【Tensorflow2.0】8、tensorflow2.0_hdf5_savedmodel_pb模型转换[2]
2024-06-05 07:29:03
文章目录
- 1、生成分类模型
- 2、分别用h5和saved model做推理
- 2.1 h5模型推理
- 2.2 saved_model进行推理
- 3、模型转换pb
- 3.1 h5转pb
- 3.2 saved model转pb
- 4、调用pb模型进行推理
- 4.1 h5 pb的推理
- 4.2 saved pb的推理
- 5、h5和saved 尝试互转
- 5.1 h5转saved
- 5.1.1 h5 转saved后的模型推理
- 5.1.2 h5转saved后的模型转pb
- 5.1.3 h5转saved模型后转pb推理
- 5.2 saved model转h5
- 5.2.1 saved转h5后的模型推理
- 5.2.2 saved 转h5后的模型转pb
- 5.2.3 saved转h5后模型转pb后的推理
- 6、互转部分总结
- 6.1 精度对比
- 6.2 模型大小对比
- 总结：
前期工作（一年前，2019年11月）参见https://blog.csdn.net/u011119817/article/details/103264080，随着框架版本的发展，有些结果不尽相同，因此更新本次博客，建意两者都看看，以便给自己的工作提供指导。
本文更新于2020年11月2日，所用代码全部在tensorflow gpu 2.2版本下进行，对于pb的使用，理论上tensorflow 1.14及以上版本都是可以的。关于本次实验的结论，大家可以直接去看结论部分，总的来说就是h5 和savedmodel两种格式可以互相转换，转换后可以继续进行转pb,所有可以互转的工作全部可以做。
另外，本博客的结论是基于一个很简单的模型来做的，所以对于更加复杂的模型，更少见的算子等情况，还要自己动手做实验，毕竟模型转换是一种实验性工作。
1、生成分类模型
import tensorflow as tf
print(tf.__version__)
2.2.0
#use tf2.x to train a model,this code can be run both by tf1.14 and tf2.x# This file contains functions for training a TensorFlow model
import tensorflow as tf
import numpy as np
import os
print("tensorflow version",tf.__version__)
def process_dataset():# Import the data(x_train, y_train),(x_test, y_test) = tf.keras.datasets.mnist.load_data()x_train, x_test = x_train / 255.0, x_test / 255.0# Reshape the dataNUM_TRAIN = 60000NUM_TEST = 10000x_train = np.reshape(x_train, (NUM_TRAIN, 28, 28, 1))x_test = np.reshape(x_test, (NUM_TEST, 28, 28, 1))return x_train, y_train, x_test, y_testdef create_model():model = tf.keras.models.Sequential()model.add(tf.keras.layers.InputLayer(input_shape=[28,28, 1]))model.add(tf.keras.layers.Flatten())model.add(tf.keras.layers.Dense(512, activation=tf.nn.relu))model.add(tf.keras.layers.Dense(10, activation=tf.nn.softmax))model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])return modeldef main():x_train, y_train, x_test, y_test = process_dataset()model = create_model()model.summary()# Train the model on the datamodel.fit(x_train, y_train, epochs = 5, verbose = 1)# Evaluate the model on test datamodel.evaluate(x_test, y_test)model.save("models/lenet5.h5",save_format='h5')model.save("models/lenet5",save_format='tf')# save(model, filename="models/lenet5.pb")# if __name__ == '__main__':
# main()
if not os.path.exists('models'):os.mkdir('models')
main()
tensorflow version 2.2.0
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
flatten (Flatten) (None, 784) 0
_________________________________________________________________
dense (Dense) (None, 512) 401920
_________________________________________________________________
dense_1 (Dense) (None, 10) 5130
=================================================================
Total params: 407,050
Trainable params: 407,050
Non-trainable params: 0
_________________________________________________________________
Epoch 1/5
1875/1875 [==============================] - 3s 1ms/step - loss: 0.2001 - accuracy: 0.9414
Epoch 2/5
1875/1875 [==============================] - 3s 1ms/step - loss: 0.0796 - accuracy: 0.9760
Epoch 3/5
1875/1875 [==============================] - 3s 1ms/step - loss: 0.0525 - accuracy: 0.9833
Epoch 4/5
1875/1875 [==============================] - 3s 1ms/step - loss: 0.0357 - accuracy: 0.9886
Epoch 5/5
1875/1875 [==============================] - 3s 1ms/step - loss: 0.0283 - accuracy: 0.9905
313/313 [==============================] - 0s 1ms/step - loss: 0.0695 - accuracy: 0.9793
WARNING:tensorflow:From /home/tl/miniconda3/envs/tf2gpu/lib/python3.7/site-packages/tensorflow/python/ops/resource_variable_ops.py:1817: calling BaseResourceVariable.__init__ (from tensorflow.python.ops.resource_variable_ops) with constraint is deprecated and will be removed in a future version.
Instructions for updating:
If using Keras pass *_constraint arguments to layers.
INFO:tensorflow:Assets written to: models/lenet5/assets
2、分别用h5和saved model做推理
从测试集中随意选一张图片，保存成png格式，用来以下的测试
2.1 h5模型推理
#生成保存结果的路径
if not os.path.exists('out_results'):os.mkdir('out_results')
import tensorflow as tf
import numpy as np
from PIL import Image
import matplotlib.pyplot as pltdef h5_infer(model_path=None,img=None):model = tf.keras.models.load_model(model_path)result= model.predict(img)return resultdef main():model_path = "models/lenet5.h5"img = Image.open('5.png')plt.imshow(img)img= np.array(img).reshape(1,28,28,1)/255.0result = h5_infer(model_path,img)print("result.shape",result.shape)print(result)np.save('out_results/h5_result.npy',result)#保存结果，方便以后比较
main()
result.shape (1, 10)
[[2.6573987e-14 1.6878349e-10 1.6627098e-07 1.4200980e-04 1.8240679e-239.9985778e-01 4.3281314e-16 4.1571907e-12 2.4773154e-12 3.5959408e-10]]
2.2 saved_model进行推理
import tensorflow as tf
import numpy as np
from PIL import Image
import matplotlib.pyplot as pltdef savedmodel_infer(model_path=None,img=None):model = tf.keras.models.load_model(model_path)result= model.predict(img)return resultdef main():model_path = "models/lenet5"img = Image.open('5.png')plt.imshow(img)img= np.array(img).reshape(1,28,28,1)/255.0result = savedmodel_infer(model_path,img)print("result.shape",result.shape)print(result)np.save('out_results/savedmodel_result.npy',result)#保存结果，方便以后比较
main()
result.shape (1, 10)
[[2.6573987e-14 1.6878349e-10 1.6627098e-07 1.4200980e-04 1.8240679e-239.9985778e-01 4.3281314e-16 4.1571907e-12 2.4773154e-12 3.5959408e-10]]
还有另一种推理方式
import tensorflow as tf
import numpy as np
from PIL import Image
import matplotlib.pyplot as plt
from tensorflow.python.saved_model import tag_constantsdef savedmodel_infer(model_path=None,img=None):saved_model_loaded = tf.saved_model.load(model_path, tags=[tag_constants.SERVING])model = saved_model_loaded.signatures['serving_default']result= model(img)return resultdef main():model_path = "models/lenet5"img = Image.open('5.png')plt.imshow(img)img= tf.constant(np.array(img).astype(np.float32).reshape(1,28,28,1)/255.0)result = savedmodel_infer(model_path,img)print(result) #获取的结果是个字典for k,v in result.items(): #key 是layername value=v.numpy()print('k=%s with value:'% k,value)np.save('out_results/savedmodelanother_result.npy',value)#保存结果，方便以后比较
main()
{'dense_1': <tf.Tensor: shape=(1, 10), dtype=float32, numpy=
array([[2.6573987e-14, 1.6878349e-10, 1.6627098e-07, 1.4200980e-04,1.8240679e-23, 9.9985778e-01, 4.3281314e-16, 4.1571907e-12,2.4773154e-12, 3.5959408e-10]], dtype=float32)>}
k=dense_1 with value: [[2.6573987e-14 1.6878349e-10 1.6627098e-07 1.4200980e-04 1.8240679e-239.9985778e-01 4.3281314e-16 4.1571907e-12 2.4773154e-12 3.5959408e-10]]
3、模型转换pb
有两个思路，分别是h5转pb、saved model转pb
3.1 h5转pb
# use tf2.x(tf1.14 also ok) to convert the model hdf5 trained by tf2.x to pb
import tensorflow as tfdef freeze_session(model_path=None,clear_devices=True):tf.compat.v1.reset_default_graph()session=tf.compat.v1.keras.backend.get_session()graph = session.graphwith graph.as_default():model = tf.keras.models.load_model(model_path)output_names = [out.op.name for out in model.outputs]print("output_names",output_names)input_names =[innode.op.name for innode in model.inputs]print("input_names",input_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 = tf.compat.v1.graph_util.convert_variables_to_constants(session, input_graph_def,output_names)outgraph = tf.compat.v1.graph_util.remove_training_nodes(frozen_graph)#去掉与推理无关的内容print("##################################################################")for node in outgraph.node:print('node:', node.name)print("length of node",len(outgraph.node))tf.io.write_graph(frozen_graph, "./models", "lenet5_h5.pb", as_text=False)return outgraphdef main(): freeze_session("models/lenet5.h5",True)main()
output_names ['dense_1/Softmax']
input_names ['flatten_input']
node: flatten_input
node: flatten/Const
node: flatten/Reshape
node: dense/kernel/Initializer/random_uniform/shape
node: dense/kernel/Initializer/random_uniform/min
node: dense/kernel/Initializer/random_uniform/max
node: dense/kernel/Initializer/random_uniform/RandomUniform
node: dense/kernel/Initializer/random_uniform/sub
node: dense/kernel/Initializer/random_uniform/mul
node: dense/kernel/Initializer/random_uniform
node: dense/kernel
node: dense/kernel/IsInitialized/VarIsInitializedOp
node: dense/kernel/Assign
node: dense/kernel/Read/ReadVariableOp
node: dense/bias/Initializer/zeros
node: dense/bias
node: dense/bias/IsInitialized/VarIsInitializedOp
node: dense/bias/Assign
node: dense/bias/Read/ReadVariableOp
node: dense/MatMul/ReadVariableOp
node: dense/MatMul
node: dense/BiasAdd/ReadVariableOp
node: dense/BiasAdd
node: dense/Relu
node: dense_1/kernel/Initializer/random_uniform/shape
node: dense_1/kernel/Initializer/random_uniform/min
node: dense_1/kernel/Initializer/random_uniform/max
node: dense_1/kernel/Initializer/random_uniform/RandomUniform
node: dense_1/kernel/Initializer/random_uniform/sub
node: dense_1/kernel/Initializer/random_uniform/mul
node: dense_1/kernel/Initializer/random_uniform
node: dense_1/kernel
node: dense_1/kernel/IsInitialized/VarIsInitializedOp
node: dense_1/kernel/Assign
node: dense_1/kernel/Read/ReadVariableOp
node: dense_1/bias/Initializer/zeros
node: dense_1/bias
node: dense_1/bias/IsInitialized/VarIsInitializedOp
node: dense_1/bias/Assign
node: dense_1/bias/Read/ReadVariableOp
node: dense_1/MatMul/ReadVariableOp
node: dense_1/MatMul
node: dense_1/BiasAdd/ReadVariableOp
node: dense_1/BiasAdd
node: dense_1/Softmax
node: Placeholder
node: AssignVariableOp
node: ReadVariableOp
node: Placeholder_1
node: AssignVariableOp_1
node: ReadVariableOp_1
node: Placeholder_2
node: AssignVariableOp_2
node: ReadVariableOp_2
node: Placeholder_3
node: AssignVariableOp_3
node: ReadVariableOp_3
node: VarIsInitializedOp
node: VarIsInitializedOp_1
node: VarIsInitializedOp_2
node: VarIsInitializedOp_3
node: init
node: dense_1_target
node: total/Initializer/zeros
node: total
node: total/IsInitialized/VarIsInitializedOp
node: total/Assign
node: total/Read/ReadVariableOp
node: count/Initializer/zeros
node: count
node: count/IsInitialized/VarIsInitializedOp
node: count/Assign
node: count/Read/ReadVariableOp
node: metrics/accuracy/Squeeze
node: metrics/accuracy/ArgMax/dimension
node: metrics/accuracy/ArgMax
node: metrics/accuracy/Cast
node: metrics/accuracy/Equal
node: metrics/accuracy/Cast_1
node: metrics/accuracy/Const
node: metrics/accuracy/Sum
node: metrics/accuracy/AssignAddVariableOp
node: metrics/accuracy/ReadVariableOp
node: metrics/accuracy/Size
node: metrics/accuracy/Cast_2
node: metrics/accuracy/AssignAddVariableOp_1
node: metrics/accuracy/ReadVariableOp_1
node: metrics/accuracy/div_no_nan/ReadVariableOp
node: metrics/accuracy/div_no_nan/ReadVariableOp_1
node: metrics/accuracy/div_no_nan
node: metrics/accuracy/Identity
node: loss/dense_1_loss/Cast
node: loss/dense_1_loss/Shape
node: loss/dense_1_loss/Reshape/shape
node: loss/dense_1_loss/Reshape
node: loss/dense_1_loss/strided_slice/stack
node: loss/dense_1_loss/strided_slice/stack_1
node: loss/dense_1_loss/strided_slice/stack_2
node: loss/dense_1_loss/strided_slice
node: loss/dense_1_loss/Reshape_1/shape/0
node: loss/dense_1_loss/Reshape_1/shape
node: loss/dense_1_loss/Reshape_1
node: loss/dense_1_loss/SparseSoftmaxCrossEntropyWithLogits/Shape
node: loss/dense_1_loss/SparseSoftmaxCrossEntropyWithLogits/SparseSoftmaxCrossEntropyWithLogits
node: loss/dense_1_loss/weighted_loss/Cast/x
node: loss/dense_1_loss/weighted_loss/Mul
node: loss/dense_1_loss/Const
node: loss/dense_1_loss/Sum
node: loss/dense_1_loss/num_elements
node: loss/dense_1_loss/num_elements/Cast
node: loss/dense_1_loss/Const_1
node: loss/dense_1_loss/Sum_1
node: loss/dense_1_loss/value
node: loss/mul/x
node: loss/mul
node: iter/Initializer/zeros
node: iter
node: iter/IsInitialized/VarIsInitializedOp
node: iter/Assign
node: iter/Read/ReadVariableOp
node: beta_1/Initializer/initial_value
node: beta_1
node: beta_1/IsInitialized/VarIsInitializedOp
node: beta_1/Assign
node: beta_1/Read/ReadVariableOp
node: beta_2/Initializer/initial_value
node: beta_2
node: beta_2/IsInitialized/VarIsInitializedOp
node: beta_2/Assign
node: beta_2/Read/ReadVariableOp
node: decay/Initializer/initial_value
node: decay
node: decay/IsInitialized/VarIsInitializedOp
node: decay/Assign
node: decay/Read/ReadVariableOp
node: learning_rate/Initializer/initial_value
node: learning_rate
node: learning_rate/IsInitialized/VarIsInitializedOp
node: learning_rate/Assign
node: learning_rate/Read/ReadVariableOp
node: dense/kernel/m/Initializer/zeros/shape_as_tensor
node: dense/kernel/m/Initializer/zeros/Const
node: dense/kernel/m/Initializer/zeros
node: dense/kernel/m
node: dense/kernel/m/IsInitialized/VarIsInitializedOp
node: dense/kernel/m/Assign
node: dense/kernel/m/Read/ReadVariableOp
node: dense/bias/m/Initializer/zeros
node: dense/bias/m
node: dense/bias/m/IsInitialized/VarIsInitializedOp
node: dense/bias/m/Assign
node: dense/bias/m/Read/ReadVariableOp
node: dense_1/kernel/m/Initializer/zeros/shape_as_tensor
node: dense_1/kernel/m/Initializer/zeros/Const
node: dense_1/kernel/m/Initializer/zeros
node: dense_1/kernel/m
node: dense_1/kernel/m/IsInitialized/VarIsInitializedOp
node: dense_1/kernel/m/Assign
node: dense_1/kernel/m/Read/ReadVariableOp
node: dense_1/bias/m/Initializer/zeros
node: dense_1/bias/m
node: dense_1/bias/m/IsInitialized/VarIsInitializedOp
node: dense_1/bias/m/Assign
node: dense_1/bias/m/Read/ReadVariableOp
node: dense/kernel/v/Initializer/zeros/shape_as_tensor
node: dense/kernel/v/Initializer/zeros/Const
node: dense/kernel/v/Initializer/zeros
node: dense/kernel/v
node: dense/kernel/v/IsInitialized/VarIsInitializedOp
node: dense/kernel/v/Assign
node: dense/kernel/v/Read/ReadVariableOp
node: dense/bias/v/Initializer/zeros
node: dense/bias/v
node: dense/bias/v/IsInitialized/VarIsInitializedOp
node: dense/bias/v/Assign
node: dense/bias/v/Read/ReadVariableOp
node: dense_1/kernel/v/Initializer/zeros/shape_as_tensor
node: dense_1/kernel/v/Initializer/zeros/Const
node: dense_1/kernel/v/Initializer/zeros
node: dense_1/kernel/v
node: dense_1/kernel/v/IsInitialized/VarIsInitializedOp
node: dense_1/kernel/v/Assign
node: dense_1/kernel/v/Read/ReadVariableOp
node: dense_1/bias/v/Initializer/zeros
node: dense_1/bias/v
node: dense_1/bias/v/IsInitialized/VarIsInitializedOp
node: dense_1/bias/v/Assign
node: dense_1/bias/v/Read/ReadVariableOp
node: VarIsInitializedOp_4
node: VarIsInitializedOp_5
node: VarIsInitializedOp_6
node: VarIsInitializedOp_7
node: VarIsInitializedOp_8
node: VarIsInitializedOp_9
node: VarIsInitializedOp_10
node: VarIsInitializedOp_11
node: VarIsInitializedOp_12
node: VarIsInitializedOp_13
node: VarIsInitializedOp_14
node: VarIsInitializedOp_15
node: VarIsInitializedOp_16
node: VarIsInitializedOp_17
node: VarIsInitializedOp_18
node: init_1
node: Placeholder_4
node: AssignVariableOp_4
node: ReadVariableOp_4
node: Placeholder_5
node: AssignVariableOp_5
node: ReadVariableOp_5
node: Placeholder_6
node: AssignVariableOp_6
node: ReadVariableOp_6
node: Placeholder_7
node: AssignVariableOp_7
node: ReadVariableOp_7
node: Placeholder_8
node: AssignVariableOp_8
node: ReadVariableOp_8
node: Placeholder_9
node: AssignVariableOp_9
node: ReadVariableOp_9
node: Placeholder_10
node: AssignVariableOp_10
node: ReadVariableOp_10
node: Placeholder_11
node: AssignVariableOp_11
node: ReadVariableOp_11
node: Placeholder_12
node: AssignVariableOp_12
node: ReadVariableOp_12
len node1 231
INFO:tensorflow:Froze 4 variables.
INFO:tensorflow:Converted 4 variables to const ops.
##################################################################
node: flatten_input
node: flatten/Const
node: flatten/Reshape
node: dense/kernel
node: dense/bias
node: dense/MatMul
node: dense/BiasAdd
node: dense/Relu
node: dense_1/kernel
node: dense_1/bias
node: dense_1/MatMul
node: dense_1/BiasAdd
node: dense_1/Softmax
length of node 13
3.2 saved model转pb
import tensorflow as tfdef freeze_session(model_path=None,clear_devices=True):tf.compat.v1.reset_default_graph()session=tf.compat.v1.keras.backend.get_session()graph = session.graphwith graph.as_default():model = tf.keras.models.load_model(model_path)output_names = [out.op.name for out in model.outputs]print("output_names",output_names)input_names =[innode.op.name for innode in model.inputs]print("input_names",input_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 = tf.compat.v1.graph_util.convert_variables_to_constants(session, input_graph_def,output_names)outgraph = tf.compat.v1.graph_util.remove_training_nodes(frozen_graph)#去掉与推理无关的内容print("##################################################################")for node in outgraph.node:print('node:', node.name)print("length of node",len(outgraph.node))tf.io.write_graph(frozen_graph, "./models", "lenet5_savedmodel.pb", as_text=False)def main(): freeze_session("models/lenet5",True)main()
output_names ['dense_1/Softmax']
input_names ['input_1']
node: kernel/Initializer/random_uniform/shape
node: kernel/Initializer/random_uniform/min
node: kernel/Initializer/random_uniform/max
node: kernel/Initializer/random_uniform/RandomUniform
node: kernel/Initializer/random_uniform/sub
node: kernel/Initializer/random_uniform/mul
node: kernel/Initializer/random_uniform
node: kernel
node: kernel/IsInitialized/VarIsInitializedOp
node: kernel/Assign
node: kernel/Read/ReadVariableOp
node: bias/Initializer/zeros
node: bias
node: bias/IsInitialized/VarIsInitializedOp
node: bias/Assign
node: bias/Read/ReadVariableOp
node: kernel_1/Initializer/random_uniform/shape
node: kernel_1/Initializer/random_uniform/min
node: kernel_1/Initializer/random_uniform/max
node: kernel_1/Initializer/random_uniform/RandomUniform
node: kernel_1/Initializer/random_uniform/sub
node: kernel_1/Initializer/random_uniform/mul
node: kernel_1/Initializer/random_uniform
node: kernel_1
node: kernel_1/IsInitialized/VarIsInitializedOp
node: kernel_1/Assign
node: kernel_1/Read/ReadVariableOp
node: bias_1/Initializer/zeros
node: bias_1
node: bias_1/IsInitialized/VarIsInitializedOp
node: bias_1/Assign
node: bias_1/Read/ReadVariableOp
node: total/Initializer/zeros
node: total
node: total/IsInitialized/VarIsInitializedOp
node: total/Assign
node: total/Read/ReadVariableOp
node: count/Initializer/zeros
node: count
node: count/IsInitialized/VarIsInitializedOp
node: count/Assign
node: count/Read/ReadVariableOp
node: total_1/Initializer/zeros
node: total_1
node: total_1/IsInitialized/VarIsInitializedOp
node: total_1/Assign
node: total_1/Read/ReadVariableOp
node: count_1/Initializer/zeros
node: count_1
node: count_1/IsInitialized/VarIsInitializedOp
node: count_1/Assign
node: count_1/Read/ReadVariableOp
node: Adam/iter
node: Adam/iter/Read/ReadVariableOp
node: Adam/beta_1
node: Adam/beta_1/Read/ReadVariableOp
node: Adam/beta_2
node: Adam/beta_2/Read/ReadVariableOp
node: Adam/decay
node: Adam/decay/Read/ReadVariableOp
node: Adam/learning_rate
node: Adam/learning_rate/Read/ReadVariableOp
node: dense/kernel/m/Initializer/zeros/shape_as_tensor
node: dense/kernel/m/Initializer/zeros/Const
node: dense/kernel/m/Initializer/zeros
node: dense/kernel/m
node: dense/kernel/m/IsInitialized/VarIsInitializedOp
node: dense/kernel/m/Assign
node: dense/kernel/m/Read/ReadVariableOp
node: dense/bias/m/Initializer/zeros
node: dense/bias/m
node: dense/bias/m/IsInitialized/VarIsInitializedOp
node: dense/bias/m/Assign
node: dense/bias/m/Read/ReadVariableOp
node: dense_1/kernel/m/Initializer/zeros/shape_as_tensor
node: dense_1/kernel/m/Initializer/zeros/Const
node: dense_1/kernel/m/Initializer/zeros
node: dense_1/kernel/m
node: dense_1/kernel/m/IsInitialized/VarIsInitializedOp
node: dense_1/kernel/m/Assign
node: dense_1/kernel/m/Read/ReadVariableOp
node: dense_1/bias/m/Initializer/zeros
node: dense_1/bias/m
node: dense_1/bias/m/IsInitialized/VarIsInitializedOp
node: dense_1/bias/m/Assign
node: dense_1/bias/m/Read/ReadVariableOp
node: dense/kernel/v/Initializer/zeros/shape_as_tensor
node: dense/kernel/v/Initializer/zeros/Const
node: dense/kernel/v/Initializer/zeros
node: dense/kernel/v
node: dense/kernel/v/IsInitialized/VarIsInitializedOp
node: dense/kernel/v/Assign
node: dense/kernel/v/Read/ReadVariableOp
node: dense/bias/v/Initializer/zeros
node: dense/bias/v
node: dense/bias/v/IsInitialized/VarIsInitializedOp
node: dense/bias/v/Assign
node: dense/bias/v/Read/ReadVariableOp
node: dense_1/kernel/v/Initializer/zeros/shape_as_tensor
node: dense_1/kernel/v/Initializer/zeros/Const
node: dense_1/kernel/v/Initializer/zeros
node: dense_1/kernel/v
node: dense_1/kernel/v/IsInitialized/VarIsInitializedOp
node: dense_1/kernel/v/Assign
node: dense_1/kernel/v/Read/ReadVariableOp
node: dense_1/bias/v/Initializer/zeros
node: dense_1/bias/v
node: dense_1/bias/v/IsInitialized/VarIsInitializedOp
node: dense_1/bias/v/Assign
node: dense_1/bias/v/Read/ReadVariableOp
node: input_1
node: flatten/Const
node: flatten/Reshape
node: dense/MatMul/ReadVariableOp
node: dense/MatMul
node: dense/BiasAdd/ReadVariableOp
node: dense/BiasAdd
node: dense/Relu
node: dense_1/MatMul/ReadVariableOp
node: dense_1/MatMul
node: dense_1/BiasAdd/ReadVariableOp
node: dense_1/BiasAdd
node: dense_1/Softmax
node: Const
node: RestoreV2/tensor_names
node: RestoreV2/shape_and_slices
node: RestoreV2
node: Identity
node: AssignVariableOp
node: RestoreV2_1/tensor_names
node: RestoreV2_1/shape_and_slices
node: RestoreV2_1
node: Identity_1
node: AssignVariableOp_1
node: RestoreV2_2/tensor_names
node: RestoreV2_2/shape_and_slices
node: RestoreV2_2
node: Identity_2
node: AssignVariableOp_2
node: RestoreV2_3/tensor_names
node: RestoreV2_3/shape_and_slices
node: RestoreV2_3
node: Identity_3
node: AssignVariableOp_3
node: RestoreV2_4/tensor_names
node: RestoreV2_4/shape_and_slices
node: RestoreV2_4
node: Identity_4
node: AssignVariableOp_4
node: RestoreV2_5/tensor_names
node: RestoreV2_5/shape_and_slices
node: RestoreV2_5
node: Identity_5
node: AssignVariableOp_5
node: RestoreV2_6/tensor_names
node: RestoreV2_6/shape_and_slices
node: RestoreV2_6
node: Identity_6
node: AssignVariableOp_6
node: RestoreV2_7/tensor_names
node: RestoreV2_7/shape_and_slices
node: RestoreV2_7
node: Identity_7
node: AssignVariableOp_7
node: RestoreV2_8/tensor_names
node: RestoreV2_8/shape_and_slices
node: RestoreV2_8
node: Identity_8
node: AssignVariableOp_8
node: Identity_9
node: AssignVariableOp_9
node: Identity_10
node: AssignVariableOp_10
node: Identity_11
node: AssignVariableOp_11
node: Identity_12
node: AssignVariableOp_12
node: Identity_13
node: AssignVariableOp_13
node: Identity_14
node: AssignVariableOp_14
node: Identity_15
node: AssignVariableOp_15
node: Identity_16
node: AssignVariableOp_16
node: Identity_17
node: AssignVariableOp_17
node: Identity_18
node: AssignVariableOp_18
node: Identity_19
node: AssignVariableOp_19
node: Identity_20
node: AssignVariableOp_20
node: dense_1_target
node: total_2/Initializer/zeros
node: total_2
node: total_2/IsInitialized/VarIsInitializedOp
node: total_2/Assign
node: total_2/Read/ReadVariableOp
node: count_2/Initializer/zeros
node: count_2
node: count_2/IsInitialized/VarIsInitializedOp
node: count_2/Assign
node: count_2/Read/ReadVariableOp
node: metrics/accuracy/Squeeze
node: metrics/accuracy/ArgMax/dimension
node: metrics/accuracy/ArgMax
node: metrics/accuracy/Cast
node: metrics/accuracy/Equal
node: metrics/accuracy/Cast_1
node: metrics/accuracy/Const
node: metrics/accuracy/Sum
node: metrics/accuracy/AssignAddVariableOp
node: metrics/accuracy/ReadVariableOp
node: metrics/accuracy/Size
node: metrics/accuracy/Cast_2
node: metrics/accuracy/AssignAddVariableOp_1
node: metrics/accuracy/ReadVariableOp_1
node: metrics/accuracy/div_no_nan/ReadVariableOp
node: metrics/accuracy/div_no_nan/ReadVariableOp_1
node: metrics/accuracy/div_no_nan
node: metrics/accuracy/Identity
node: loss/dense_1_loss/Cast
node: loss/dense_1_loss/Shape
node: loss/dense_1_loss/Reshape/shape
node: loss/dense_1_loss/Reshape
node: loss/dense_1_loss/strided_slice/stack
node: loss/dense_1_loss/strided_slice/stack_1
node: loss/dense_1_loss/strided_slice/stack_2
node: loss/dense_1_loss/strided_slice
node: loss/dense_1_loss/Reshape_1/shape/0
node: loss/dense_1_loss/Reshape_1/shape
node: loss/dense_1_loss/Reshape_1
node: loss/dense_1_loss/SparseSoftmaxCrossEntropyWithLogits/Shape
node: loss/dense_1_loss/SparseSoftmaxCrossEntropyWithLogits/SparseSoftmaxCrossEntropyWithLogits
node: loss/dense_1_loss/weighted_loss/Cast/x
node: loss/dense_1_loss/weighted_loss/Mul
node: loss/dense_1_loss/Const
node: loss/dense_1_loss/Sum
node: loss/dense_1_loss/num_elements
node: loss/dense_1_loss/num_elements/Cast
node: loss/dense_1_loss/Const_1
node: loss/dense_1_loss/Sum_1
node: loss/dense_1_loss/value
node: loss/mul/x
node: loss/mul
node: VarIsInitializedOp
node: VarIsInitializedOp_1
node: VarIsInitializedOp_2
node: VarIsInitializedOp_3
node: VarIsInitializedOp_4
node: VarIsInitializedOp_5
node: VarIsInitializedOp_6
node: VarIsInitializedOp_7
node: VarIsInitializedOp_8
node: VarIsInitializedOp_9
node: VarIsInitializedOp_10
node: VarIsInitializedOp_11
node: VarIsInitializedOp_12
node: VarIsInitializedOp_13
node: VarIsInitializedOp_14
node: VarIsInitializedOp_15
node: VarIsInitializedOp_16
node: VarIsInitializedOp_17
node: init
len node1 265
INFO:tensorflow:Froze 4 variables.
INFO:tensorflow:Converted 4 variables to const ops.
##################################################################
node: kernel
node: bias
node: kernel_1
node: bias_1
node: input_1
node: flatten/Const
node: flatten/Reshape
node: dense/MatMul
node: dense/BiasAdd
node: dense/Relu
node: dense_1/MatMul
node: dense_1/BiasAdd
node: dense_1/Softmax
length of node 13
4、调用pb模型进行推理
同样，有两个可以进行推理的pb,分别来自不同的转换源
4.1 h5 pb的推理
输入输出结点名称要参考上文中转pb过程中确定的input name 和output name
#use tf1.14 and tf2.x with pb to load the data are both ok
import tensorflow as tf
from PIL import Image
import numpy as npdef load_graph(file_path):with tf.io.gfile.GFile(file_path,'rb') as f:graph_def = tf.compat.v1.GraphDef()graph_def.ParseFromString(f.read())with tf.compat.v1.Graph().as_default() as graph:tf.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
def main():file_path='./models/lenet5_h5.pb'img= np.array(Image.open('5.png')).reshape(1,28,28,1)/255.0graph,graph_nodes = load_graph(file_path)print("num nodes",len(graph_nodes))for node in graph_nodes:print('node:', node.name)input_node = graph.get_tensor_by_name('flatten_input:0')print("input_node.shape:",input_node.shape)output = graph.get_tensor_by_name('dense_1/Softmax:0') config = tf.compat.v1.ConfigProto()config.gpu_options.allow_growth = Truewith tf.compat.v1.Session(graph=graph,config=config) as sess:logits = sess.run(output, feed_dict = {input_node:img})print("logits:",logits)np.save('out_results/h5pb_result.npy',logits)
main()WARNING:tensorflow:From <ipython-input-16-1008fb2da38b>:11: calling import_graph_def (from tensorflow.python.framework.importer) with op_dict is deprecated and will be removed in a future version.
Instructions for updating:
Please file an issue at https://github.com/tensorflow/tensorflow/issues if you depend on this feature.
num nodes 17
node: flatten_input
node: flatten/Const
node: flatten/Reshape
node: dense/kernel
node: dense/bias
node: dense/MatMul/ReadVariableOp
node: dense/MatMul
node: dense/BiasAdd/ReadVariableOp
node: dense/BiasAdd
node: dense/Relu
node: dense_1/kernel
node: dense_1/bias
node: dense_1/MatMul/ReadVariableOp
node: dense_1/MatMul
node: dense_1/BiasAdd/ReadVariableOp
node: dense_1/BiasAdd
node: dense_1/Softmax
input_node.shape: (None, 28, 28, 1)
logits: [[2.6573987e-14 1.6878349e-10 1.6627098e-07 1.4200980e-04 1.8240679e-239.9985778e-01 4.3281314e-16 4.1571907e-12 2.4773154e-12 3.5959408e-10]]
4.2 saved pb的推理
#use tf1.14 and tf2.x with pb to load the data are both ok
import tensorflow as tf
from PIL import Image
import numpy as npdef load_graph(file_path):with tf.io.gfile.GFile(file_path,'rb') as f:graph_def = tf.compat.v1.GraphDef()graph_def.ParseFromString(f.read())with tf.compat.v1.Graph().as_default() as graph:tf.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
def main():file_path='./models/lenet5_savedmodel.pb'img= np.array(Image.open('5.png')).reshape(1,28,28,1)/255.0graph,graph_nodes = load_graph(file_path)print("num nodes",len(graph_nodes))for node in graph_nodes:print('node:', node.name)input_node = graph.get_tensor_by_name('input_1:0')print("input_node.shape:",input_node.shape)output = graph.get_tensor_by_name('dense_1/Softmax:0') with tf.compat.v1.Session(graph=graph) as sess:logits = sess.run(output, feed_dict = {input_node:img})print("logits:",logits)np.save('out_results/savedmodelpb_result.npy',logits)
main()num nodes 17
node: kernel
node: bias
node: kernel_1
node: bias_1
node: input_1
node: flatten/Const
node: flatten/Reshape
node: dense/MatMul/ReadVariableOp
node: dense/MatMul
node: dense/BiasAdd/ReadVariableOp
node: dense/BiasAdd
node: dense/Relu
node: dense_1/MatMul/ReadVariableOp
node: dense_1/MatMul
node: dense_1/BiasAdd/ReadVariableOp
node: dense_1/BiasAdd
node: dense_1/Softmax
input_node.shape: (None, 28, 28, 1)
logits: [[2.6573987e-14 1.6878349e-10 1.6627098e-07 1.4200980e-04 1.8240679e-239.9985778e-01 4.3281314e-16 4.1571907e-12 2.4773154e-12 3.5959408e-10]]
5、h5和saved 尝试互转
- 从以上可以看出，h5 可以转pb,saved model也可以转pb,并且，都可以进行推理，并且结果正确
- 还未看到h5模型是否能和saved model互转下面将试一下
5.1 h5转saved
def main():file_path='./models/lenet5.h5'model = tf.keras.models.load_model(file_path)model.save('./models/h5tosaved',save_format='tf')main()
INFO:tensorflow:Assets written to: ./models/h5tosaved/assets
看起来是转成功了，那么用这个转换后的模型进行推理来难证正确性
5.1.1 h5 转saved后的模型推理
import tensorflow as tf
import numpy as np
from PIL import Image
import matplotlib.pyplot as pltdef savedmodel_infer(model_path=None,img=None):model = tf.keras.models.load_model(model_path)result= model.predict(img)return resultdef main():model_path = "models/h5tosaved"img = Image.open('5.png')plt.imshow(img)img= np.array(img).reshape(1,28,28,1)/255.0result = savedmodel_infer(model_path,img)print("result.shape",result.shape)print(result)np.save('out_results/h5tosavedmodel_result.npy',result)#保存结果，方便以后比较
main()
result.shape (1, 10)
[[2.6573987e-14 1.6878349e-10 1.6627098e-07 1.4200980e-04 1.8240679e-239.9985778e-01 4.3281314e-16 4.1571907e-12 2.4773154e-12 3.5959408e-10]]
另一种方法
import tensorflow as tf
import numpy as np
from PIL import Image
import matplotlib.pyplot as plt
from tensorflow.python.saved_model import tag_constantsdef savedmodel_infer(model_path=None,img=None):saved_model_loaded = tf.saved_model.load(model_path, tags=[tag_constants.SERVING])model = saved_model_loaded.signatures['serving_default']result= model(img)return resultdef main():model_path = "models/h5tosaved"img = Image.open('5.png')plt.imshow(img)img= tf.constant(np.array(img).astype(np.float32).reshape(1,28,28,1)/255.0)result = savedmodel_infer(model_path,img)print(result) #获取的结果是个字典for k,v in result.items(): #key 是layername value=v.numpy()print('k=%s with value:'% k,value)np.save('out_results/h5tosavedmodelanother_result.npy',value)#保存结果，方便以后比较
main()
{'dense_1': <tf.Tensor: shape=(1, 10), dtype=float32, numpy=
array([[2.6573987e-14, 1.6878349e-10, 1.6627098e-07, 1.4200980e-04,1.8240679e-23, 9.9985778e-01, 4.3281314e-16, 4.1571907e-12,2.4773154e-12, 3.5959408e-10]], dtype=float32)>}
k=dense_1 with value: [[2.6573987e-14 1.6878349e-10 1.6627098e-07 1.4200980e-04 1.8240679e-239.9985778e-01 4.3281314e-16 4.1571907e-12 2.4773154e-12 3.5959408e-10]]
5.1.2 h5转saved后的模型转pb
import tensorflow as tfdef freeze_session(model_path=None,clear_devices=True):tf.compat.v1.reset_default_graph()session=tf.compat.v1.keras.backend.get_session()graph = session.graphwith graph.as_default():model = tf.keras.models.load_model(model_path)output_names = [out.op.name for out in model.outputs]print("output_names",output_names)input_names =[innode.op.name for innode in model.inputs]print("input_names",input_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 = tf.compat.v1.graph_util.convert_variables_to_constants(session, input_graph_def,output_names)outgraph = tf.compat.v1.graph_util.remove_training_nodes(frozen_graph)#去掉与推理无关的内容print("##################################################################")for node in outgraph.node:print('node:', node.name)print("length of node",len(outgraph.node))tf.io.write_graph(frozen_graph, "./models", "h5tosaved_savedmodel.pb", as_text=False)def main(): freeze_session("models/h5tosaved",True)main()
output_names ['dense_1/Softmax']
input_names ['input_1']
node: kernel/Initializer/random_uniform/shape
node: kernel/Initializer/random_uniform/min
node: kernel/Initializer/random_uniform/max
node: kernel/Initializer/random_uniform/RandomUniform
node: kernel/Initializer/random_uniform/sub
node: kernel/Initializer/random_uniform/mul
node: kernel/Initializer/random_uniform
node: kernel
node: kernel/IsInitialized/VarIsInitializedOp
node: kernel/Assign
node: kernel/Read/ReadVariableOp
node: bias/Initializer/zeros
node: bias
node: bias/IsInitialized/VarIsInitializedOp
node: bias/Assign
node: bias/Read/ReadVariableOp
node: kernel_1/Initializer/random_uniform/shape
node: kernel_1/Initializer/random_uniform/min
node: kernel_1/Initializer/random_uniform/max
node: kernel_1/Initializer/random_uniform/RandomUniform
node: kernel_1/Initializer/random_uniform/sub
node: kernel_1/Initializer/random_uniform/mul
node: kernel_1/Initializer/random_uniform
node: kernel_1
node: kernel_1/IsInitialized/VarIsInitializedOp
node: kernel_1/Assign
node: kernel_1/Read/ReadVariableOp
node: bias_1/Initializer/zeros
node: bias_1
node: bias_1/IsInitialized/VarIsInitializedOp
node: bias_1/Assign
node: bias_1/Read/ReadVariableOp
node: iter
node: iter/Read/ReadVariableOp
node: beta_1
node: beta_1/Read/ReadVariableOp
node: beta_2
node: beta_2/Read/ReadVariableOp
node: decay
node: decay/Read/ReadVariableOp
node: learning_rate
node: learning_rate/Read/ReadVariableOp
node: dense_6/kernel/m/Initializer/zeros/shape_as_tensor
node: dense_6/kernel/m/Initializer/zeros/Const
node: dense_6/kernel/m/Initializer/zeros
node: dense_6/kernel/m
node: dense_6/kernel/m/IsInitialized/VarIsInitializedOp
node: dense_6/kernel/m/Assign
node: dense_6/kernel/m/Read/ReadVariableOp
node: dense_6/bias/m/Initializer/zeros
node: dense_6/bias/m
node: dense_6/bias/m/IsInitialized/VarIsInitializedOp
node: dense_6/bias/m/Assign
node: dense_6/bias/m/Read/ReadVariableOp
node: dense_1_5/kernel/m/Initializer/zeros/shape_as_tensor
node: dense_1_5/kernel/m/Initializer/zeros/Const
node: dense_1_5/kernel/m/Initializer/zeros
node: dense_1_5/kernel/m
node: dense_1_5/kernel/m/IsInitialized/VarIsInitializedOp
node: dense_1_5/kernel/m/Assign
node: dense_1_5/kernel/m/Read/ReadVariableOp
node: dense_1_5/bias/m/Initializer/zeros
node: dense_1_5/bias/m
node: dense_1_5/bias/m/IsInitialized/VarIsInitializedOp
node: dense_1_5/bias/m/Assign
node: dense_1_5/bias/m/Read/ReadVariableOp
node: dense_6/kernel/v/Initializer/zeros/shape_as_tensor
node: dense_6/kernel/v/Initializer/zeros/Const
node: dense_6/kernel/v/Initializer/zeros
node: dense_6/kernel/v
node: dense_6/kernel/v/IsInitialized/VarIsInitializedOp
node: dense_6/kernel/v/Assign
node: dense_6/kernel/v/Read/ReadVariableOp
node: dense_6/bias/v/Initializer/zeros
node: dense_6/bias/v
node: dense_6/bias/v/IsInitialized/VarIsInitializedOp
node: dense_6/bias/v/Assign
node: dense_6/bias/v/Read/ReadVariableOp
node: dense_1_5/kernel/v/Initializer/zeros/shape_as_tensor
node: dense_1_5/kernel/v/Initializer/zeros/Const
node: dense_1_5/kernel/v/Initializer/zeros
node: dense_1_5/kernel/v
node: dense_1_5/kernel/v/IsInitialized/VarIsInitializedOp
node: dense_1_5/kernel/v/Assign
node: dense_1_5/kernel/v/Read/ReadVariableOp
node: dense_1_5/bias/v/Initializer/zeros
node: dense_1_5/bias/v
node: dense_1_5/bias/v/IsInitialized/VarIsInitializedOp
node: dense_1_5/bias/v/Assign
node: dense_1_5/bias/v/Read/ReadVariableOp
node: input_1
node: flatten/Const
node: flatten/Reshape
node: dense/MatMul/ReadVariableOp
node: dense/MatMul
node: dense/BiasAdd/ReadVariableOp
node: dense/BiasAdd
node: dense/Relu
node: dense_1/MatMul/ReadVariableOp
node: dense_1/MatMul
node: dense_1/BiasAdd/ReadVariableOp
node: dense_1/BiasAdd
node: dense_1/Softmax
node: Const
node: RestoreV2/tensor_names
node: RestoreV2/shape_and_slices
node: RestoreV2
node: Identity
node: AssignVariableOp
node: RestoreV2_1/tensor_names
node: RestoreV2_1/shape_and_slices
node: RestoreV2_1
node: Identity_1
node: AssignVariableOp_1
node: RestoreV2_2/tensor_names
node: RestoreV2_2/shape_and_slices
node: RestoreV2_2
node: Identity_2
node: AssignVariableOp_2
node: RestoreV2_3/tensor_names
node: RestoreV2_3/shape_and_slices
node: RestoreV2_3
node: Identity_3
node: AssignVariableOp_3
node: RestoreV2_4/tensor_names
node: RestoreV2_4/shape_and_slices
node: RestoreV2_4
node: Identity_4
node: AssignVariableOp_4
node: RestoreV2_5/tensor_names
node: RestoreV2_5/shape_and_slices
node: RestoreV2_5
node: Identity_5
node: AssignVariableOp_5
node: RestoreV2_6/tensor_names
node: RestoreV2_6/shape_and_slices
node: RestoreV2_6
node: Identity_6
node: AssignVariableOp_6
node: RestoreV2_7/tensor_names
node: RestoreV2_7/shape_and_slices
node: RestoreV2_7
node: Identity_7
node: AssignVariableOp_7
node: RestoreV2_8/tensor_names
node: RestoreV2_8/shape_and_slices
node: RestoreV2_8
node: Identity_8
node: AssignVariableOp_8
node: Identity_9
node: AssignVariableOp_9
node: Identity_10
node: AssignVariableOp_10
node: Identity_11
node: AssignVariableOp_11
node: Identity_12
node: AssignVariableOp_12
node: Identity_13
node: AssignVariableOp_13
node: Identity_14
node: AssignVariableOp_14
node: Identity_15
node: AssignVariableOp_15
node: Identity_16
node: AssignVariableOp_16
node: dense_1_target
node: total/Initializer/zeros
node: total
node: total/IsInitialized/VarIsInitializedOp
node: total/Assign
node: total/Read/ReadVariableOp
node: count/Initializer/zeros
node: count
node: count/IsInitialized/VarIsInitializedOp
node: count/Assign
node: count/Read/ReadVariableOp
node: metrics/accuracy/Squeeze
node: metrics/accuracy/ArgMax/dimension
node: metrics/accuracy/ArgMax
node: metrics/accuracy/Cast
node: metrics/accuracy/Equal
node: metrics/accuracy/Cast_1
node: metrics/accuracy/Const
node: metrics/accuracy/Sum
node: metrics/accuracy/AssignAddVariableOp
node: metrics/accuracy/ReadVariableOp
node: metrics/accuracy/Size
node: metrics/accuracy/Cast_2
node: metrics/accuracy/AssignAddVariableOp_1
node: metrics/accuracy/ReadVariableOp_1
node: metrics/accuracy/div_no_nan/ReadVariableOp
node: metrics/accuracy/div_no_nan/ReadVariableOp_1
node: metrics/accuracy/div_no_nan
node: metrics/accuracy/Identity
node: loss/dense_1_loss/Cast
node: loss/dense_1_loss/Shape
node: loss/dense_1_loss/Reshape/shape
node: loss/dense_1_loss/Reshape
node: loss/dense_1_loss/strided_slice/stack
node: loss/dense_1_loss/strided_slice/stack_1
node: loss/dense_1_loss/strided_slice/stack_2
node: loss/dense_1_loss/strided_slice
node: loss/dense_1_loss/Reshape_1/shape/0
node: loss/dense_1_loss/Reshape_1/shape
node: loss/dense_1_loss/Reshape_1
node: loss/dense_1_loss/SparseSoftmaxCrossEntropyWithLogits/Shape
node: loss/dense_1_loss/SparseSoftmaxCrossEntropyWithLogits/SparseSoftmaxCrossEntropyWithLogits
node: loss/dense_1_loss/weighted_loss/Cast/x
node: loss/dense_1_loss/weighted_loss/Mul
node: loss/dense_1_loss/Const
node: loss/dense_1_loss/Sum
node: loss/dense_1_loss/num_elements
node: loss/dense_1_loss/num_elements/Cast
node: loss/dense_1_loss/Const_1
node: loss/dense_1_loss/Sum_1
node: loss/dense_1_loss/value
node: loss/mul/x
node: loss/mul
node: VarIsInitializedOp
node: VarIsInitializedOp_1
node: VarIsInitializedOp_2
node: VarIsInitializedOp_3
node: VarIsInitializedOp_4
node: VarIsInitializedOp_5
node: VarIsInitializedOp_6
node: VarIsInitializedOp_7
node: VarIsInitializedOp_8
node: VarIsInitializedOp_9
node: VarIsInitializedOp_10
node: VarIsInitializedOp_11
node: VarIsInitializedOp_12
node: VarIsInitializedOp_13
node: init
len node1 233
INFO:tensorflow:Froze 4 variables.
INFO:tensorflow:Converted 4 variables to const ops.
##################################################################
node: kernel
node: bias
node: kernel_1
node: bias_1
node: input_1
node: flatten/Const
node: flatten/Reshape
node: dense/MatMul
node: dense/BiasAdd
node: dense/Relu
node: dense_1/MatMul
node: dense_1/BiasAdd
node: dense_1/Softmax
length of node 13
5.1.3 h5转saved模型后转pb推理
#use tf1.14 and tf2.x with pb to load the data are both ok
import tensorflow as tf
from PIL import Image
import numpy as npdef load_graph(file_path):with tf.io.gfile.GFile(file_path,'rb') as f:graph_def = tf.compat.v1.GraphDef()graph_def.ParseFromString(f.read())with tf.compat.v1.Graph().as_default() as graph:tf.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
def main():file_path='./models/h5tosaved_savedmodel.pb'img= np.array(Image.open('5.png')).reshape(1,28,28,1)/255.0graph,graph_nodes = load_graph(file_path)print("num nodes",len(graph_nodes))for node in graph_nodes:print('node:', node.name)input_node = graph.get_tensor_by_name('input_1:0')print("input_node.shape:",input_node.shape)output = graph.get_tensor_by_name('dense_1/Softmax:0') with tf.compat.v1.Session(graph=graph) as sess:logits = sess.run(output, feed_dict = {input_node:img})print("logits:",logits)np.save('out_results/h5tosaved_savedmodelpb_result.npy',logits)
main()num nodes 17
node: kernel
node: bias
node: kernel_1
node: bias_1
node: input_1
node: flatten/Const
node: flatten/Reshape
node: dense/MatMul/ReadVariableOp
node: dense/MatMul
node: dense/BiasAdd/ReadVariableOp
node: dense/BiasAdd
node: dense/Relu
node: dense_1/MatMul/ReadVariableOp
node: dense_1/MatMul
node: dense_1/BiasAdd/ReadVariableOp
node: dense_1/BiasAdd
node: dense_1/Softmax
input_node.shape: (None, 28, 28, 1)
logits: [[2.6573987e-14 1.6878349e-10 1.6627098e-07 1.4200980e-04 1.8240679e-239.9985778e-01 4.3281314e-16 4.1571907e-12 2.4773154e-12 3.5959408e-10]]
以上过程说明，h5转saved没问题，以后saved转pb也是可以
5.2 saved model转h5
def main():file_path='./models/lenet5'model = tf.keras.models.load_model(file_path)model.save('./models/savedtoh5.h5',save_format='h5')main()
5.2.1 saved转h5后的模型推理
import tensorflow as tf
import numpy as np
from PIL import Image
import matplotlib.pyplot as pltdef h5_infer(model_path=None,img=None):model = tf.keras.models.load_model(model_path)result= model.predict(img)return resultdef main():model_path = "models/savedtoh5.h5"img = Image.open('5.png')plt.imshow(img)img= np.array(img).reshape(1,28,28,1)/255.0result = h5_infer(model_path,img)print("result.shape",result.shape)print(result)np.save('out_results/savedtoh5_result.npy',result)#保存结果，方便以后比较
main()
result.shape (1, 10)
[[2.6573987e-14 1.6878349e-10 1.6627098e-07 1.4200980e-04 1.8240679e-239.9985778e-01 4.3281314e-16 4.1571907e-12 2.4773154e-12 3.5959408e-10]]
5.2.2 saved 转h5后的模型转pb
# use tf2.x(tf1.14 also ok) to convert the model hdf5 trained by tf2.x to pb
import tensorflow as tfdef freeze_session(model_path=None,clear_devices=True):tf.compat.v1.reset_default_graph()session=tf.compat.v1.keras.backend.get_session()graph = session.graphwith graph.as_default():model = tf.keras.models.load_model(model_path)output_names = [out.op.name for out in model.outputs]print("output_names",output_names)input_names =[innode.op.name for innode in model.inputs]print("input_names",input_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 = tf.compat.v1.graph_util.convert_variables_to_constants(session, input_graph_def,output_names)outgraph = tf.compat.v1.graph_util.remove_training_nodes(frozen_graph)#去掉与推理无关的内容print("##################################################################")for node in outgraph.node:print('node:', node.name)print("length of node",len(outgraph.node))tf.io.write_graph(frozen_graph, "./models", "savedtoh5_h5.pb", as_text=False)return outgraphdef main(): freeze_session("models/savedtoh5.h5",True)main()
output_names ['dense_1/Softmax']
input_names ['flatten_input']
node: flatten_input
node: flatten/Const
node: flatten/Reshape
node: dense/kernel/Initializer/random_uniform/shape
node: dense/kernel/Initializer/random_uniform/min
node: dense/kernel/Initializer/random_uniform/max
node: dense/kernel/Initializer/random_uniform/RandomUniform
node: dense/kernel/Initializer/random_uniform/sub
node: dense/kernel/Initializer/random_uniform/mul
node: dense/kernel/Initializer/random_uniform
node: dense/kernel
node: dense/kernel/IsInitialized/VarIsInitializedOp
node: dense/kernel/Assign
node: dense/kernel/Read/ReadVariableOp
node: dense/bias/Initializer/zeros
node: dense/bias
node: dense/bias/IsInitialized/VarIsInitializedOp
node: dense/bias/Assign
node: dense/bias/Read/ReadVariableOp
node: dense/MatMul/ReadVariableOp
node: dense/MatMul
node: dense/BiasAdd/ReadVariableOp
node: dense/BiasAdd
node: dense/Relu
node: dense_1/kernel/Initializer/random_uniform/shape
node: dense_1/kernel/Initializer/random_uniform/min
node: dense_1/kernel/Initializer/random_uniform/max
node: dense_1/kernel/Initializer/random_uniform/RandomUniform
node: dense_1/kernel/Initializer/random_uniform/sub
node: dense_1/kernel/Initializer/random_uniform/mul
node: dense_1/kernel/Initializer/random_uniform
node: dense_1/kernel
node: dense_1/kernel/IsInitialized/VarIsInitializedOp
node: dense_1/kernel/Assign
node: dense_1/kernel/Read/ReadVariableOp
node: dense_1/bias/Initializer/zeros
node: dense_1/bias
node: dense_1/bias/IsInitialized/VarIsInitializedOp
node: dense_1/bias/Assign
node: dense_1/bias/Read/ReadVariableOp
node: dense_1/MatMul/ReadVariableOp
node: dense_1/MatMul
node: dense_1/BiasAdd/ReadVariableOp
node: dense_1/BiasAdd
node: dense_1/Softmax
node: Placeholder
node: AssignVariableOp
node: ReadVariableOp
node: Placeholder_1
node: AssignVariableOp_1
node: ReadVariableOp_1
node: Placeholder_2
node: AssignVariableOp_2
node: ReadVariableOp_2
node: Placeholder_3
node: AssignVariableOp_3
node: ReadVariableOp_3
node: VarIsInitializedOp
node: VarIsInitializedOp_1
node: VarIsInitializedOp_2
node: VarIsInitializedOp_3
node: init
node: dense_1_target
node: total/Initializer/zeros
node: total
node: total/IsInitialized/VarIsInitializedOp
node: total/Assign
node: total/Read/ReadVariableOp
node: count/Initializer/zeros
node: count
node: count/IsInitialized/VarIsInitializedOp
node: count/Assign
node: count/Read/ReadVariableOp
node: metrics/accuracy/Squeeze
node: metrics/accuracy/ArgMax/dimension
node: metrics/accuracy/ArgMax
node: metrics/accuracy/Cast
node: metrics/accuracy/Equal
node: metrics/accuracy/Cast_1
node: metrics/accuracy/Const
node: metrics/accuracy/Sum
node: metrics/accuracy/AssignAddVariableOp
node: metrics/accuracy/ReadVariableOp
node: metrics/accuracy/Size
node: metrics/accuracy/Cast_2
node: metrics/accuracy/AssignAddVariableOp_1
node: metrics/accuracy/ReadVariableOp_1
node: metrics/accuracy/div_no_nan/ReadVariableOp
node: metrics/accuracy/div_no_nan/ReadVariableOp_1
node: metrics/accuracy/div_no_nan
node: metrics/accuracy/Identity
node: loss/dense_1_loss/Cast
node: loss/dense_1_loss/Shape
node: loss/dense_1_loss/Reshape/shape
node: loss/dense_1_loss/Reshape
node: loss/dense_1_loss/strided_slice/stack
node: loss/dense_1_loss/strided_slice/stack_1
node: loss/dense_1_loss/strided_slice/stack_2
node: loss/dense_1_loss/strided_slice
node: loss/dense_1_loss/Reshape_1/shape/0
node: loss/dense_1_loss/Reshape_1/shape
node: loss/dense_1_loss/Reshape_1
node: loss/dense_1_loss/SparseSoftmaxCrossEntropyWithLogits/Shape
node: loss/dense_1_loss/SparseSoftmaxCrossEntropyWithLogits/SparseSoftmaxCrossEntropyWithLogits
node: loss/dense_1_loss/weighted_loss/Cast/x
node: loss/dense_1_loss/weighted_loss/Mul
node: loss/dense_1_loss/Const
node: loss/dense_1_loss/Sum
node: loss/dense_1_loss/num_elements
node: loss/dense_1_loss/num_elements/Cast
node: loss/dense_1_loss/Const_1
node: loss/dense_1_loss/Sum_1
node: loss/dense_1_loss/value
node: loss/mul/x
node: loss/mul
len node1 115
INFO:tensorflow:Froze 4 variables.
INFO:tensorflow:Converted 4 variables to const ops.
##################################################################
node: flatten_input
node: flatten/Const
node: flatten/Reshape
node: dense/kernel
node: dense/bias
node: dense/MatMul
node: dense/BiasAdd
node: dense/Relu
node: dense_1/kernel
node: dense_1/bias
node: dense_1/MatMul
node: dense_1/BiasAdd
node: dense_1/Softmax
length of node 13
5.2.3 saved转h5后模型转pb后的推理
#use tf1.14 and tf2.x with pb to load the data are both ok
import tensorflow as tf
from PIL import Image
import numpy as npdef load_graph(file_path):with tf.io.gfile.GFile(file_path,'rb') as f:graph_def = tf.compat.v1.GraphDef()graph_def.ParseFromString(f.read())with tf.compat.v1.Graph().as_default() as graph:tf.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
def main():file_path='./models/savedtoh5_h5.pb'img= np.array(Image.open('5.png')).reshape(1,28,28,1)/255.0graph,graph_nodes = load_graph(file_path)print("num nodes",len(graph_nodes))for node in graph_nodes:print('node:', node.name)input_node = graph.get_tensor_by_name('flatten_input:0')print("input_node.shape:",input_node.shape)output = graph.get_tensor_by_name('dense_1/Softmax:0') with tf.compat.v1.Session(graph=graph) as sess:logits = sess.run(output, feed_dict = {input_node:img})print("logits:",logits)np.save('out_results/savedtoh5_pb_result.npy',logits)
main()num nodes 17
node: flatten_input
node: flatten/Const
node: flatten/Reshape
node: dense/kernel
node: dense/bias
node: dense/MatMul/ReadVariableOp
node: dense/MatMul
node: dense/BiasAdd/ReadVariableOp
node: dense/BiasAdd
node: dense/Relu
node: dense_1/kernel
node: dense_1/bias
node: dense_1/MatMul/ReadVariableOp
node: dense_1/MatMul
node: dense_1/BiasAdd/ReadVariableOp
node: dense_1/BiasAdd
node: dense_1/Softmax
input_node.shape: (None, 28, 28, 1)
logits: [[2.6573987e-14 1.6878349e-10 1.6627098e-07 1.4200980e-04 1.8240679e-239.9985778e-01 4.3281314e-16 4.1571907e-12 2.4773154e-12 3.5959408e-10]]
6、互转部分总结
对于所有模型的互转我们在精度上和模型大小上进行总结。因为这前把所有推理结果都保存了，自然就直接可以对比的
6.1 精度对比
所有模型的结果都保存在out_results中，都读取出来然后进行比较
import numpy as np
# h5模型结果
h5_result=np.load('out_results/h5_result.npy')
# h5转pb后的pb模型的结果
h5pb_result=np.load('out_results/h5pb_result.npy')
# h5转savedmodel的模型的结果
h5tosavedmodel_result=np.load('out_results/h5tosavedmodel_result.npy')
# h5转savedmodel的模型的另一种调用方法做的结果，也就是说模型相同，但调用方法不同
h5tosavedmodelanother_result=np.load('out_results/h5tosavedmodelanother_result.npy')
############################分割线################################
#saved model 结果
savedmodel_result=np.load('out_results/savedmodel_result.npy')
#saved model 另一种调用方式的结果
savedmodelanother_result=np.load('out_results/savedmodelanother_result.npy')
#saved model 转pb 的结果
savedmodelpb_result=np.load('out_results/savedmodelpb_result.npy')
#saved 模型转h5模型的推理结果
savedtoh5_result=np.load('out_results/savedtoh5_result.npy')
#saved 模型转 h5后转pb的结果
savedtoh5_pb_result=np.load('out_results/savedtoh5_pb_result.npy')
以上一共有10个结果我们都与第一个做对比即可
np.testing.assert_array_almost_equal(h5_result,h5pb_result)
np.testing.assert_array_almost_equal(h5_result,h5tosavedmodel_result)
np.testing.assert_array_almost_equal(h5_result,h5tosavedmodelanother_result)
np.testing.assert_array_almost_equal(h5_result,savedmodel_result)
np.testing.assert_array_almost_equal(h5_result,savedmodelanother_result)
np.testing.assert_array_almost_equal(h5_result,savedmodelpb_result)
np.testing.assert_array_almost_equal(h5_result,savedtoh5_result)
np.testing.assert_array_almost_equal(h5_result,savedtoh5_pb_result)
以上代码执行过程中，没有报任何信息，说明所有结果全部相同
6.2 模型大小对比
一共有8个模型我们对比模型大小以及md5sum值来查看
%%bash
#各个h5模型的比较，有两个
#第一个原始的h5模型
du models/lenet5.h5
md5sum -b models/lenet5.h5
#第二个模型 saved 转换成的h5模型
du models/savedtoh5.h5
md5sum -b models/savedtoh5.h5
4800 models/lenet5.h5
810d4ceb0ac0c5de599528378b84c1a9 *models/lenet5.h5
1604 models/savedtoh5.h5
19067c2a5d4587521e2d77df582abd30 *models/savedtoh5.h5
可以看到，saved model转的h5模型小不少，但是结果是一样的
%%bash
#所有saved model的比较，有两个
#第一个是原始的saved model
du models/lenet5
md5sum -b models/lenet5/saved_model.pb
#第二个是h5转换的saved 模型
du models/h5tosaved
md5sum -b models/h5tosaved/saved_model.pb
4784 models/lenet5/variables
4 models/lenet5/assets
4868 models/lenet5
5064cad4d06f58b98c2013a34e65a658 *models/lenet5/saved_model.pb
4784 models/h5tosaved/variables
4 models/h5tosaved/assets
4864 models/h5tosaved
6a9a41e1f3e7ee543934faa793be2962 *models/h5tosaved/saved_model.pb
从以上可以看出二都差不多大
%%bash
#所有pb模型的比较，共有四个
#第一个h5转的pb
du models/lenet5_h5.pb
md5sum -b models/lenet5_h5.pb
#第二个saved model转pb模型
du models/lenet5_savedmodel.pb
md5sum -b models/lenet5_savedmodel.pb
#第三个模型是h5转savedmodel后再转的pb
du models/h5tosaved_savedmodel.pb
md5sum -b models/h5tosaved_savedmodel.pb
#第四个saved 模型转h5再转pb模型
du models/savedtoh5_h5.pb
md5sum -b models/savedtoh5_h5.pb
1592 models/lenet5_h5.pb
1c76f6947f41758bee1c0892877dfea2 *models/lenet5_h5.pb
1636 models/lenet5_savedmodel.pb
c37c6991c6192b11d88b890f13c953f0 *models/lenet5_savedmodel.pb
1632 models/h5tosaved_savedmodel.pb
a00a3395a985482bf2a3cc365505ec48 *models/h5tosaved_savedmodel.pb
1592 models/savedtoh5_h5.pb
a7239f1e842e761aaf75fb79ddfd0e04 *models/savedtoh5_h5.pb
以上四个模型大小相差不大，但不完全相同，还不清楚是什么是原因
总结：
以上为本文第一部分,所用环境为tensorflow2.2gpu版
- h5和saved可以互转
- 互转后的模型可以顺利完成后续所有工作，总的来说精度是一样的，大小saved model转h5模型最小，和所有pb差不多大
- pb模型结点名称各种h5转来的一样，saved model 转pb的名称一样
【Tensorflow2.0】8、tensorflow2.0_hdf5_savedmodel_pb模型转换[2]相关推荐
- TensorFlow2.0学习笔记-3.模型训练
3.模型训练 3.1.Keras版本模型训练 • 构建模型(顺序模型.函数式模型.子类模型) • 模型训练: model.fit() • 模型验证: model.evaluate() • 模型预测: ...
- 【TensorFlow2.0】TensorFlow2.0专栏上线，你来吗？
如今深度学习框架尤其多,几乎每个科技公司巨头都有其深度学习框架,Google的TensorFlow是目前用户最多的框架,几乎所有人都在用.另外在2019年的3月6日-7日TensoFlow开发者大会上 ...
- TensorFlow2.0(八)--tf.function函数转换
tf.function函数转换 1. 关于tf.function 2. tf.function 的实现 3. 关于@tf.function 1. 关于tf.function tf.function的官 ...
- Tensorflow2.0(Keras)转换TFlite
Tensorflow 2.0(Keras)转换TFlite 目录 Tensorflow 2.0(Keras)转换TFlite 1. TensorFlow Lite 指南 (1)TensorFlow L ...
- 图注意力网络(Graph Attention Network, GAT) 模型解读与代码实现(tensorflow2.0)
前面的文章,我们讲解了图神经网络三剑客GCN.GraphSAGE.GAT中的两个: 图卷积神经网络(GCN)理解与tensorflow2.0代码实现 GraphSAGE 模型解读与tensorflow ...
- 【Python深度学习】基于Tensorflow2.0构建CNN模型尝试分类音乐类型(一)
音乐分类 前言 复现代码 MP3转mel CNN模型 训练结果 总结 前言 我在逛github的时候,偶然发现了一个项目:基于深度学习的音乐推荐.[VikramShenoy97].作者是基于CNN做的 ...
- 【TensorFlow2.0】以后我们再也离不开Keras了？
TensorFlow2.0 Alpha版已经发布,在2.0中最重要的API或者说到处都出现的API是谁,那无疑是Keras.因此用过2.0的人都会吐槽全世界都是Keras.今天我们就来说说Keras这 ...
- internetreadfile读取数据长度为0_【完结】TensorFlow2.0 快速上手手册
大家好,这是专栏<TensorFlow2.0>的第五篇文章,我们对专栏<TensorFlow2.0>进行一个总结. 我们知道全新的TensorFlow2.0 Alpha已经于2 ...
- 【TensorFlow2.0】如何搭建网络模型？
大家好,这是专栏<TensorFlow2.0>的第四篇文章,讲述网络模型的搭建. 我们知道在不考虑输入层的情况下,一个典型的卷积神经网络通常由若干个卷积层.激活层.池化层及全连接层组成,无 ...
- 《30天吃掉那只 TensorFlow2.0》 1-4 时间序列数据建模流程范例 (国内新冠疫情结束时间预测问题)
1-4 时间序列数据建模流程范例 (国内新冠疫情结束时间预测问题) 文章目录 1-4 时间序列数据建模流程范例 (国内新冠疫情结束时间预测问题) 一,准备数据 二,定义模型 三,训练模型 四,评估模型 ...
最新文章
- C++继承详解三 ----菱形继承、虚继承
- 【winfrom】事件与委托
- html万花筒图片轮播代码,jQuery实现可拖拽3D万花筒旋转特效
- OSPF Vlink peer的配置
- 由一次不断GC并耗费过长的时间所想到的工具 - jvmkill与jvmquake
- 蓝牙速率详细分析和提升方式【附IOS,安卓速率测试报告】
- android 代码保护 高级混淆
- 学习资料怎么打印,能用手机直接打印学习资料
- ViewBinding和DataBinding的理解和区别
- vscode配置opencv环境，包括opencv源码编译（mingw64 + cmake）
- python inchat库下载_LearnPython/python_wechat.py at master · oxtaw/LearnPython · GitHub
- 配置http和https
- 如何给esxi中的群晖NAS系统添加U盘、USB移动硬盘设备
- C#【保留小数点后位数】与【字符串数字格式化】
- expect的基本用法
- GEE开发2——GPP/ET数据的下载
- 日志采集 - Filebeat
- c语言break语句作用,解析c语言switch中break语句的具体作用
- Java版碰撞球游戏
- [新版系统]软件著作权登记申请流程及注意事项
热门文章
- 104道精选iOS面试题PDF
- 基于Java的卡诺图化简
- LeetCode||颜色分类--给定一个包含红色、白色和蓝色，一共 *n* 个元素的数组，**原地**对它们进行排序，使得相同颜色的元素相邻，并按照红色、白色、蓝色顺序排列。
- iOS更新系统服务器出错,iPhone 更新失败怎么办？更新 iOS 常见的错误代码及解决方法...
- 5G无线技术基础自学系列 | 传统无线网络架构
- 供应商太多，怎么才能高效比价？
- IMS+金蝶K3搭建简易版本供应商协同管理平台（SRM）
- 使用log4j失误导致系统假死，记录一下
- matlab教学ppt,matlab教程ppt(完整版).ppt
- Android报unable to instantiate application怎么解决