2022年7月更新：现在tensorflow2版本已经发展到2.9,这些模型间的互转可以看官方文档中h5 saved_model各自的缺限，默认使用saved model来保存，推荐这种方式。
本文于2020年11月2日（本博客一年之后）增加更新实验，思路与本文大体相似，理解模型间的转换建议两者都看看，参见https://blog.csdn.net/u011119817/article/details/109447755。
1、训练模型
我们将把训练好的模型分别保存成hdf5和saved model格式，然后完成它们之间的互相转换以及分别转tensorflow1.x的pb格式，具体有：
- hdf5转saved model,并验证转换后的saved model与直接保存的saved model的无差异性（大小，精度）
- saved model转hdf5,并验证转换后的hdf5与直接保存的hdf5的无差异性（大小，精度）
- hdf5转pb,并验证转换后的pb与直接原始的的hdf5的无差异性（大小，精度）
- saved mode转pb,并验证转换后的pb与直接原始的的saved mode的无差异性（大小，精度）
- 对比hdf5所转pb与saved model所转pb的区别
import tensorflow as tf
import os
from functools import partial
import numpy as np
import shutil
print("tf.__version__")
tf.__version__
batch_size=64
epochs=6
regularizer=1e-3
total_train_samples=60000
total_test_samples=10000
output_folder="/tmp/test/hdf5_model"
output_folder1="/tmp/test/saved_model"
output_folder2="/tmp/test/pb_model"
for m in (output_folder,output_folder1,output_folder2):
if os.path.exists(m):
inc=input("The model(%s) saved path has exist,Do you want to delete and remake it?(y/n)"%m)
while(inc.lower() not in ['y','n']):
inc=input("The model saved path has exist,Do you want to delete and remake it?(y/n)")
if inc.lower()=='y':
shutil.rmtree(m)
os.makedirs(m)
elif not os.path.exists(m):
os.makedirs(m)
The model(/tmp/test/hdf5_model) saved path has exist,Do you want to delete and remake it?(y/n)y
The model(/tmp/test/saved_model) saved path has exist,Do you want to delete and remake it?(y/n)y
#指定显卡
physical_devices = tf.config.experimental.list_physical_devices('GPU')#列出所有可见显卡
print("All the available GPUs:\n",physical_devices)
if physical_devices:
gpu=physical_devices[0]#显示第一块显卡
tf.config.experimental.set_memory_growth(gpu, True)#根据需要自动增长显存
tf.config.experimental.set_visible_devices(gpu, 'GPU')#只选择第一块
All the available GPUs:
[PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]
#准备数据
fashion_mnist=tf.keras.datasets.fashion_mnist
(train_x,train_y),(test_x,test_y)=fashion_mnist.load_data()
train_x,test_x = train_x[...,np.newaxis]/255.0,test_x[...,np.newaxis]/255.0
class_names = ['T-shirt/top', 'Trouser', 'Pullover', 'Dress', 'Coat','Sandal', 'Shirt', 'Sneaker', 'Bag', 'Ankle boot']
train_ds = tf.data.Dataset.from_tensor_slices((train_x,train_y))
test_ds = tf.data.Dataset.from_tensor_slices((test_x,test_y))
train_ds=train_ds.shuffle(buffer_size=batch_size*10).batch(batch_size).prefetch(buffer_size = tf.data.experimental.AUTOTUNE).repeat()
test_ds = test_ds.batch(batch_size).prefetch(buffer_size = tf.data.experimental.AUTOTUNE)#不加repeat，执行一次就行
#定义模型
l2 = tf.keras.regularizers.l2(regularizer)#定义模型正则化方法
ini = tf.keras.initializers.he_normal()#定义参数初始化方法
conv2d = partial(tf.keras.layers.Conv2D,activation='relu',padding='same',kernel_regularizer=l2,bias_regularizer=l2)
fc = partial(tf.keras.layers.Dense,activation='relu',kernel_regularizer=l2,bias_regularizer=l2)
maxpool=tf.keras.layers.MaxPooling2D
dropout=tf.keras.layers.Dropout
def test_model():
x_input = tf.keras.layers.Input(shape=(28,28,1),name='input_node')
x = conv2d(128,(5,5))(x_input)
x = maxpool((2,2))(x)
x = conv2d(256,(5,5))(x)
x = maxpool((2,2))(x)
x = tf.keras.layers.Flatten()(x)
x = fc(128)(x)
x_output=fc(10,activation=None,name='output_node')(x)
model = tf.keras.models.Model(inputs=x_input,outputs=x_output)
return model
model = test_model()
print(model.summary())
Model: "model"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_node (InputLayer) [(None, 28, 28, 1)] 0
_________________________________________________________________
conv2d (Conv2D) (None, 28, 28, 128) 3328
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 14, 14, 128) 0
_________________________________________________________________
conv2d_1 (Conv2D) (None, 14, 14, 256) 819456
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 7, 7, 256) 0
_________________________________________________________________
flatten (Flatten) (None, 12544) 0
_________________________________________________________________
dense (Dense) (None, 128) 1605760
_________________________________________________________________
output_node (Dense) (None, 10) 1290
=================================================================
Total params: 2,429,834
Trainable params: 2,429,834
Non-trainable params: 0
_________________________________________________________________
None
#编译模型
initial_learning_rate=0.01
optimizer = tf.keras.optimizers.SGD(learning_rate=initial_learning_rate,momentum=0.95)
loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
metrics=['accuracy','sparse_categorical_crossentropy']
model.compile(optimizer=optimizer,loss=loss,metrics=metrics)
#训练模型
H=model.fit(train_ds,epochs=6,
steps_per_epoch=np.floor(len(train_x)/batch_size).astype(np.int32),
validation_data=test_ds,
validation_steps=np.ceil(len(test_x)/batch_size).astype(np.int32),
verbose=1)
Train for 937 steps, validate for 157 steps
Epoch 1/6
937/937 [==============================] - 78s 84ms/step - loss: 0.9485 - accuracy: 0.8000 - sparse_categorical_crossentropy: 1.5262 - val_loss: 0.6896 - val_accuracy: 0.8527 - val_sparse_categorical_crossentropy: 1.3110
Epoch 2/6
937/937 [==============================] - 76s 82ms/step - loss: 0.5635 - accuracy: 0.8824 - sparse_categorical_crossentropy: 1.2657 - val_loss: 0.5024 - val_accuracy: 0.8892 - val_sparse_categorical_crossentropy: 1.2297
Epoch 3/6
937/937 [==============================] - 76s 81ms/step - loss: 0.4517 - accuracy: 0.8971 - sparse_categorical_crossentropy: 1.2169 - val_loss: 0.4438 - val_accuracy: 0.8928 - val_sparse_categorical_crossentropy: 1.1814
Epoch 4/6
937/937 [==============================] - 76s 81ms/step - loss: 0.4077 - accuracy: 0.9015 - sparse_categorical_crossentropy: 1.1756 - val_loss: 0.4136 - val_accuracy: 0.8995 - val_sparse_categorical_crossentropy: 1.1607
Epoch 5/6
937/937 [==============================] - 76s 81ms/step - loss: 0.3850 - accuracy: 0.9051 - sparse_categorical_crossentropy: 1.1504 - val_loss: 0.3992 - val_accuracy: 0.8978 - val_sparse_categorical_crossentropy: 1.1329
Epoch 6/6
937/937 [==============================] - 76s 81ms/step - loss: 0.3719 - accuracy: 0.9093 - sparse_categorical_crossentropy: 1.1388 - val_loss: 0.4391 - val_accuracy: 0.8844 - val_sparse_categorical_crossentropy: 1.1219
#分别保存两种格式的模型
model.save(filepath=os.path.join(output_folder,'hdf5_model.h5'),save_format='h5')
model.save(filepath=output_folder1,save_format='tf')
#报的warning信息在tensorflow官网的例子中同样存在
INFO:tensorflow:Assets written to: /tmp/test/saved_model/assets
%%bash
echo -e "hdf5 model information...\n"
tree "/tmp/test/hdf5_model"
du -ah "/tmp/test/hdf5_model"
echo -e "saved model information...\n"
tree "/tmp/test/saved_model"
du -ah "/tmp/test/saved_model"
hdf5 model information...
/tmp/test/hdf5_model
└── hdf5_model.h5
0 directories, 1 file
19M /tmp/test/hdf5_model/hdf5_model.h5
19M /tmp/test/hdf5_model
saved model information...
/tmp/test/saved_model
├── assets
├── saved_model.pb
└── variables
├── variables.data-00000-of-00002
├── variables.data-00001-of-00002
└── variables.index
2 directories, 4 files
4.0K /tmp/test/saved_model/assets
216K /tmp/test/saved_model/saved_model.pb
19M /tmp/test/saved_model/variables/variables.data-00001-of-00002
4.0K /tmp/test/saved_model/variables/variables.index
8.0K /tmp/test/saved_model/variables/variables.data-00000-of-00002
19M /tmp/test/saved_model/variables
19M /tmp/test/saved_model
#选取第一个样本来做为测试样本,用来评估不同模型之间的精度
test_sample = train_x[0:1]
test_y=train_y[0]
out = model.predict(test_sample)
print("probs:",out[0])
print("true label:{} pred label:{}".format(test_y,np.argmax(out)))
probs: [-2.0793445e+00 -2.2612031e+00 -1.8440809e+00 -1.1460640e+00
-1.9762940e+00 8.9537799e-03 -3.4592066e+00 3.3828874e+00
1.5507856e-01 9.2633562e+00]
true label:9 pred label:9
2、各种模型间互转并验证
2.1 hdf5转saved model
rm -rf /tmp/test/hdf52saved_model
ls /tmp/test/hdf5_model
hdf5_model.h5
tf.keras.backend.clear_session()
hdf5_model = tf.keras.models.load_model("/tmp/test/hdf5_model/hdf5_model.h5")
hdf5_model.save("/tmp/test/hdf52saved_model",save_format='tf')
INFO:tensorflow:Assets written to: /tmp/test/hdf52saved_model/assets
%%bash
echo -e "hdf5 model information...\n"
tree "/tmp/test/hdf5_model"
du -ah "/tmp/test/hdf5_model"
echo -e "saved model information...\n"
tree "/tmp/test/saved_model"
du -ah "/tmp/test/saved_model"
echo -e "saved model information...\n"
tree "/tmp/test/hdf52saved_model"
du -ah "/tmp/test/hdf52saved_model"
hdf5 model information...
/tmp/test/hdf5_model
└── hdf5_model.h5
0 directories, 1 file
19M /tmp/test/hdf5_model/hdf5_model.h5
19M /tmp/test/hdf5_model
saved model information...
/tmp/test/saved_model
├── assets
├── saved_model.pb
└── variables
├── variables.data-00000-of-00002
├── variables.data-00001-of-00002
└── variables.index
2 directories, 4 files
4.0K /tmp/test/saved_model/assets
216K /tmp/test/saved_model/saved_model.pb
19M /tmp/test/saved_model/variables/variables.data-00001-of-00002
4.0K /tmp/test/saved_model/variables/variables.index
8.0K /tmp/test/saved_model/variables/variables.data-00000-of-00002
19M /tmp/test/saved_model/variables
19M /tmp/test/saved_model
saved model information...
/tmp/test/hdf52saved_model
├── assets
├── saved_model.pb
└── variables
├── variables.data-00000-of-00002
├── variables.data-00001-of-00002
└── variables.index
2 directories, 4 files
4.0K /tmp/test/hdf52saved_model/assets
224K /tmp/test/hdf52saved_model/saved_model.pb
8.0K /tmp/test/hdf52saved_model/variables/variables.data-00001-of-00002
4.0K /tmp/test/hdf52saved_model/variables/variables.index
19M /tmp/test/hdf52saved_model/variables/variables.data-00000-of-00002
19M /tmp/test/hdf52saved_model/variables
19M /tmp/test/hdf52saved_model
2.2 saved model转hdf5
tf.keras.backend.clear_session()
saved_model = tf.keras.models.load_model("/tmp/test/saved_model")
saved_model.save("/tmp/test/hdf5_model/saved2hdf5_model.h5",save_format='h5')
---------------------------------------------------------------------------
NotImplementedError Traceback (most recent call last)
<ipython-input-144-322a2edc9a6e> in <module>
1 tf.keras.backend.clear_session()
2 saved_model = tf.keras.models.load_model("/tmp/test/saved_model")
----> 3 saved_model.save("/tmp/test/hdf5_model/saved2hdf5_model.h5",save_format='h5')
~/miniconda3/envs/tf2/lib/python3.6/site-packages/tensorflow_core/python/keras/engine/network.py in save(self, filepath, overwrite, include_optimizer, save_format, signatures, options)
973 """
974 saving.save_model(self, filepath, overwrite, include_optimizer, save_format,
--> 975 signatures, options)
976
977 def save_weights(self, filepath, overwrite=True, save_format=None):
~/miniconda3/envs/tf2/lib/python3.6/site-packages/tensorflow_core/python/keras/saving/save.py in save_model(model, filepath, overwrite, include_optimizer, save_format, signatures, options)
103 not isinstance(model, sequential.Sequential)):
104 raise NotImplementedError(
--> 105 'Saving the model to HDF5 format requires the model to be a '
106 'Functional model or a Sequential model. It does not work for '
107 'subclassed models, because such models are defined via the body of '
NotImplementedError: Saving the model to HDF5 format requires the model to be a Functional model or a Sequential model. It does not work for subclassed models, because such models are defined via the body of a Python method, which isn't safely serializable. Consider saving to the Tensorflow SavedModel format (by setting save_format="tf") or using `save_weights`.
以上信息说明，从saved model是无法转换成hdf5模型的，所以个人感觉在训练过程中保存hdf5格式的模型比较好
2.3 所有模型精度测试
测试三个模型的精度，原始hdf5模型，原始saved model，hdf5转换的saved model
origin_hdf5_model = tf.keras.models.load_model("/tmp/test/hdf5_model/hdf5_model.h5")
origin_saved_model = tf.keras.models.load_model("/tmp/test/saved_model")
converted_saved_model = tf.keras.models.load_model("/tmp/test/hdf52saved_model")
out1 = origin_hdf5_model.predict(test_sample)
out2 = origin_saved_model.predict(test_sample)
out3 = converted_saved_model.predict(test_sample)
print("probs:",out1[0])
print("true label:{} pred label:{}".format(test_y,np.argmax(out1)))
print("probs:",out2[0])
print("true label:{} pred label:{}".format(test_y,np.argmax(out2)))
print("probs:",out3[0])
print("true label:{} pred label:{}".format(test_y,np.argmax(out3)))
np.testing.assert_array_almost_equal(out,out1)
np.testing.assert_array_almost_equal(out,out2)
np.testing.assert_array_almost_equal(out,out3)
probs: [-2.0793445e+00 -2.2612031e+00 -1.8440809e+00 -1.1460640e+00
-1.9762940e+00 8.9537799e-03 -3.4592066e+00 3.3828874e+00
1.5507856e-01 9.2633562e+00]
true label:9 pred label:9
probs: [-2.0793445e+00 -2.2612031e+00 -1.8440809e+00 -1.1460640e+00
-1.9762940e+00 8.9537799e-03 -3.4592066e+00 3.3828874e+00
1.5507856e-01 9.2633562e+00]
true label:9 pred label:9
probs: [-2.0793445e+00 -2.2612031e+00 -1.8440809e+00 -1.1460640e+00
-1.9762940e+00 8.9537799e-03 -3.4592066e+00 3.3828874e+00
1.5507856e-01 9.2633562e+00]
true label:9 pred label:9
可以看到结果完全一致，模型转换没有问题
2.4 hdf5和saved模型转tensorflow1.x pb模型
我们需要在tensorflow2.0中使用tensorflow1.x内容
以下是hdf5转pb模型
import tensorflow.compat.v1 as tf1
tf1.reset_default_graph()
tf1.keras.backend.set_learning_phase(0) #调用模型前一定要执行该命令
tf1.disable_v2_behavior() #禁止tensorflow2.0的行为
#加载hdf5模型
hdf5_pb_model = tf.keras.models.load_model("/tmp/test/hdf5_model/hdf5_model.h5")
def freeze_session(session,keep_var_names=None,output_names=None,clear_devices=True):
graph = session.graph
with graph.as_default():
# freeze_var_names = list(set(v.op.name for v in tf1.global_variables()).difference(keep_var_names or []))
output_names = output_names or []
# output_names += [v.op.name for v in tf1.global_variables()]
print("output_names",output_names)
input_graph_def = graph.as_graph_def()
# for node in input_graph_def.node:
# print('node:', node.name)
print("len node1",len(input_graph_def.node))
if clear_devices:
for node in input_graph_def.node:
node.device = ""
frozen_graph = tf1.graph_util.convert_variables_to_constants(session, input_graph_def,
output_names)
outgraph = tf1.graph_util.remove_training_nodes(frozen_graph)#云掉与推理无关的内容
print("##################################################################")
for node in outgraph.node:
print('node:', node.name)
print("len node1",len(outgraph.node))
return outgraph
frozen_graph = freeze_session(tf1.keras.backend.get_session(),output_names=[out.op.name for out in hdf5_pb_model.outputs])
tf1.train.write_graph(frozen_graph, output_folder2, "hdf52pb.pb", as_text=False)
output_names ['output_node/BiasAdd']
len node1 626
INFO:tensorflow:Froze 8 variables.
INFO:tensorflow:Converted 8 variables to const ops.
##################################################################
node: input_node
node: conv2d/kernel
node: conv2d/bias
node: conv2d/Conv2D
node: conv2d/BiasAdd
node: conv2d/Relu
node: max_pooling2d/MaxPool
node: conv2d_1/kernel
node: conv2d_1/bias
node: conv2d_1/Conv2D
node: conv2d_1/BiasAdd
node: conv2d_1/Relu
node: max_pooling2d_1/MaxPool
node: flatten/Reshape/shape
node: flatten/Reshape
node: dense/kernel
node: dense/bias
node: dense/MatMul
node: dense/BiasAdd
node: dense/Relu
node: output_node/kernel
node: output_node/bias
node: output_node/MatMul
node: output_node/BiasAdd
len node1 24
'/tmp/test/pb_model/hdf52pb.pb'
以下是saved model转pb模型
import tensorflow.compat.v1 as tf1
tf1.reset_default_graph()
tf1.keras.backend.set_learning_phase(0) #调用模型前一定要执行该命令
tf1.disable_v2_behavior() #禁止tensorflow2.0的行为
#加载hdf5模型
saved_pb_model = tf.keras.models.load_model("/tmp/test/saved_model")
def freeze_session(session,keep_var_names=None,output_names=None,clear_devices=True):
graph = session.graph
with graph.as_default():
# freeze_var_names = list(set(v.op.name for v in tf1.global_variables()).difference(keep_var_names or []))
output_names = output_names or []
# output_names += [v.op.name for v in tf1.global_variables()]
print("output_names",output_names)
input_graph_def = graph.as_graph_def()
# for node in input_graph_def.node:
# print('node:', node.name)
print("len node1",len(input_graph_def.node))
if clear_devices:
for node in input_graph_def.node:
node.device = ""
frozen_graph = tf1.graph_util.convert_variables_to_constants(session, input_graph_def,
output_names)
outgraph = tf1.graph_util.remove_training_nodes(frozen_graph)#云掉与推理无关的内容
print("##################################################################")
for node in outgraph.node:
print('node:', node.name)
print("len node1",len(outgraph.node))
return outgraph
frozen_graph = freeze_session(tf1.keras.backend.get_session(),output_names=[out.op.name for out in saved_pb_model.outputs])
tf1.train.write_graph(frozen_graph, output_folder2, "saved2pb.pb", as_text=False)
output_names ['model/StatefulPartitionedCall']
len node1 304
INFO:tensorflow:Froze 8 variables.
INFO:tensorflow:Converted 8 variables to const ops.
##################################################################
node: conv2d/kernel
node: conv2d/bias
node: conv2d_1/kernel
node: conv2d_1/bias
node: dense/kernel
node: dense/bias
node: output_node/kernel
node: output_node/bias
node: input_1
node: model/StatefulPartitionedCall
len node1 10
'/tmp/test/pb_model/saved2pb.pb'
以下是转换后的saved model转pb
import tensorflow.compat.v1 as tf1
tf1.reset_default_graph()
tf1.keras.backend.set_learning_phase(0) #调用模型前一定要执行该命令
tf1.disable_v2_behavior() #禁止tensorflow2.0的行为
#加载hdf5模型
hdf52saved_pb_model = tf.keras.models.load_model("/tmp/test/hdf52saved_model/")
def freeze_session(session,keep_var_names=None,output_names=None,clear_devices=True):
graph = session.graph
with graph.as_default():
# freeze_var_names = list(set(v.op.name for v in tf1.global_variables()).difference(keep_var_names or []))
output_names = output_names or []
# output_names += [v.op.name for v in tf1.global_variables()]
print("output_names",output_names)
input_graph_def = graph.as_graph_def()
# for node in input_graph_def.node:
# print('node:', node.name)
print("len node1",len(input_graph_def.node))
if clear_devices:
for node in input_graph_def.node:
node.device = ""
frozen_graph = tf1.graph_util.convert_variables_to_constants(session, input_graph_def,
output_names)
outgraph = tf1.graph_util.remove_training_nodes(frozen_graph)#云掉与推理无关的内容
print("##################################################################")
for node in outgraph.node:
print('node:', node.name)
print("len node1",len(outgraph.node))
return outgraph
frozen_graph = freeze_session(tf1.keras.backend.get_session(),output_names=[out.op.name for out in saved_pb_model.outputs])
tf1.train.write_graph(frozen_graph, output_folder2, "hdf52saved2pb.pb", as_text=False)
output_names ['model/StatefulPartitionedCall']
len node1 304
INFO:tensorflow:Froze 8 variables.
INFO:tensorflow:Converted 8 variables to const ops.
##################################################################
node: conv2d/kernel
node: conv2d/bias
node: conv2d_1/kernel
node: conv2d_1/bias
node: dense/kernel
node: dense/bias
node: output_node/kernel
node: output_node/bias
node: input_1
node: model/StatefulPartitionedCall
len node1 10
'/tmp/test/pb_model/hdf52saved2pb.pb'
%%bash
tree /tmp/test
du -ah /tmp/test
/tmp/test
├── hdf52saved_model
│ ├── assets
│ ├── saved_model.pb
│ └── variables
│ ├── variables.data-00000-of-00002
│ ├── variables.data-00001-of-00002
│ └── variables.index
├── hdf5_model
│ └── hdf5_model.h5
├── pb_model
│ ├── hdf52pb.pb
│ ├── hdf52saved2pb.pb
│ └── saved2pb.pb
└── saved_model
├── assets
├── saved_model.pb
└── variables
├── variables.data-00000-of-00002
├── variables.data-00001-of-00002
└── variables.index
8 directories, 12 files
19M /tmp/test/hdf5_model/hdf5_model.h5
19M /tmp/test/hdf5_model
9.3M /tmp/test/pb_model/hdf52saved2pb.pb
9.3M /tmp/test/pb_model/saved2pb.pb
9.3M /tmp/test/pb_model/hdf52pb.pb
28M /tmp/test/pb_model
4.0K /tmp/test/saved_model/assets
216K /tmp/test/saved_model/saved_model.pb
19M /tmp/test/saved_model/variables/variables.data-00001-of-00002
4.0K /tmp/test/saved_model/variables/variables.index
8.0K /tmp/test/saved_model/variables/variables.data-00000-of-00002
19M /tmp/test/saved_model/variables
19M /tmp/test/saved_model
4.0K /tmp/test/hdf52saved_model/assets
224K /tmp/test/hdf52saved_model/saved_model.pb
8.0K /tmp/test/hdf52saved_model/variables/variables.data-00001-of-00002
4.0K /tmp/test/hdf52saved_model/variables/variables.index
19M /tmp/test/hdf52saved_model/variables/variables.data-00000-of-00002
19M /tmp/test/hdf52saved_model/variables
19M /tmp/test/hdf52saved_model
84M /tmp/test
可以看到三个pb模型的大小是相同的，但了节点名称不一样，且打印出来的名称顺序在hdf5模型体现更好。
另外，hdf5模型不经保存也可以直接保存成pb,但是保存再读取和直接保存节点名称会变，但精度还是相同的。所以可以把节点名称打印出来分析。
2.5 加载并测试pb模型
有三个pb模型分别进行测试
import tensorflow.compat.v1 as tf1
import numpy as np
def load_graph(file_path):
with tf1.gfile.GFile(file_path,'rb') as f:
graph_def = tf1.GraphDef()
graph_def.ParseFromString(f.read())
with tf1.Graph().as_default() as graph:
tf1.import_graph_def(graph_def,input_map = None,return_elements = None,name = "",op_dict = None,producer_op_list = None)
graph_nodes = [n for n in graph_def.node]
return graph,graph_nodes
三个模型依次调用：
- /tmp/test/pb_model/hdf52saved2pb.pb
- /tmp/test/pb_model/saved2pb.pb
- /tmp/test/pb_model/hdf52pb.pb
第一个模型
file_path='/tmp/test/pb_model/hdf52pb.pb'
graph,graph_nodes = load_graph(file_path)
print("num nodes",len(graph_nodes))
for node in graph_nodes:
print('node:', node.name)
num nodes 24
node: input_node
node: conv2d/kernel
node: conv2d/bias
node: conv2d/Conv2D
node: conv2d/BiasAdd
node: conv2d/Relu
node: max_pooling2d/MaxPool
node: conv2d_1/kernel
node: conv2d_1/bias
node: conv2d_1/Conv2D
node: conv2d_1/BiasAdd
node: conv2d_1/Relu
node: max_pooling2d_1/MaxPool
node: flatten/Reshape/shape
node: flatten/Reshape
node: dense/kernel
node: dense/bias
node: dense/MatMul
node: dense/BiasAdd
node: dense/Relu
node: output_node/kernel
node: output_node/bias
node: output_node/MatMul
node: output_node/BiasAdd
input_node = graph.get_tensor_by_name('input_node:0')
output = graph.get_tensor_by_name('output_node/BiasAdd:0')
config = tf1.ConfigProto()
config.gpu_options.allow_growth = True
# config.gpu_options.per_process_gpu_memory_fraction = 0.25# 设定GPU使用占比
config.gpu_options.visible_device_list = '0' # '0,1'
config.allow_soft_placement = True
config.log_device_placement = False
with tf1.Session(config=config,graph=graph) as sess:
logits = sess.run(output, feed_dict = {input_node:test_sample})
print("logits:",logits)
np.testing.assert_array_almost_equal(out,logits)
logits: [[-2.0793445e+00 -2.2612031e+00 -1.8440809e+00 -1.1460640e+00
-1.9762940e+00 8.9537799e-03 -3.4592066e+00 3.3828874e+00
1.5507856e-01 9.2633562e+00]]
从以上结果可以看到，hdf5转换的pb结果完全正确
另外经过验证，使用tensorflow2.x中tf.compat.v1 api转换成的pb模型，只能用tensorflow1.14和tensorflow1.15两个版本调用使用。
第二个模型
file_path='/tmp/test/pb_model/saved2pb.pb'
graph,graph_nodes = load_graph(file_path)
print("num nodes",len(graph_nodes))
for node in graph_nodes:
print('node:', node.name)
---------------------------------------------------------------------------
InvalidArgumentError Traceback (most recent call last)
~/miniconda3/envs/tf2/lib/python3.6/site-packages/tensorflow_core/python/framework/importer.py in _import_graph_def_internal(graph_def, input_map, return_elements, validate_colocation_constraints, name, op_dict, producer_op_list)
500 results = c_api.TF_GraphImportGraphDefWithResults(
--> 501 graph._c_graph, serialized, options) # pylint: disable=protected-access
502 results = c_api_util.ScopedTFImportGraphDefResults(results)
InvalidArgumentError: Input 1 of node model/StatefulPartitionedCall was passed float from conv2d/kernel:0 incompatible with expected resource.
During handling of the above exception, another exception occurred:
ValueError Traceback (most recent call last)
<ipython-input-108-4ad5ab9ba9bb> in <module>
1 file_path='/tmp/test/pb_model/saved2pb.pb'
----> 2 graph,graph_nodes = load_graph(file_path)
3 print("num nodes",len(graph_nodes))
4 for node in graph_nodes:
5 print('node:', node.name)
<ipython-input-103-6c7963fc55a7> in load_graph(file_path)
7 graph_def.ParseFromString(f.read())
8 with tf1.Graph().as_default() as graph:
----> 9 tf1.import_graph_def(graph_def,input_map = None,return_elements = None,name = "",op_dict = None,producer_op_list = None)
10 graph_nodes = [n for n in graph_def.node]
11 return graph,graph_nodes
~/miniconda3/envs/tf2/lib/python3.6/site-packages/tensorflow_core/python/util/deprecation.py in new_func(*args, **kwargs)
505 'in a future version' if date is None else ('after %s' % date),
506 instructions)
--> 507 return func(*args, **kwargs)
508
509 doc = _add_deprecated_arg_notice_to_docstring(
~/miniconda3/envs/tf2/lib/python3.6/site-packages/tensorflow_core/python/framework/importer.py in import_graph_def(graph_def, input_map, return_elements, name, op_dict, producer_op_list)
403 name=name,
404 op_dict=op_dict,
--> 405 producer_op_list=producer_op_list)
406
407
~/miniconda3/envs/tf2/lib/python3.6/site-packages/tensorflow_core/python/framework/importer.py in _import_graph_def_internal(graph_def, input_map, return_elements, validate_colocation_constraints, name, op_dict, producer_op_list)
503 except errors.InvalidArgumentError as e:
504 # Convert to ValueError for backwards compatibility.
--> 505 raise ValueError(str(e))
506
507 # Create _DefinedFunctions for any imported functions.
ValueError: Input 1 of node model/StatefulPartitionedCall was passed float from conv2d/kernel:0 incompatible with expected resource.
可以看出,不论是原始的saved model还是hdf5转成的saved_model,都可以转pb，但使用中还是会报错。
另外一种调用hdf5转换的pb(或在tensorflow2.x中调用tensorflow1.x转的pb)
tf1.reset_default_graph()
tf1.enable_v2_behavior()#tensorflow2.x中调用tensorflow1.x的内容需要激活tensorflow2.x的特性
tf.keras.backend.clear_session()
def wrap_frozen_graph(graph_def, inputs, outputs):
def _imports_graph_def():
tf1.import_graph_def(graph_def, name="")
wrapped_import = tf1.wrap_function(_imports_graph_def, [])
import_graph = wrapped_import.graph
return wrapped_import.prune(
tf.nest.map_structure(import_graph.as_graph_element, inputs),
tf.nest.map_structure(import_graph.as_graph_element, outputs))
file_path='/tmp/test/pb_model/hdf52pb.pb'
with open(file_path,'rb') as f:
graph_def = tf1.GraphDef()
graph_def.ParseFromString(f.read())
for node in graph_def.node:
print("node.name",node.name)
node.name input_node
node.name conv2d/kernel
node.name conv2d/bias
node.name conv2d/Conv2D
node.name conv2d/BiasAdd
node.name conv2d/Relu
node.name max_pooling2d/MaxPool
node.name conv2d_1/kernel
node.name conv2d_1/bias
node.name conv2d_1/Conv2D
node.name conv2d_1/BiasAdd
node.name conv2d_1/Relu
node.name max_pooling2d_1/MaxPool
node.name flatten/Reshape/shape
node.name flatten/Reshape
node.name dense/kernel
node.name dense/bias
node.name dense/MatMul
node.name dense/BiasAdd
node.name dense/Relu
node.name output_node/kernel
node.name output_node/bias
node.name output_node/MatMul
node.name output_node/BiasAdd
model_func = wrap_frozen_graph(
graph_def, inputs='input_node:0',
outputs='output_node/BiasAdd:0')
o=model_func(tf.constant(test_sample,dtype=tf.float32))
print(o)
np.testing.assert_array_almost_equal(out,o.numpy())
tf.Tensor(
[[-2.0793445e+00 -2.2612031e+00 -1.8440809e+00 -1.1460640e+00
-1.9762940e+00 8.9537799e-03 -3.4592066e+00 3.3828874e+00
1.5507856e-01 9.2633562e+00]], shape=(1, 10), dtype=float32)
总结
- tensorflow2.x保存的hdf5模型可以转tensorflow1.x的pb ,也可以转tensorflow2.x saved model
- saved model可以转pb ,但是转换后无法使用
- saved model不可以转换成hdf5模型
- 在tensorflow2.x中可以使用tensorflow1.x或tensorflow2.x的语法来调用，从而选则不同版本
- tensorflow2.x训练的模型 ，转换成pb后，只能用tensorflow1.14和1.15来调用。
所以我们在以后可以只保存hdf5模型，这样可以使用tensorflow2.x来训练模型，如果要用tensorflow2.x推荐的格式，就把hdf5转换成save_model来用；如果要用旧的tensorflow1.x版本，可以把hdf5转换成pb来用。
