keras读取h5文件load_weights、load代码操作

关于保存h5模型、权重网上的示例非常多，也非常简单。主要有以下两个函数：

1、keras.models.load_model() 读取网络、权重

2、keras.models.load_weights() 仅读取权重

load_model代码包含load_weights的代码，区别在于load_weights时需要先有网络、并且load_weights需要将权重数据写入到对应网络层的tensor中。

下面以resnet50加载h5权重为例，示例代码如下

import keras
from keras.preprocessing import image
import numpy as np

from network.resnet50 import ResNet50
#修改过，不加载权重（默认官方加载亦可）
model = ResNet50() 

# 参数默认 by_name = Fasle， 否则只读取匹配的权重
# 这里h5的层和权重文件中层名是对应的（除input层）
model.load_weights(r'\models\resnet50_weights_tf_dim_ordering_tf_kernels_v2.h5') 

模型通过 model.summary()输出

resnet50_v2 这个权重文件，仅一个attr “layer_names”, 该attr包含177个string的Array，Array中每个元素就是层的名字（这里是严格对应在keras进行保存权重时网络中每一层的name值，且层的顺序也严格对应）。

对于每一个key（层名），都有一个属性"weights_names"，（value值可能为空)。

例如：

conv1的"weights_names"有"conv1_W:0"和"conv1_b:0"，

flatten_1的"weights_names"为null。

filtered_layers为当前模型resnet50过滤（input、paddind、activation、merge/add、flastten等）层后剩下107层的list

2、从hdf5文件中获取包含权重数据的层的名字

前面通过HDFView看过每一层有一个[“weight_names”]属性，如果不为空，就说明该层存在权重数据。

先看一下控制台对h5py对象f的基本操作（需要的去查看相关数据结构定义）：

>>> f
<HDF5 file "resnet50_weights_tf_dim_ordering_tf_kernels_v2.h5" (mode r)>

>>> f.filename
'E:\\DeepLearning\\keras_test\\models\\resnet50_weights_tf_dim_ordering_tf_kernels_v2.h5'

>>> f.name  
'/'

>>> f.attrs.keys()   # f属性列表 #
<KeysViewHDF5 ['layer_names']>

>>> f.keys() #无顺序
<KeysViewHDF5 ['activation_1', 'activation_10', 'activation_11', 'activation_12', 
...,'activation_8', 'activation_9', 'avg_pool', 'bn2a_branch1', 'bn2a_branch2a', 
...,'res5c_branch2a', 'res5c_branch2b', 'res5c_branch2c', 'zeropadding2d_1']>

>>> f.attrs['layer_names']  #*** 有顺序, 和summary()对应 ****
array([b'input_1', b'zeropadding2d_1', b'conv1', b'bn_conv1',
  b'activation_1', b'maxpooling2d_1', b'res2a_branch2a',
  ..., b'res2a_branch1', b'bn2a_branch2c', b'bn2a_branch1', 
  b'merge_1', b'activation_47', b'res5c_branch2b', b'bn5c_branch2b',
  ..., b'activation_48', b'res5c_branch2c', b'bn5c_branch2c', 
  b'merge_16', b'activation_49', b'avg_pool', b'flatten_1', b'fc1000'],
  dtype='|S15')

>>> f['input_1']
<HDF5 group "/input_1" (0 members)>

>>> f['input_1'].attrs.keys() # 在keras中，每一个层都有‘weight_names'属性 #
<KeysViewHDF5 ['weight_names']>

>>> f['input_1'].attrs['weight_names'] # input层无权重 #
array([], dtype=float64)

>>> f['conv1']
<HDF5 group "/conv1" (2 members)>

>>> f['conv1'].attrs.keys()
<KeysViewHDF5 ['weight_names']>

>>> f['conv1'].attrs['weight_names'] # conv层有权重w、b #
array([b'conv1_W:0', b'conv1_b:0'], dtype='|S9')

从文件中读取具有权重数据的层的名字列表

 # 获取后hdf5文本文件中层的名字，顺序对应
 layer_names = load_attributes_from_hdf5_group(f, 'layer_names')
 #上一句实现 layer_names = [n.decode('utf8') for n in f.attrs['layer_names']]
 filtered_layer_names = []
 for name in layer_names:
  g = f[name]
  weight_names = load_attributes_from_hdf5_group(g, 'weight_names')
  #上一句实现 weight_names = [n.decode('utf8') for n in f[name].attrs['weight_names']]
  #保留有权重层的名字
  if weight_names:
   filtered_layer_names.append(name)
 layer_names = filtered_layer_names
 # 验证模型中有有权重tensor的层 与 从h5中读取有权重层名字的 数量 保持一致。
 if len(layer_names) != len(filtered_layers):
  raise ValueError('You are trying to load a weight file '
       'containing ' + str(len(layer_names)) +
       ' layers into a model with ' +
       str(len(filtered_layers)) + ' layers.')

3、从hdf5文件中读取的权重数据、和keras模型层tf.Variable打包对应

先看一下权重数据、层的权重变量(Tensor tf.Variable)对象，以conv1为例

>>> f['conv1']['conv1_W:0'] # conv1_W:0 权重数据数据集
<HDF5 dataset "conv1_W:0": shape (7, 7, 3, 64), type "<f4">

>>> f['conv1']['conv1_W:0'].value # conv1_W:0 权重数据的值， 是一个标准的4d array
array([[[[ 2.82526277e-02, -1.18737184e-02, 1.51488732e-03, ...,
   -1.07003953e-02, -5.27982824e-02, -1.36667420e-03],
   [ 5.86827798e-03, 5.04415408e-02, 3.46324709e-03, ...,
   1.01423981e-02, 1.39493728e-02, 1.67549420e-02],
   [-2.44090753e-03, -4.86173332e-02, 2.69966386e-03, ...,
   -3.44439060e-04, 3.48098315e-02, 6.28910400e-03]],
  [[ 1.81872323e-02, -7.20698107e-03, 4.80302610e-03, ...,
 …. ]]]])

>>> conv1_w = np.asarray(f['conv1']['conv1_W:0']) # 直接转换成numpy格式 
>>> conv1_w.shape
(7, 7, 3, 64)

# 卷积层
>>> filtered_layers[0]
<keras.layers.convolutional.Conv2D object at 0x000001F7487C0E10>

>>> filtered_layers[0].name
'conv1'

>>> filtered_layers[0].input
<tf.Tensor 'conv1_pad/Pad:0' shape=(?, 230, 230, 3) dtype=float32>

#卷积层权重数据
>>> filtered_layers[0].weights
[<tf.Variable 'conv1/kernel:0' shape=(7, 7, 3, 64) dtype=float32_ref>, 
 <tf.Variable 'conv1/bias:0' shape=(64,) dtype=float32_ref>]

将模型权重数据变量Tensor(tf.Variable)、读取的权重数据打包对应，便于后续将数据写入到权重变量中.

weight_value_tuples = []
# 枚举过滤后的层
for k, name in enumerate(layer_names):
 g = f[name]
 weight_names = load_attributes_from_hdf5_group(g, 'weight_names')
 # 获取文件中当前层的权重数据list， 数据类型转换为numpy array 
 weight_values = [np.asarray(g[weight_name]) for weight_name in weight_names]
 # 获取keras模型中层具有的权重数据tf.Variable个数
 layer = filtered_layers[k]
 symbolic_weights = layer.weights
 # 权重数据预处理
 weight_values = preprocess_weights_for_loading(layer, weight_values,
       original_keras_version, original_backend,reshape=reshape)
 # 验证权重数据、tf.Variable数据是否相同
 if len(weight_values) != len(symbolic_weights):
  raise ValueError('Layer #' + str(k) + '(named "' + layer.name + 
    '" in the current model) was found to correspond to layer ' + name + 
    ' in the save file. However the new layer ' + layer.name + ' expects ' + 
    str(len(symbolic_weights)) + 'weights, but the saved weights have ' + 
    str(len(weight_values)) + ' elements.')
 # tf.Variable 和 权重数据 打包
 weight_value_tuples += zip(symbolic_weights, weight_values)

4、将读取的权重数据写入到层的权重变量中

在3中已经对应好每一层的权重变量Tensor和权重数据，后面将使用tensorflow的sess.run方法进新写入，后面一行代码。

K.batch_set_value(weight_value_tuples)

实际实现

def batch_set_value(tuples):
 if tuples:
  assign_ops = []
  feed_dict = {}
  for x, value in tuples: 
   # 获取权重数据类型  
   value = np.asarray(value, dtype=dtype(x))
   tf_dtype = tf.as_dtype(x.dtype.name.split('_')[0])
   if hasattr(x, '_assign_placeholder'):
    assign_placeholder = x._assign_placeholder
    assign_op = x._assign_op
   else:
    # 权重的tf.placeholder
    assign_placeholder = tf.placeholder(tf_dtype, shape=value.shape)
    # 对权重变量Tensor的赋值 assign的operation
    assign_op = x.assign(assign_placeholder)
    x._assign_placeholder = assign_placeholder # 用处?
    x._assign_op = assign_op     # 用处?
   assign_ops.append(assign_op)
   feed_dict[assign_placeholder] = value
  # 利用tensorflow的tf.Session().run()对tensor进行assign批次赋值
  get_session().run(assign_ops, feed_dict=feed_dict)

至此，先有网络模型，后从h5中加载权重文件结束。后面就可以直接利用模型进行predict了。

三、模型加载 load_model()

这里基本和前面类似，多了一个加载网络而已，后面的权重加载方式一样。

首先将前面加载权重的模型使用 model.save()保存为res50_model.h5，使用HDFView查看

当前res50_model.h5没有额外的参数设置。

处理代码如下

if compile:
 training_config = f.get('training_config')
 if training_config is None:
 warnings.warn('No training configuration found in save file: '
     'the model was *not* compiled. Compile it manually.')
  return model
 training_config = json.loads(training_config.decode('utf-8'))
 optimizer_config = training_config['optimizer_config']
 optimizer = optimizers.deserialize(optimizer_config, custom_objects=custom_objects)
 # Recover loss functions and metrics.
 loss = convert_custom_objects(training_config['loss'])
 metrics = convert_custom_objects(training_config['metrics'])
 sample_weight_mode = training_config['sample_weight_mode']
 loss_weights = training_config['loss_weights']
 # Compile model.
 model.compile(optimizer=optimizer, loss=loss, metrics=metrics,
   loss_weights=loss_weights, sample_weight_mode=sample_weight_mode)
 # Set optimizer weights.
 if 'optimizer_weights' in f:
  # Build train function (to get weight updates).
  model._make_train_function()
  optimizer_weights_group = f['optimizer_weights']
  optimizer_weight_names = [ 
   n.decode('utf8') for n in ptimizer_weights_group['weight_names']]
  optimizer_weight_values = [
   optimizer_weights_group[n] for n in optimizer_weight_names]
  try:
   model.optimizer.set_weights(optimizer_weight_values)
  except ValueError:
   warnings.warn('Error in loading the saved optimizer state. As a result,'
    'your model is starting with a freshly initialized optimizer.')

以上这篇keras读取h5文件load_weights、load代码操作就是小编分享给大家的全部内容了，希望能给大家一个参考，也希望大家多多支持来客网。