關於儲存h5模型、權重網上的示例非常多，也非常簡單。主要有以下兩個函式：

1、keras.models.load_model() 讀取網路、權重

2、keras.models.load_weights() 僅讀取權重

load_model程式碼包含load_weights的程式碼，區別在於load_weights時需要先有網路、並且load_weights需要將權重資料寫入到對應網路層的tensor中。

下面以resnet50載入h5權重為例，示例程式碼如下

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

from network.resnet50 import ResNet50
#修改過，不載入權重（預設官方載入亦可）
model = ResNet50() 

# 引數預設 by_name = Fasle， 否則只讀取匹配的權重
# 這裡h5的層和權重檔案中層名是對應的（除input層）
model.load_weights(r'\models\resnet50_weights_tf_dim_ordering_tf_kernels_v2.h5') 

模型通過 model.summary()輸出

一、模型載入權重 load_weights()

def load_weights(self,filepath,by_name=False,skip_mismatch=False,reshape=False):
 if h5py is None:
  raise ImportError('`load_weights` requires h5py.')
 with h5py.File(filepath,mode='r') as f:
  if 'layer_names' not in f.attrs and 'model_weights' in f:
   f = f['model_weights']
  if by_name:
   saving.load_weights_from_hdf5_group_by_name(
    f,self.layers,skip_mismatch=skip_mismatch,reshape=reshape)
  else:
   saving.load_weights_from_hdf5_group(f,reshape=reshape)

這裡關心函式saving.load_weights_from_hdf5_group(f,reshape=reshape)即可，引數 f 傳遞了一個h5py檔案物件。

讀取h5檔案使用 h5py 包，簡單使用HDFView看一下resnet50的權重檔案。

resnet50_v2 這個權重檔案，僅一個attr “layer_names”,該attr包含177個string的Array，Array中每個元素就是層的名字（這裡是嚴格對應在keras進行儲存權重時網路中每一層的name值，且層的順序也嚴格對應）。

對於每一個key（層名），都有一個屬性"weights_names"，（value值可能為空)。

例如：

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

flatten_1的"weights_names"為null。

這裡就簡單介紹，後面在程式碼中說明h5py如何讀取權重資料。

二、從hdf5檔案中載入權重 load_weights_from_hdf5_group()

1、找出keras模型層中具有weight的Tensor(tf.Variable)的層

def load_weights_from_hdf5_group(f,layers,reshape=False):
 # keras模型resnet50的model.layers的過濾
 # 僅保留layer.weights不為空的層，過濾掉無學習引數的層
 filtered_layers = []
 for layer in layers:
  weights = layer.weights
  if weights:
   filtered_layers.append(layer)

filtered_layers為當前模型resnet50過濾（input、paddind、activation、merge/add、flastten等）層後剩下107層的list

2、從hdf5檔案中獲取包含權重資料的層的名字

前面通過HDFView看過每一層有一個[“weight_names”]屬性，如果不為空，就說明該層存在權重資料。

先看一下控制檯對h5py物件f的基本操作（需要的去檢視相關資料結構定義）：

>>> 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屬性列表 #
<KeysViewHDF5 ['layer_names']>

>>> f.keys() #無順序
<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']  #*** 有順序,和summary()對應 ****
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，每一個層都有‘weight_names'屬性 #
<KeysViewHDF5 ['weight_names']>

>>> f['input_1'].attrs['weight_names'] # input層無權重 #
array([],dtype=float64)

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

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

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

從檔案中讀取具有權重資料的層的名字列表

 # 獲取後hdf5文字檔案中層的名字，順序對應
 layer_names = load_attributes_from_hdf5_group(f,'layer_names')
 #上一句實現 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')
  #上一句實現 weight_names = [n.decode('utf8') for n in f[name].attrs['weight_names']]
  #保留有權重層的名字
  if weight_names:
   filtered_layer_names.append(name)
 layer_names = filtered_layer_names
 # 驗證模型中有有權重tensor的層 與 從h5中讀取有權重層名字的 數量 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、從hdf5檔案中讀取的權重資料、和keras模型層tf.Variable打包對應

先看一下權重資料、層的權重變數(Tensor tf.Variable)物件，以conv1為例

>>> f['conv1']['conv1_W:0'] # conv1_W:0 權重資料資料集
<HDF5 dataset "conv1_W:0": shape (7,7,3,64),type "<f4">

>>> f['conv1']['conv1_W:0'].value # conv1_W:0 權重資料的值， 是一個標準的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']) # 直接轉換成numpy格式 
>>> conv1_w.shape
(7,64)

# 卷積層
>>> 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=(?,230,3) dtype=float32>

#卷積層權重資料
>>> filtered_layers[0].weights
[<tf.Variable 'conv1/kernel:0' shape=(7,64) dtype=float32_ref>,<tf.Variable 'conv1/bias:0' shape=(64,) dtype=float32_ref>]

將模型權重資料變數Tensor(tf.Variable)、讀取的權重資料打包對應，便於後續將資料寫入到權重變數中.

weight_value_tuples = []
# 列舉過濾後的層
for k,name in enumerate(layer_names):
 g = f[name]
 weight_names = load_attributes_from_hdf5_group(g,'weight_names')
 # 獲取檔案中當前層的權重資料list， 資料型別轉換為numpy array 
 weight_values = [np.asarray(g[weight_name]) for weight_name in weight_names]
 # 獲取keras模型中層具有的權重資料tf.Variable個數
 layer = filtered_layers[k]
 symbolic_weights = layer.weights
 # 權重資料預處理
 weight_values = preprocess_weights_for_loading(layer,weight_values,original_keras_version,original_backend,reshape=reshape)
 # 驗證權重資料、tf.Variable資料是否相同
 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 權重資料 打包
 weight_value_tuples += zip(symbolic_weights,weight_values)

4、將讀取的權重資料寫入到層的權重變數中

在3中已經對應好每一層的權重變數Tensor和權重資料，後面將使用tensorflow的sess.run方法進新寫入，後面一行程式碼。

K.batch_set_value(weight_value_tuples)

實際實現

def batch_set_value(tuples):
 if tuples:
  assign_ops = []
  feed_dict = {}
  for x,value in tuples: 
   # 獲取權重資料型別  
   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:
    # 權重的tf.placeholder
    assign_placeholder = tf.placeholder(tf_dtype,shape=value.shape)
    # 對權重變數Tensor的賦值 assign的operation
    assign_op = x.assign(assign_placeholder)
    x._assign_placeholder = assign_placeholder # 用處?
    x._assign_op = assign_op     # 用處?
   assign_ops.append(assign_op)
   feed_dict[assign_placeholder] = value
  # 利用tensorflow的tf.Session().run()對tensor進行assign批次賦值
  get_session().run(assign_ops,feed_dict=feed_dict)

至此，先有網路模型，後從h5中載入權重檔案結束。後面就可以直接利用模型進行predict了。

三、模型載入 load_model()

這裡基本和前面類似，多了一個載入網路而已，後面的權重載入方式一樣。

首先將前面載入權重的模型使用 model.save()儲存為res50_model.h5，使用HDFView檢視

屬性成了3個，backend,keras_version和model_config，用於說明模型檔案由某種後端生成，後端版本，以及json格式的網路模型結構。

有一個key鍵"model_weights",相較於屬性有前面的h5模型，屬性多了2個為['backend','keras_version','layer_names'] 該key鍵下面的鍵值是一個list,和前面的h5模型的權重資料完全一致。

類似的，先利用python程式碼檢視下檔案結構

>>> ff
<HDF5 file "res50_model.h5" (mode r)>

>>> ff.attrs.keys()
<KeysViewHDF5 ['backend','model_config']>

>>> ff.keys()
<KeysViewHDF5 ['model_weights']>

>>> ff['model_weights'].attrs.keys() ## ff['model_weights']有三個屬性
<KeysViewHDF5 ['backend','layer_names']>

>>> ff['model_weights'].keys() ## 無順序
<KeysViewHDF5 ['activation_1',…,'bn2a_branch2b','bn5c_branch2c','bn_conv1','conv1','conv1_pad','fc1000','input_1','c_branch2a','res5c_branch2c']>

>>> ff['model_weights'].attrs['layer_names'] ## 有順序
array([b'input_1',b'conv1_pad',b'pool1_pad',b'max_pooling2d_1',b'bn2a_branch2a',b'activation_2',b'res2a_branch2b',... 省略
  b'activation_48',b'add_16',dtype='|S15')

1、載入模型主函式load_model

def load_model(filepath,custom_objects=None,compile=True):
 if h5py is None:
  raise ImportError('`load_model` requires h5py.')
 model = None
 opened_new_file = not isinstance(filepath,h5py.Group)
 # h5載入後轉換為一個 h5dict 類，編譯通過鍵取值
 f = h5dict(filepath,'r')
 try:
  # 序列化並compile
  model = _deserialize_model(f,custom_objects,compile)
 finally:
  if opened_new_file:
   f.close()
 return model

2、序列化並編譯_deserialize_model

函式def _deserialize_model(f,compile=True)的程式碼顯示主要部分

第一步，載入網路結構，實現完全同keras.models.model_from_json()

# 從h5中讀取網路結構的json描述字串
model_config = f['model_config']
model_config = json.loads(model_config.decode('utf-8'))
# 根據json構建網路模型結構
model = model_from_config(model_config,custom_objects=custom_objects)

第二步，載入網路權重，完全同model.load_weights()

# 獲取有順序的網路層名,網路層
model_weights_group = f['model_weights']
layer_names = model_weights_group['layer_names'] 
layers = model.layers
# 過濾 有權重Tensor的層
for layer in layers:
 weights = layer.weights
 if weights:
  filtered_layers.append(layer)
# 過濾有權重的資料
filtered_layer_names = []
for name in layer_names:
 layer_weights = model_weights_group[name]
 weight_names = layer_weights['weight_names']
 if weight_names:
  filtered_layer_names.append(name)
# 打包資料 weight_value_tuples
weight_value_tuples = []
for k,name in enumerate(layer_names):
 layer_weights = model_weights_group[name]
 weight_names = layer_weights['weight_names']
 weight_values = [layer_weights[weight_name] for weight_name in weight_names]
 layer = filtered_layers[k]
 symbolic_weights = layer.weights
 weight_values = preprocess_weights_for_loading(...)
 weight_value_tuples += zip(symbolic_weights,weight_values) 
# 批寫入 
K.batch_set_value(weight_value_tuples)

第三步，compile並返回模型

正常情況，模型網路建立、載入權重後 compile之後就完成。若還有其他設定，則可以再進行額外的處理。（模型訓練後save會有額外是引數設定）。

例如，一個只有dense層的網路訓練儲存後檢視，屬性多了"training_config"，鍵多了"optimizer_weights"，如下圖。

當前res50_model.h5沒有額外的引數設定。

處理程式碼如下

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.')

以上這篇keras讀取h5檔案load_weights、load程式碼操作就是小編分享給大家的全部內容了，希望能給大家一個參考，也希望大家多多支援我們。