride),
                        padding='same')(self.prev_layer)
        self.all_layers.append(maxpool_layer)
        self.prev_layer = maxpool_layer

    def shortcut(self,block):
        index = int(block['from'])
        activation = block['activation']
        assert activation == 'linear', 'Only linear activation supported.'
        shortcut_layer = Add()([self.all_layers[index],self.prev_layer])
        self.all_layers.append(shortcut_layer)
        self.prev_layer = shortcut_layer

    def route(self,block):
        layers_ids = block['layers']
        ids = [int(i) for i in layers_ids.split(',')]
        layers = [self.all_layers[i] for i in ids]
        if len(layers) > 1:
            print('Concatenating route layers:', layers)
            concatenate_layer = Concatenate()(layers)
            self.all_layers.append(concatenate_layer)
            self.prev_layer = concatenate_layer
        else:
            skip_layer = layers[0]
            self.all_layers.append(skip_layer)
            self.prev_layer = skip_layer

    def upsample(self,block):
        stride = int(block['stride'])
        assert stride == 2, 'Only stride=2 supported.'
        upsample_layer = UpSampling2D(stride)(self.prev_layer)
        self.all_layers.append(upsample_layer)
        self.prev_layer = self.all_layers[-1]
    
    def yolo(self,block):
        self.out_index.append(len(self.all_layers)-1)
        self.all_layers.append(None)
        self.prev_layer = self.all_layers[-1]

    def net(self, block):
        self.weight_decay = block['decay']


    def create_and_save(self,weights_only,output_path):
        if len(self.out_index) == 0:
            self.out_index.append( len(self.all_layers)-1 )

        output_layers = [self.all_layers[i] for i in self.out_index]
        model = Model(inputs=self.input_layer,
                      outputs=output_layers)
        print(model.summary())

        if weights_only:
            model.save_weights(output_path)
            print('Saved Keras weights to {}'.format(output_path))
        else:
            model.save(output_path)
            print('Saved Keras model to {}'.format(output_path))

    def close(self):
        self.weight_loader.close()
        

def main():

    args = parser()
    print('loading weights...')

    cfg_parser = DarkNetParser(args.cfg_path,args.weights_path)

    print('creating keras model...')

    layers_fun = {'convolutional':cfg_parser.conv,
                 'net':cfg_parser.net,
                 'yolo':cfg_parser.yolo,
                 'route':cfg_parser.route,
                 'upsample':cfg_parser.upsample,
                 'maxpool':cfg_parser.maxpool,
                 'shortcut':cfg_parser.shortcut
                 }

    print('Parsing Darknet config.')
    for ind,block in enumerate(cfg_parser.block_gen):
        type = block['type']
        layers_fun[type](block)

    cfg_parser.create_and_save(args.weights_only, args.output_path)
    cfg_parser.close()
        

if __name__ == '__main__':
    main()

運行結果python

python yolov3_darknet_to_keras.py -cfg_path text.cfg -weights_path yolov3.weights -output_path yolov3c_d2k.h5

<center/>![](https://img2018.cnblogs.com/blog/441382/201903/441382-20190321153317547-1618024530.png) ![](https://img2018.cnblogs.com/blog/441382/201903/441382-20190321153425958-831228099.png)</center> ### 2 將kears(tf)的h5格式轉換成darknet格式的yolov3.weights 其中上面的與下面的名稱轉換 ``` bias -> beta scale -> gamma mean -> moving_mean var -> moving_variance ``` 基於此寫的腳本爲： ```python # -*- coding: utf-8 -*- ''' yolov3_keras_to_darknet.py''' import argparse import numpy import numpy as np import keras from keras.models import load_model from keras import backend as Kgit

def parser(): parser = argparse.ArgumentParser(description="Darknet's yolov3.cfg and yolov3.weights
converted into Keras's yolov3.h5!") parser.add_argument('-cfg_path', help='yolov3.cfg') parser.add_argument('-h5_path', help='yolov3.h5') parser.add_argument('-output_path', help='yolov3.weights') return parser.parse_args()github

class WeightSaver(object):網絡

def __init__(self,h5_path,output_path):
    self.model = load_model(h5_path)
   # 若是要讀取keras調用save_weights的h5文件，能夠先讀取一次save的h5，
   # 而後取消下面的註釋，讀取save_weights的h5

self.model.load_weights('text.h5')

self.layers = {weight.name:weight for weight in self.model.weights}
    self.sess = K.get_session()
    self.fhandle = open(output_path,'wb')
    self._write_head()

def _write_head(self):
    numpy_data = numpy.ndarray(shape=(3,),
                      dtype='int32',
                      buffer=np.array([0,2,0],dtype='int32') )
    self.save(numpy_data)
    numpy_data = numpy.ndarray(shape=(1,),
                      dtype='int64',
                      buffer=np.array([320000],dtype='int64'))
    self.save(numpy_data)

def get_bn_layername(self,num):
    layer_name = 'batch_normalization_{num}'.format(num=num)
    bias = self.layers['{0}/beta:0'.format(layer_name)]
    scale = self.layers['{0}/gamma:0'.format(layer_name)]
    mean = self.layers['{0}/moving_mean:0'.format(layer_name)]
    var = self.layers['{0}/moving_variance:0'.format(layer_name)]
   
    bias_np = self.get_numpy(bias)
    scale_np = self.get_numpy(scale)
    mean_np = self.get_numpy(mean)
    var_np = self.get_numpy(var)
    return bias_np,scale_np,mean_np,var_np

def get_convbias_layername(self,num):
    layer_name = 'conv2d_{num}'.format(num=num)
    bias = self.layers['{0}/bias:0'.format(layer_name)]
  
    bias_np = self.get_numpy(bias)
    return bias_np

def get_conv_layername(self,num):
    layer_name = 'conv2d_{num}'.format(num=num)
    conv = self.layers['{0}/kernel:0'.format(layer_name)]
   
    conv_np = self.get_numpy(conv)
    return conv_np


def get_numpy(self,layer_name):
    numpy_data = self.sess.run(layer_name)
    return numpy_data

def save(self,numpy_data):
    bytes_data = numpy_data.tobytes()
    self.fhandle.write(bytes_data)
    self.fhandle.flush()

def close(self):
    self.fhandle.close()

class KerasParser(object):session

def __init__(self, cfg_path, h5_path, output_path):
    self.block_gen = self._get_block(cfg_path)
    self.weights_saver = WeightSaver(h5_path, output_path)
    self.count_conv = 0
    self.count_bn = 0

def _get_block(self,cfg_path):

    block = {}
    with open(cfg_path,'r', encoding='utf-8') as fr:
        for line in fr:
            line = line.strip()
            if '[' in line and ']' in line:
                if block:
                    yield block
                block = {}
                block['type'] = line.strip(' []')
            elif not line or '#' in line:
                continue
            else:
                key,val = line.strip().replace(' ','').split('=')
                key,val = key.strip(), val.strip()
                block[key] = val

        yield block

def close(self):
    self.weights_saver.close()

def conv(self, block):
    self.count_conv += 1
    batch_normalize = 'batch_normalize' in block
    print('handing.. ',self.count_conv)

    # 若是bn存在，則先處理bn，順序爲bias，scale，mean，var
    if batch_normalize:
        bias,scale,mean,var = self.bn()
        self.weights_saver.save(bias)
        
        scale = scale.reshape(1,-1)
        mean = mean.reshape(1,-1)
        var = var.reshape(1,-1)
        remain = np.concatenate([scale,mean,var],axis=0)
        self.weights_saver.save(remain)

    # 不然，先處理biase
    else:
        conv_bias = self.weights_saver.get_convbias_layername(self.count_conv)
        self.weights_saver.save(conv_bias)

    # 接着處理weights
    conv_weights = self.weights_saver.get_conv_layername(self.count_conv)
    # 須要將(height, width, in_dim, out_dim)轉換成(out_dim, in_dim, height, width)
    conv_weights = np.transpose(conv_weights,[3,2,0,1])
    self.weights_saver.save(conv_weights)

def bn(self):
    self.count_bn += 1
    bias,scale,mean,var = self.weights_saver.get_bn_layername(self.count_bn) 
    return bias,scale,mean,var

def main(): args = parser() keras_loader = KerasParser(args.cfg_path, args.h5_path, args.output_path)app

for block in keras_loader.block_gen:
    if 'convolutional' in block['type']:
        keras_loader.conv(block)
keras_loader.close()

if name == "main": main()ide

經過讀取keras保存的h5文件，並讀取其權重，其以下所示,

[<tf.Variable 'conv2d_1/kernel:0' shape=(3, 3, 3, 32) dtype=float32_ref>, <tf.Variable 'batch_normalization_1/gamma:0' shape=(32,) dtype=float32_ref>, <tf.Variable 'batch_normalization_1/beta:0' shape=(32,) dtype=float32_ref>, <tf.Variable 'batch_normalization_1/moving_mean:0' shape=(32,) dtype=float32_ref>, <tf.Variable 'batch_normalization_1/moving_variance:0' shape=(32,) dtype=float32_ref>,spa

<tf.Variable 'conv2d_2/kernel:0' shape=(3, 3, 32, 64) dtype=float32_ref>, <tf.Variable 'batch_normalization_2/gamma:0' shape=(64,) dtype=float32_ref>, <tf.Variable 'batch_normalization_2/beta:0' shape=(64,) dtype=float32_ref>, <tf.Variable 'batch_normalization_2/moving_mean:0' shape=(64,) dtype=float32_ref>, <tf.Variable 'batch_normalization_2/moving_variance:0' shape=(64,) dtype=float32_ref>,.net

<tf.Variable 'conv2d_3/kernel:0' shape=(1, 1, 64, 32) dtype=float32_ref>, <tf.Variable 'batch_normalization_3/gamma:0' shape=(32,) dtype=float32_ref>, <tf.Variable 'batch_normalization_3/beta:0' shape=(32,) dtype=float32_ref>, <tf.Variable 'batch_normalization_3/moving_mean:0' shape=(32,) dtype=float32_ref>, <tf.Variable 'batch_normalization_3/moving_variance:0' shape=(32,) dtype=float32_ref>,

<tf.Variable 'conv2d_4/kernel:0' shape=(3, 3, 32, 64) dtype=float32_ref>, <tf.Variable 'batch_normalization_4/gamma:0' shape=(64,) dtype=float32_ref>, <tf.Variable 'batch_normalization_4/beta:0' shape=(64,) dtype=float32_ref>, <tf.Variable 'batch_normalization_4/moving_mean:0' shape=(64,) dtype=float32_ref>, <tf.Variable 'batch_normalization_4/moving_variance:0' shape=(64,) dtype=float32_ref>,

<tf.Variable 'conv2d_5/kernel:0' shape=(3, 3, 64, 128) dtype=float32_ref>, <tf.Variable 'batch_normalization_5/gamma:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_5/beta:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_5/moving_mean:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_5/moving_variance:0' shape=(128,) dtype=float32_ref>,

<tf.Variable 'conv2d_6/kernel:0' shape=(1, 1, 128, 64) dtype=float32_ref>, <tf.Variable 'batch_normalization_6/gamma:0' shape=(64,) dtype=float32_ref>, <tf.Variable 'batch_normalization_6/beta:0' shape=(64,) dtype=float32_ref>, <tf.Variable 'batch_normalization_6/moving_mean:0' shape=(64,) dtype=float32_ref>, <tf.Variable 'batch_normalization_6/moving_variance:0' shape=(64,) dtype=float32_ref>,

<tf.Variable 'conv2d_7/kernel:0' shape=(3, 3, 64, 128) dtype=float32_ref>, <tf.Variable 'batch_normalization_7/gamma:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_7/beta:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_7/moving_mean:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_7/moving_variance:0' shape=(128,) dtype=float32_ref>,

<tf.Variable 'conv2d_8/kernel:0' shape=(1, 1, 128, 64) dtype=float32_ref>, <tf.Variable 'batch_normalization_8/gamma:0' shape=(64,) dtype=float32_ref>, <tf.Variable 'batch_normalization_8/beta:0' shape=(64,) dtype=float32_ref>, <tf.Variable 'batch_normalization_8/moving_mean:0' shape=(64,) dtype=float32_ref>, <tf.Variable 'batch_normalization_8/moving_variance:0' shape=(64,) dtype=float32_ref>,

<tf.Variable 'conv2d_9/kernel:0' shape=(3, 3, 64, 128) dtype=float32_ref>, <tf.Variable 'batch_normalization_9/gamma:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_9/beta:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_9/moving_mean:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_9/moving_variance:0' shape=(128,) dtype=float32_ref>,

<tf.Variable 'conv2d_10/kernel:0' shape=(3, 3, 128, 256) dtype=float32_ref>, <tf.Variable 'batch_normalization_10/gamma:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_10/beta:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_10/moving_mean:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_10/moving_variance:0' shape=(256,) dtype=float32_ref>,

<tf.Variable 'conv2d_11/kernel:0' shape=(1, 1, 256, 128) dtype=float32_ref>, <tf.Variable 'batch_normalization_11/gamma:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_11/beta:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_11/moving_mean:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_11/moving_variance:0' shape=(128,) dtype=float32_ref>,

<tf.Variable 'conv2d_12/kernel:0' shape=(3, 3, 128, 256) dtype=float32_ref>, <tf.Variable 'batch_normalization_12/gamma:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_12/beta:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_12/moving_mean:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_12/moving_variance:0' shape=(256,) dtype=float32_ref>,

<tf.Variable 'conv2d_13/kernel:0' shape=(1, 1, 256, 128) dtype=float32_ref>, <tf.Variable 'batch_normalization_13/gamma:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_13/beta:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_13/moving_mean:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_13/moving_variance:0' shape=(128,) dtype=float32_ref>,

<tf.Variable 'conv2d_14/kernel:0' shape=(3, 3, 128, 256) dtype=float32_ref>, <tf.Variable 'batch_normalization_14/gamma:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_14/beta:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_14/moving_mean:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_14/moving_variance:0' shape=(256,) dtype=float32_ref>,

<tf.Variable 'conv2d_15/kernel:0' shape=(1, 1, 256, 128) dtype=float32_ref>, <tf.Variable 'batch_normalization_15/gamma:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_15/beta:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_15/moving_mean:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_15/moving_variance:0' shape=(128,) dtype=float32_ref>,

<tf.Variable 'conv2d_16/kernel:0' shape=(3, 3, 128, 256) dtype=float32_ref>, <tf.Variable 'batch_normalization_16/gamma:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_16/beta:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_16/moving_mean:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_16/moving_variance:0' shape=(256,) dtype=float32_ref>,

<tf.Variable 'conv2d_17/kernel:0' shape=(1, 1, 256, 128) dtype=float32_ref>, <tf.Variable 'batch_normalization_17/gamma:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_17/beta:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_17/moving_mean:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_17/moving_variance:0' shape=(128,) dtype=float32_ref>,

<tf.Variable 'conv2d_18/kernel:0' shape=(3, 3, 128, 256) dtype=float32_ref>, <tf.Variable 'batch_normalization_18/gamma:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_18/beta:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_18/moving_mean:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_18/moving_variance:0' shape=(256,) dtype=float32_ref>,

<tf.Variable 'conv2d_19/kernel:0' shape=(1, 1, 256, 128) dtype=float32_ref>, <tf.Variable 'batch_normalization_19/gamma:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_19/beta:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_19/moving_mean:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_19/moving_variance:0' shape=(128,) dtype=float32_ref>,

<tf.Variable 'conv2d_20/kernel:0' shape=(3, 3, 128, 256) dtype=float32_ref>, <tf.Variable 'batch_normalization_20/gamma:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_20/beta:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_20/moving_mean:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_20/moving_variance:0' shape=(256,) dtype=float32_ref>,

<tf.Variable 'conv2d_21/kernel:0' shape=(1, 1, 256, 128) dtype=float32_ref>, <tf.Variable 'batch_normalization_21/gamma:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_21/beta:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_21/moving_mean:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_21/moving_variance:0' shape=(128,) dtype=float32_ref>,

<tf.Variable 'conv2d_22/kernel:0' shape=(3, 3, 128, 256) dtype=float32_ref>, <tf.Variable 'batch_normalization_22/gamma:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_22/beta:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_22/moving_mean:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_22/moving_variance:0' shape=(256,) dtype=float32_ref>,

<tf.Variable 'conv2d_23/kernel:0' shape=(1, 1, 256, 128) dtype=float32_ref>, <tf.Variable 'batch_normalization_23/gamma:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_23/beta:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_23/moving_mean:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_23/moving_variance:0' shape=(128,) dtype=float32_ref>,

<tf.Variable 'conv2d_24/kernel:0' shape=(3, 3, 128, 256) dtype=float32_ref>, <tf.Variable 'batch_normalization_24/gamma:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_24/beta:0' shape=(256,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_25/kernel:0' shape=(1, 1, 256, 128) dtype=float32_ref>, <tf.Variable 'batch_normalization_25/gamma:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_25/beta:0' shape=(128,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_26/kernel:0' shape=(3, 3, 128, 256) dtype=float32_ref>, <tf.Variable 'batch_normalization_26/gamma:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_26/beta:0' shape=(256,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_27/kernel:0' shape=(3, 3, 256, 512) dtype=float32_ref>, <tf.Variable 'batch_normalization_27/gamma:0' shape=(512,) dtype=float32_ref>, <tf.Variable 'batch_normalization_27/beta:0' shape=(512,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_28/kernel:0' shape=(1, 1, 512, 256) dtype=float32_ref>, <tf.Variable 'batch_normalization_28/gamma:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_28/beta:0' shape=(256,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_29/kernel:0' shape=(3, 3, 256, 512) dtype=float32_ref>, <tf.Variable 'batch_normalization_29/gamma:0' shape=(512,) dtype=float32_ref>, <tf.Variable 'batch_normalization_29/beta:0' shape=(512,) dtype=float32_ref> ......

<tf.Variable 'conv2d_30/kernel:0' shape=(1, 1, 512, 256) dtype=float32_ref>, <tf.Variable 'batch_normalization_30/gamma:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_30/beta:0' shape=(256,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_31/kernel:0' shape=(3, 3, 256, 512) dtype=float32_ref>, <tf.Variable 'batch_normalization_31/gamma:0' shape=(512,) dtype=float32_ref>, <tf.Variable 'batch_normalization_31/beta:0' shape=(512,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_32/kernel:0' shape=(1, 1, 512, 256) dtype=float32_ref>, <tf.Variable 'batch_normalization_32/gamma:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_32/beta:0' shape=(256,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_33/kernel:0' shape=(3, 3, 256, 512) dtype=float32_ref>, <tf.Variable 'batch_normalization_33/gamma:0' shape=(512,) dtype=float32_ref>, <tf.Variable 'batch_normalization_33/beta:0' shape=(512,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_34/kernel:0' shape=(1, 1, 512, 256) dtype=float32_ref>, <tf.Variable 'batch_normalization_34/gamma:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_34/beta:0' shape=(256,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_35/kernel:0' shape=(3, 3, 256, 512) dtype=float32_ref>, <tf.Variable 'batch_normalization_35/gamma:0' shape=(512,) dtype=float32_ref>, <tf.Variable 'batch_normalization_35/beta:0' shape=(512,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_36/kernel:0' shape=(1, 1, 512, 256) dtype=float32_ref>, <tf.Variable 'batch_normalization_36/gamma:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_36/beta:0' shape=(256,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_37/kernel:0' shape=(3, 3, 256, 512) dtype=float32_ref>, <tf.Variable 'batch_normalization_37/gamma:0' shape=(512,) dtype=float32_ref>, <tf.Variable 'batch_normalization_37/beta:0' shape=(512,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_38/kernel:0' shape=(1, 1, 512, 256) dtype=float32_ref>, <tf.Variable 'batch_normalization_38/gamma:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_38/beta:0' shape=(256,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_39/kernel:0' shape=(3, 3, 256, 512) dtype=float32_ref>, <tf.Variable 'batch_normalization_39/gamma:0' shape=(512,) dtype=float32_ref>, <tf.Variable 'batch_normalization_39/beta:0' shape=(512,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_40/kernel:0' shape=(1, 1, 512, 256) dtype=float32_ref>, <tf.Variable 'batch_normalization_40/gamma:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_40/beta:0' shape=(256,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_41/kernel:0' shape=(3, 3, 256, 512) dtype=float32_ref>, <tf.Variable 'batch_normalization_41/gamma:0' shape=(512,) dtype=float32_ref>, <tf.Variable 'batch_normalization_41/beta:0' shape=(512,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_42/kernel:0' shape=(1, 1, 512, 256) dtype=float32_ref>, <tf.Variable 'batch_normalization_42/gamma:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_42/beta:0' shape=(256,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_43/kernel:0' shape=(3, 3, 256, 512) dtype=float32_ref>, <tf.Variable 'batch_normalization_43/gamma:0' shape=(512,) dtype=float32_ref>, <tf.Variable 'batch_normalization_43/beta:0' shape=(512,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_44/kernel:0' shape=(3, 3, 512, 1024) dtype=float32_ref>, <tf.Variable 'batch_normalization_44/gamma:0' shape=(1024,) dtype=float32_ref>, <tf.Variable 'batch_normalization_44/beta:0' shape=(1024,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_45/kernel:0' shape=(1, 1, 1024, 512) dtype=float32_ref>, <tf.Variable 'batch_normalization_45/gamma:0' shape=(512,) dtype=float32_ref>, <tf.Variable 'batch_normalization_45/beta:0' shape=(512,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_46/kernel:0' shape=(3, 3, 512, 1024) dtype=float32_ref>, <tf.Variable 'batch_normalization_46/gamma:0' shape=(1024,) dtype=float32_ref>, <tf.Variable 'batch_normalization_46/beta:0' shape=(1024,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_47/kernel:0' shape=(1, 1, 1024, 512) dtype=float32_ref>, <tf.Variable 'batch_normalization_47/gamma:0' shape=(512,) dtype=float32_ref>, <tf.Variable 'batch_normalization_47/beta:0' shape=(512,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_48/kernel:0' shape=(3, 3, 512, 1024) dtype=float32_ref>, <tf.Variable 'batch_normalization_48/gamma:0' shape=(1024,) dtype=float32_ref>, <tf.Variable 'batch_normalization_48/beta:0' shape=(1024,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_49/kernel:0' shape=(1, 1, 1024, 512) dtype=float32_ref>, <tf.Variable 'batch_normalization_49/gamma:0' shape=(512,) dtype=float32_ref>, <tf.Variable 'batch_normalization_49/beta:0' shape=(512,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_50/kernel:0' shape=(3, 3, 512, 1024) dtype=float32_ref>, <tf.Variable 'batch_normalization_50/gamma:0' shape=(1024,) dtype=float32_ref>, <tf.Variable 'batch_normalization_50/beta:0' shape=(1024,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_51/kernel:0' shape=(1, 1, 1024, 512) dtype=float32_ref>, <tf.Variable 'batch_normalization_51/gamma:0' shape=(512,) dtype=float32_ref>, <tf.Variable 'batch_normalization_51/beta:0' shape=(512,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_52/kernel:0' shape=(3, 3, 512, 1024) dtype=float32_ref>, <tf.Variable 'batch_normalization_52/gamma:0' shape=(1024,) dtype=float32_ref>, <tf.Variable 'batch_normalization_52/beta:0' shape=(1024,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_53/kernel:0' shape=(1, 1, 1024, 512) dtype=float32_ref>, <tf.Variable 'batch_normalization_53/gamma:0' shape=(512,) dtype=float32_ref>, <tf.Variable 'batch_normalization_53/beta:0' shape=(512,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_54/kernel:0' shape=(3, 3, 512, 1024) dtype=float32_ref>, <tf.Variable 'batch_normalization_54/gamma:0' shape=(1024,) dtype=float32_ref>, <tf.Variable 'batch_normalization_54/beta:0' shape=(1024,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_55/kernel:0' shape=(1, 1, 1024, 512) dtype=float32_ref>, <tf.Variable 'batch_normalization_55/gamma:0' shape=(512,) dtype=float32_ref>, <tf.Variable 'batch_normalization_55/beta:0' shape=(512,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_56/kernel:0' shape=(3, 3, 512, 1024) dtype=float32_ref>, <tf.Variable 'batch_normalization_56/gamma:0' shape=(1024,) dtype=float32_ref>, <tf.Variable 'batch_normalization_56/beta:0' shape=(1024,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_57/kernel:0' shape=(1, 1, 1024, 512) dtype=float32_ref>, <tf.Variable 'batch_normalization_57/gamma:0' shape=(512,) dtype=float32_ref>, <tf.Variable 'batch_normalization_57/beta:0' shape=(512,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_58/kernel:0' shape=(3, 3, 512, 1024) dtype=float32_ref>, <tf.Variable 'batch_normalization_58/gamma:0' shape=(1024,) dtype=float32_ref>, <tf.Variable 'batch_normalization_58/beta:0' shape=(1024,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_59/kernel:0' shape=(1, 1, 1024, 21) dtype=float32_ref>, <tf.Variable 'conv2d_59/bias:0' shape=(21,) dtype=float32_ref>,

<tf.Variable 'conv2d_60/kernel:0' shape=(1, 1, 512, 256) dtype=float32_ref>, <tf.Variable 'batch_normalization_59/gamma:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_59/beta:0' shape=(256,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_61/kernel:0' shape=(1, 1, 768, 256) dtype=float32_ref>, <tf.Variable 'batch_normalization_60/gamma:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_60/beta:0' shape=(256,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_62/kernel:0' shape=(3, 3, 256, 512) dtype=float32_ref>, <tf.Variable 'batch_normalization_61/gamma:0' shape=(512,) dtype=float32_ref>, <tf.Variable 'batch_normalization_61/beta:0' shape=(512,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_63/kernel:0' shape=(1, 1, 512, 256) dtype=float32_ref>, <tf.Variable 'batch_normalization_62/gamma:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_62/beta:0' shape=(256,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_64/kernel:0' shape=(3, 3, 256, 512) dtype=float32_ref>, <tf.Variable 'batch_normalization_63/gamma:0' shape=(512,) dtype=float32_ref>, <tf.Variable 'batch_normalization_63/beta:0' shape=(512,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_65/kernel:0' shape=(1, 1, 512, 256) dtype=float32_ref>, <tf.Variable 'batch_normalization_64/gamma:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_64/beta:0' shape=(256,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_66/kernel:0' shape=(3, 3, 256, 512) dtype=float32_ref>, <tf.Variable 'batch_normalization_65/gamma:0' shape=(512,) dtype=float32_ref>, <tf.Variable 'batch_normalization_65/beta:0' shape=(512,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_67/kernel:0' shape=(1, 1, 512, 21) dtype=float32_ref>, <tf.Variable 'conv2d_67/bias:0' shape=(21,) dtype=float32_ref>,

<tf.Variable 'conv2d_68/kernel:0' shape=(1, 1, 256, 128) dtype=float32_ref>, <tf.Variable 'batch_normalization_66/gamma:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_66/beta:0' shape=(128,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_69/kernel:0' shape=(1, 1, 384, 128) dtype=float32_ref>, <tf.Variable 'batch_normalization_67/gamma:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_67/beta:0' shape=(128,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_70/kernel:0' shape=(3, 3, 128, 256) dtype=float32_ref>, <tf.Variable 'batch_normalization_68/gamma:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_68/beta:0' shape=(256,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_71/kernel:0' shape=(1, 1, 256, 128) dtype=float32_ref>, <tf.Variable 'batch_normalization_69/gamma:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_69/beta:0' shape=(128,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_72/kernel:0' shape=(3, 3, 128, 256) dtype=float32_ref>, <tf.Variable 'batch_normalization_70/gamma:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_70/beta:0' shape=(256,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_73/kernel:0' shape=(1, 1, 256, 128) dtype=float32_ref>, <tf.Variable 'batch_normalization_71/gamma:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_71/beta:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_71/moving_mean:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'batch_normalization_71/moving_variance:0' shape=(128,) dtype=float32_ref>,

<tf.Variable 'conv2d_74/kernel:0' shape=(3, 3, 128, 256) dtype=float32_ref>, <tf.Variable 'batch_normalization_72/gamma:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'batch_normalization_72/beta:0' shape=(256,) dtype=float32_ref>, ......

<tf.Variable 'conv2d_75/kernel:0' shape=(1, 1, 256, 21) dtype=float32_ref>, <tf.Variable 'conv2d_75/bias:0' shape=(21,) dtype=float32_ref>,

運行結果

python yolov3_keras_to_darknet.py -cfg_path text.cfg -h5_path yolov3c_d2k.h5 -output_path yolov3c_d2k_k2d.weights

<center/>![](https://img2018.cnblogs.com/blog/441382/201903/441382-20190321152939710-1147459957.png)</center>
能夠看出原始文件yolov3.weights轉成yolov3c_d2k.h5，而後再轉回來yolov3c_d2k_k2d.weights，而md5值未變，說明逆向轉換成功。

### 3 實際案例

這裏咱們給出完整的操做過程：
首先，機器環境(我的以爲這裏應該無所謂)：
1 - python3.5.6；
2 - keras 2.2.4；
3 - tensorflow-gpu 1.12.0.

而後，去github的[chineseocr](https://github.com/chineseocr/chineseocr)給出的[百度網盤](https://pan.baidu.com/s/1gTW9gwJR6hlwTuyB6nCkzQ#list/path=%2Fchineseocr)下下載:
> * text.h5: 經過keras的save_weights方式保存的
> * text.weights：darknet生成的文件
> * text.cfg： darknet中yolov3的網絡結構

若是直接執行

python yolov3_keras_to_darknet.py -cfg_path text.cfg -h5_path text.h5 -output_path test.weights

會報：
<center/>![](https://img2018.cnblogs.com/blog/441382/201904/441382-20190426130013495-2030975770.png)</center>
是由於當前h5是經過save_weights方式保存的，而非save方式。

因此咱們先執行

python yolov3_darknet_to_keras.py -cfg_path text.cfg -weights_path text.weights -output_path test.h5

此時是將darknet的結構經過keras的save方式轉換成h5.而後此時執行,就沒問題了：

python h5_to_weights.py -cfg_path text.cfg -h5_path test.h5 -output_path test.weights

相關文章
相關標籤/搜索
每日一句
    每一个你不满意的现在，都有一个你没有努力的曾经。
本站公眾號
   歡迎關注本站公眾號,獲取更多信息
聯繫我們 最近搜索 最新文章 沪ICP备13005482号-10 MyBatis教程 SQL 教程 MySQL教程 Java 教程 Thymeleaf 教程 Hibernate教程 Spring教程 Redis教程