X2Paddle实践之——Tensorflow版本VGG模型转化为paddle版本模型

最近经常会刷到很多关于paddle的文章，说实在的paddle我之前读研的时候还是有用过的，后面就没怎么再用过了，经过几年的捶打完善，感觉现在已经是一款不错的深度学习框架了，是时候捡起来学习了解一下了。

PaddlePaddle地址在这里，首页截图如下所示：

上面图中的每一个项目都是当前很火的项目，不过我们今天的主要学习内容不在上面，之前也做过不同框架之间模型转化的工作，今天也是这样的内容，因为PaddlePaddle自家提供了类似的开源工具X2Paddle，顾名思义，就是提供其他框架转化为PaddlePaddle自家模型文件的工具。

话不多说，这里先来看X2Paddle，地址在这里，首页如下：

感兴趣的话可以去仔细了解一下，我们今天主要是借助具体的实践来学习一下具体的操作使用，VGG_16是CV领域的一个经典模型，这里就以VGG_16为例，给大家展示如何将TensorFlow训练好的模型转换为飞桨模型。

首先下载模型，地址如下：

http://download.tensorflow.org/models/vgg_16_2016_08_28.tar.gz

之后解压缩，截图如下所示：

接下来我们需要重新加载参数，并将网络结构和参数一起保存为checkpoint模型，代码如下：

with tf.Session() as sess:inputs = tf.placeholder(dtype=tf.float32, shape=[None, 224, 224, 3], name="inputs")logits, endpoint = vgg.vgg_16(inputs, num_classes=1000, is_training=False)load_model = slim.assign_from_checkpoint_fn("vgg_16.ckpt", slim.get_model_variables("vgg_16"))load_model(sess)data = numpy.random.rand(5, 224, 224, 3)input_tensor = sess.graph.get_tensor_by_name("inputs:0")output_tensor = sess.graph.get_tensor_by_name("vgg_16/fc8/squeezed:0")result = sess.run([output_tensor], {input_tensor:data})numpy.save("tensorflow.npy", numpy.array(result))saver = tf.train.Saver()saver.save(sess, "checkpoint/model")

完成上述处理，我们就可以着手将模型转为PaddlePaddle模型了，代码如下：

parser = convert._get_parser()
parser.meta_file = "checkpoint/model.meta"
parser.ckpt_dir = "checkpoint"
parser.in_nodes = ["inputs"]
parser.input_shape = ["None,224,224,3"]
parser.output_nodes = ["vgg_16/fc8/squeezed"]
parser.use_cuda = "True"
parser.input_format = "NHWC"
parser.save_dir = "paddle_model"
convert.run(parser)

由于是初次学习使用，这里我附上完整的日志输出，如下：

Use tf.compat.v1.graph_util.extract_sub_graph
INFO:root:Tensorflow model loaded!
INFO:root:TotalNum:86,TraslatedNum:1,CurrentNode:inputs
INFO:root:TotalNum:86,TraslatedNum:2,CurrentNode:vgg_16/conv1/conv1_1/weights
INFO:root:TotalNum:86,TraslatedNum:3,CurrentNode:vgg_16/conv1/conv1_1/biases
INFO:root:TotalNum:86,TraslatedNum:4,CurrentNode:vgg_16/conv1/conv1_2/weights
INFO:root:TotalNum:86,TraslatedNum:5,CurrentNode:vgg_16/conv1/conv1_2/biases
INFO:root:TotalNum:86,TraslatedNum:6,CurrentNode:vgg_16/conv2/conv2_1/weights
INFO:root:TotalNum:86,TraslatedNum:7,CurrentNode:vgg_16/conv2/conv2_1/biases
INFO:root:TotalNum:86,TraslatedNum:8,CurrentNode:vgg_16/conv2/conv2_2/weights
INFO:root:TotalNum:86,TraslatedNum:9,CurrentNode:vgg_16/conv2/conv2_2/biases
INFO:root:TotalNum:86,TraslatedNum:10,CurrentNode:vgg_16/conv3/conv3_1/weights
INFO:root:TotalNum:86,TraslatedNum:11,CurrentNode:vgg_16/conv3/conv3_1/biases
INFO:root:TotalNum:86,TraslatedNum:12,CurrentNode:vgg_16/conv3/conv3_2/weights
INFO:root:TotalNum:86,TraslatedNum:13,CurrentNode:vgg_16/conv3/conv3_2/biases
INFO:root:TotalNum:86,TraslatedNum:14,CurrentNode:vgg_16/conv3/conv3_3/weights
INFO:root:TotalNum:86,TraslatedNum:15,CurrentNode:vgg_16/conv3/conv3_3/biases
INFO:root:TotalNum:86,TraslatedNum:16,CurrentNode:vgg_16/conv4/conv4_1/weights
INFO:root:TotalNum:86,TraslatedNum:17,CurrentNode:vgg_16/conv4/conv4_1/biases
INFO:root:TotalNum:86,TraslatedNum:18,CurrentNode:vgg_16/conv4/conv4_2/weights
INFO:root:TotalNum:86,TraslatedNum:19,CurrentNode:vgg_16/conv4/conv4_2/biases
INFO:root:TotalNum:86,TraslatedNum:20,CurrentNode:vgg_16/conv4/conv4_3/weights
INFO:root:TotalNum:86,TraslatedNum:21,CurrentNode:vgg_16/conv4/conv4_3/biases
INFO:root:TotalNum:86,TraslatedNum:22,CurrentNode:vgg_16/conv5/conv5_1/weights
INFO:root:TotalNum:86,TraslatedNum:23,CurrentNode:vgg_16/conv5/conv5_1/biases
INFO:root:TotalNum:86,TraslatedNum:24,CurrentNode:vgg_16/conv5/conv5_2/weights
INFO:root:TotalNum:86,TraslatedNum:25,CurrentNode:vgg_16/conv5/conv5_2/biases
INFO:root:TotalNum:86,TraslatedNum:26,CurrentNode:vgg_16/conv5/conv5_3/weights
INFO:root:TotalNum:86,TraslatedNum:27,CurrentNode:vgg_16/conv5/conv5_3/biases
INFO:root:TotalNum:86,TraslatedNum:28,CurrentNode:vgg_16/fc6/weights
INFO:root:TotalNum:86,TraslatedNum:29,CurrentNode:vgg_16/fc6/biases
INFO:root:TotalNum:86,TraslatedNum:30,CurrentNode:vgg_16/fc7/weights
INFO:root:TotalNum:86,TraslatedNum:31,CurrentNode:vgg_16/fc7/biases
INFO:root:TotalNum:86,TraslatedNum:32,CurrentNode:vgg_16/fc8/weights
INFO:root:TotalNum:86,TraslatedNum:33,CurrentNode:vgg_16/fc8/biases
INFO:root:TotalNum:86,TraslatedNum:34,CurrentNode:vgg_16/conv1/conv1_1/Conv2D
INFO:root:TotalNum:86,TraslatedNum:35,CurrentNode:vgg_16/conv1/conv1_1/BiasAdd
INFO:root:TotalNum:86,TraslatedNum:36,CurrentNode:vgg_16/conv1/conv1_1/Relu
INFO:root:TotalNum:86,TraslatedNum:37,CurrentNode:vgg_16/conv1/conv1_2/Conv2D
INFO:root:TotalNum:86,TraslatedNum:38,CurrentNode:vgg_16/conv1/conv1_2/BiasAdd
INFO:root:TotalNum:86,TraslatedNum:39,CurrentNode:vgg_16/conv1/conv1_2/Relu
INFO:root:TotalNum:86,TraslatedNum:40,CurrentNode:vgg_16/pool1/MaxPool
INFO:root:TotalNum:86,TraslatedNum:41,CurrentNode:vgg_16/conv2/conv2_1/Conv2D
INFO:root:TotalNum:86,TraslatedNum:42,CurrentNode:vgg_16/conv2/conv2_1/BiasAdd
INFO:root:TotalNum:86,TraslatedNum:43,CurrentNode:vgg_16/conv2/conv2_1/Relu
INFO:root:TotalNum:86,TraslatedNum:44,CurrentNode:vgg_16/conv2/conv2_2/Conv2D
INFO:root:TotalNum:86,TraslatedNum:45,CurrentNode:vgg_16/conv2/conv2_2/BiasAdd
INFO:root:TotalNum:86,TraslatedNum:46,CurrentNode:vgg_16/conv2/conv2_2/Relu
INFO:root:TotalNum:86,TraslatedNum:47,CurrentNode:vgg_16/pool2/MaxPool
INFO:root:TotalNum:86,TraslatedNum:48,CurrentNode:vgg_16/conv3/conv3_1/Conv2D
INFO:root:TotalNum:86,TraslatedNum:49,CurrentNode:vgg_16/conv3/conv3_1/BiasAdd
INFO:root:TotalNum:86,TraslatedNum:50,CurrentNode:vgg_16/conv3/conv3_1/Relu
INFO:root:TotalNum:86,TraslatedNum:51,CurrentNode:vgg_16/conv3/conv3_2/Conv2D
INFO:root:TotalNum:86,TraslatedNum:52,CurrentNode:vgg_16/conv3/conv3_2/BiasAdd
INFO:root:TotalNum:86,TraslatedNum:53,CurrentNode:vgg_16/conv3/conv3_2/Relu
INFO:root:TotalNum:86,TraslatedNum:54,CurrentNode:vgg_16/conv3/conv3_3/Conv2D
INFO:root:TotalNum:86,TraslatedNum:55,CurrentNode:vgg_16/conv3/conv3_3/BiasAdd
INFO:root:TotalNum:86,TraslatedNum:56,CurrentNode:vgg_16/conv3/conv3_3/Relu
INFO:root:TotalNum:86,TraslatedNum:57,CurrentNode:vgg_16/pool3/MaxPool
INFO:root:TotalNum:86,TraslatedNum:58,CurrentNode:vgg_16/conv4/conv4_1/Conv2D
INFO:root:TotalNum:86,TraslatedNum:59,CurrentNode:vgg_16/conv4/conv4_1/BiasAdd
INFO:root:TotalNum:86,TraslatedNum:60,CurrentNode:vgg_16/conv4/conv4_1/Relu
INFO:root:TotalNum:86,TraslatedNum:61,CurrentNode:vgg_16/conv4/conv4_2/Conv2D
INFO:root:TotalNum:86,TraslatedNum:62,CurrentNode:vgg_16/conv4/conv4_2/BiasAdd
INFO:root:TotalNum:86,TraslatedNum:63,CurrentNode:vgg_16/conv4/conv4_2/Relu
INFO:root:TotalNum:86,TraslatedNum:64,CurrentNode:vgg_16/conv4/conv4_3/Conv2D
INFO:root:TotalNum:86,TraslatedNum:65,CurrentNode:vgg_16/conv4/conv4_3/BiasAdd
INFO:root:TotalNum:86,TraslatedNum:66,CurrentNode:vgg_16/conv4/conv4_3/Relu
INFO:root:TotalNum:86,TraslatedNum:67,CurrentNode:vgg_16/pool4/MaxPool
INFO:root:TotalNum:86,TraslatedNum:68,CurrentNode:vgg_16/conv5/conv5_1/Conv2D
INFO:root:TotalNum:86,TraslatedNum:69,CurrentNode:vgg_16/conv5/conv5_1/BiasAdd
INFO:root:TotalNum:86,TraslatedNum:70,CurrentNode:vgg_16/conv5/conv5_1/Relu
INFO:root:TotalNum:86,TraslatedNum:71,CurrentNode:vgg_16/conv5/conv5_2/Conv2D
INFO:root:TotalNum:86,TraslatedNum:72,CurrentNode:vgg_16/conv5/conv5_2/BiasAdd
INFO:root:TotalNum:86,TraslatedNum:73,CurrentNode:vgg_16/conv5/conv5_2/Relu
INFO:root:TotalNum:86,TraslatedNum:74,CurrentNode:vgg_16/conv5/conv5_3/Conv2D
INFO:root:TotalNum:86,TraslatedNum:75,CurrentNode:vgg_16/conv5/conv5_3/BiasAdd
INFO:root:TotalNum:86,TraslatedNum:76,CurrentNode:vgg_16/conv5/conv5_3/Relu
INFO:root:TotalNum:86,TraslatedNum:77,CurrentNode:vgg_16/pool5/MaxPool
INFO:root:TotalNum:86,TraslatedNum:78,CurrentNode:vgg_16/fc6/Conv2D
INFO:root:TotalNum:86,TraslatedNum:79,CurrentNode:vgg_16/fc6/BiasAdd
INFO:root:TotalNum:86,TraslatedNum:80,CurrentNode:vgg_16/fc6/Relu
INFO:root:TotalNum:86,TraslatedNum:81,CurrentNode:vgg_16/fc7/Conv2D
INFO:root:TotalNum:86,TraslatedNum:82,CurrentNode:vgg_16/fc7/BiasAdd
INFO:root:TotalNum:86,TraslatedNum:83,CurrentNode:vgg_16/fc7/Relu
INFO:root:TotalNum:86,TraslatedNum:84,CurrentNode:vgg_16/fc8/Conv2D
INFO:root:TotalNum:86,TraslatedNum:85,CurrentNode:vgg_16/fc8/BiasAdd
INFO:root:TotalNum:86,TraslatedNum:86,CurrentNode:vgg_16/fc8/squeezed
INFO:root:Model translated!

执行成功截图如下所示：

转化成功后会出现 paddle_model 目录，如下：

截图如下所示：

转化的工作到这里其实就已经结束了，但是一般的转化都逃不开精度损失的问题，这里需要做一下验证计算。代码如下：

model = ml.ModelLoader("paddle_model", use_cuda=False)
numpy.random.seed(13)
data = numpy.random.rand(5, 224, 224, 3).astype("float32")
# NHWC -> NCHW
data = numpy.transpose(data, (0, 3, 1, 2))
results = model.inference(feed_dict={model.inputs[0]:data})
numpy.save("paddle.npy", numpy.array(results))

执行的时候报错了，错误信息如下：

AssertionError: In PaddlePaddle 2.x, we turn on dynamic graph mode by default, and 'data()' is only supported in static graph mode. So if you want to use this api, please call 'paddle.enable_static()' before this api to enter static graph mode.

代码更正为下面的，其实主要就是静态图的问题。

paddle.enable_static()
model = ml.ModelLoader("paddle_model", use_cuda=False)
numpy.random.seed(13)
data = numpy.random.rand(5, 224, 224, 3).astype("float32")
# NHWC -> NCHW
data = numpy.transpose(data, (0, 3, 1, 2))
results = model.inference(feed_dict={model.inputs[0]:data})
numpy.save("paddle.npy", numpy.array(results))

之后计算一下模型的损失，通过把两个模型文件加载进来后，通过numpy.fabs来求两个模型结果的差异即可，如下：

paddle_result = numpy.load("paddle.npy")
tensorflow_result = numpy.load("tensorflow.npy")
diff = numpy.fabs(paddle_result - tensorflow_result)
print(numpy.max(diff))

结果如下所示：

误差的数据还是比较小的，在可接受的范围内。

最后附上完整代码，如下：

#!usr/bin/env python
#encoding:utf-8
from __future__ import division'''
__Author__:沂水寒城
功能： VGG Tensorflow 转 Paddle 实战
'''import os
import tensorflow.contrib.slim as slim
from tensorflow.contrib.slim.nets import vgg
import tensorflow as tf
import numpy'''
保存模型为checkpoint格式
'''
with tf.Session() as sess:inputs = tf.placeholder(dtype=tf.float32, shape=[None, 224, 224, 3], name="inputs")logits, endpoint = vgg.vgg_16(inputs, num_classes=1000, is_training=False)load_model = slim.assign_from_checkpoint_fn("vgg_16.ckpt", slim.get_model_variables("vgg_16"))load_model(sess)data = numpy.random.rand(5, 224, 224, 3)input_tensor = sess.graph.get_tensor_by_name("inputs:0")output_tensor = sess.graph.get_tensor_by_name("vgg_16/fc8/squeezed:0")result = sess.run([output_tensor], {input_tensor:data})numpy.save("tensorflow.npy", numpy.array(result))saver = tf.train.Saver()saver.save(sess, "checkpoint/model")'''
将模型转换为飞桨模型
'''
import tf2fluid.convert as convert
import argparse
parser = convert._get_parser()
parser.meta_file = "checkpoint/model.meta"
parser.ckpt_dir = "checkpoint"
parser.in_nodes = ["inputs"]
parser.input_shape = ["None,224,224,3"]
parser.output_nodes = ["vgg_16/fc8/squeezed"]
parser.use_cuda = "True"
parser.input_format = "NHWC"
parser.save_dir = "paddle_model"
convert.run(parser)'''
预测结果差异对比
'''
import numpy
import paddle
import tf2fluid.model_loader as ml
paddle.enable_static()
model = ml.ModelLoader("paddle_model", use_cuda=False)
numpy.random.seed(13)
data = numpy.random.rand(5, 224, 224, 3).astype("float32")
# NHWC -> NCHW
data = numpy.transpose(data, (0, 3, 1, 2))
results = model.inference(feed_dict={model.inputs[0]:data})
numpy.save("paddle.npy", numpy.array(results))'''
对比模型损失
'''
import numpy
paddle_result = numpy.load("paddle.npy")
tensorflow_result = numpy.load("tensorflow.npy")
diff = numpy.fabs(paddle_result - tensorflow_result)
print(numpy.max(diff))

初次的实践使用就到这里了，后面有时间继续深入研究下。

X2Paddle实践之——Tensorflow版本VGG模型转化为paddle版本模型相关推荐

人工智能实践：TensorFlow笔记学习（八）—— 卷积神经网络实践
大纲 7.1 复现已有的卷积神经网络 7.2 用vgg16实现图片识别目标掌握复现已有网络,用vgg16实现图片识别 7.1 复现已有的卷积神经网络 VGGNet是Karen simonya ...
人工智能实践：TensorFlow笔记学习（一）—— 人工智能概述
概述一. 基本概念 1.什么是人工智能人工智能的概念:机器模拟人的意识和思维重要人物:艾伦·麦席森·图灵(Alan Mathison Turing) 人物简介:1912年6月23日-19 ...
保存模型后无法训练_模型构建到部署实践
导读在工业界一般会采用了tensorflow-serving进行模型的部署,而在模型构建时会因人而异会使用不同的深度学习框架,这就需要在使用指定深度学习框架训练出模型后,统一将模型转为pb格式,便于 ...
北京大学人工智能实践：Tensorflow笔记——曹健（writed by Enigmalgia）
可以看下我转载的一篇文章机器学习入门概括关于TensorFlow的安装 Ubuntu18.04下安装anaconda和pycharm搭建TensorFlow 贴一下TensorFlow的官方中文版教 ...
人工智能实践：TensorFlow笔记学习（六）—— 全连接网络实践
输入手写数字输出识别结果大纲 6.1 输入手写数字图片输出识别结果 6.2 制作数据集目标 1.实现断点续训 2.输入真实图片,输出预测结果 3.制作数据集,实现特定应用 6.1 输入手写数字图 ...
人工智能实践：TensorFlow笔记学习（五）—— 全连接网络基础
MNIST数据集输出手写数字识别准确率大纲 5.1 MNIST数据集 5.2 模块化搭建神经网络 5.3 手写数字识别准确率输出目标利用MNIST数据集巩固模块化搭建神经网路的八股,实践前向传播 ...
人工智能实践：TensorFlow笔记学习（四）—— 神经网络优化
神经网络优化大纲 4.1 损失函数 4.2 学习率 4.3 滑动平均 4.4 正则化 4.5 神经网络搭建八股目标掌握神经网络优化方法 4.1 损失函数神经元模型:用数学公式表示为:,f为激 ...
人工智能实践：TensorFlow笔记学习（三）——TensorFlow框架
搭建神经网络大纲 3.1 张量.计算图.会话 3.2 前向传播 3.3 反向传播目标搭建神经网络,总结搭建八股 3.1 张量.计算图.会话一.基本概念基于Tensorflow的NN:用张量表 ...
我的AI之路(5)--如何选择和正确安装跟Tensorflow版本对应的CUDA和cuDNN版本
最新的Tensorflow和CUDA cuDNN的对应关系可以从这里找到: https://tensorflow.google.cn/install/source https://tensorflow ...
SOTA模型飞入寻常百姓家-BEiT模型在AIStudio动手实践
转自AI Studio,原文链接:SOTA模型飞入寻常百姓家-BEiT模型在AIStudio动手实践 - 飞桨AI Studio 一.缘起众所周知Transformer模型精度高,但是训 ...

X2Paddle实践之——Tensorflow版本VGG模型转化为paddle版本模型

X2Paddle实践之——Tensorflow版本VGG模型转化为paddle版本模型相关推荐

最新文章

热门文章