Use the x2mindspire tool of mindstudio to convert training scripts

One 、 Reference material

X2MindSpore

X2MindSpore Function is introduced

Two 、 Related introduction

X2MindSpore The script conversion tool can convert PyTorch Scripts and TensorFlow 2.x The script is converted to MindSpore Code .

2.1 Important note

Currently only supported PyTorch and TensorFlow 2.x Training script conversion .

3、 ... and 、 Key steps

This article takes TensorFlow Version of FCN The model, for example .

3.1 Download the source code

Download link ：FCN

Blogger's code ：https://gitee.com/lljyoyo1995/fcn

3.2 Prepare the dataset

download VOC Data sets

$ wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar
$ wget http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar
$ wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar

Decompress and organize the data set

VOC           # path: /home/yang/dataset/VOC
├── test
|    └──VOCdevkit
|        └──VOC2007 (from VOCtest_06-Nov-2007.tar)
└── train
     └──VOCdevkit
         └──VOC2007 (from VOCtrainval_06-Nov-2007.tar)
         └──VOC2012 (from VOCtrainval_11-May-2012.tar)

3.3 Generate txt Mark the file

python parser_voc.py --voc_path /PATH/TO/VOC

3.4 Modify source code

parser_voc.py Used in scipy.misc.imread Method is scipy 1.2.0 The old version before API,mindspore Minimum compatibility scipy 1.5.2, So please use scipy Recommended in the official deprecation warning of imageio.imread.

 terms of settlement ：
1.  install imageio
pip install imageio

2.  modify parser_voc.py Source code 
label_image = np.array(misc.imread(label_path))
 Change it to 
label_image = np.array(imageio.imread(label_path))

3.5 Local runthrough code

If in TensorFlow 1.X Run... Under version , Please turn on emergency execution ：

import tensorflow as tf
tf.enable_eager_execution()

Pull up training

python train.py

Reasoning test

python test.py

3.6 Framework Trans

Prerequisite ： After running through the code , To carry out Framework Trans operation .

Install dependency packages

#pandas The version number must be greater than or equal to 1.2.4
pip3 install pandas
pip3 install libcst

Perform conversion

X2MindSpore Operation guide

cd Ascend-cann-toolkit The installation directory /ascend-toolkit/{
    version}/{
    arch}-linux/toolkit/tools/x2mindspore/

./run_x2mindspore.sh -i  Original script path  -o  Script conversion result output path  -f  The framework of the original script 

Four 、TensorFlow GPU2Ascend

4.1 Set up Remote CANN

restart MindStudio Activate CANN

Set up the success

4.2 Set up TensorFlow GPU2Ascend

# CANN Machine
[email protected]:32542

# Command File
/home/ma-user/Ascend/tfplugin/5.0.3/python/site-packages/npu_bridge/convert_tf2npu/main.py

# Input Paht
\PATH\TO\TensorFlow2.0-Examples\5-Image_Segmentation\FCN

# Output Path
\PATH\TO\TensorFlow2.0-Examples\5-Image_Segmentation

determine yes

5、 ... and 、FAQ

Q： Set up Remote CANN Failure

 The reason for the error ：
 The remote server is not installed ascend-toolkit development tool 

 Solution 1 ：
 install ascend-toolkit development tool 


 Solution 2 ：
 Find one installed ascend-toolkit Server for 

Q：TensorFlow Inconsistent versions lead to training errors

tensorflow.python.framework.errors_impl.InvalidArgumentError: Can not squeeze dim[2], expected a dimension of 1, got 224
	 [[{
    {
    node metrics/acc/Squeeze}}]]

result = visual_result(x[0], pred_label[0].numpy())
AttributeError: 'Tensor' object has no attribute 'numpy'

 The reason for the error ：
 Source code support TF2.X, But the blogger's environment is TF1.X

 Solution 1 ：
 stay TF2.X Running in the environment 

 Solution 2 ：
 stay TF1.X Running in the environment , Emergency execution needs to be started 
import tensorflow as tf
tf.enable_eager_execution()

Q：h5py Inconsistent versions lead to reasoning errors

Report errors ：original_keras_version = f.attrs[‘keras_version‘].decode(‘utf8‘) And problem solving

original_keras_version = f.attrs['keras_version'].decode('utf8')
AttributeError: 'str' object has no attribute 'decode'

 The reason for the error ：
 In the installation tensorflow when , Default installation h5py by 3.1.0, The error is because you installed TF An overly high version of..., is not supported h5py.

 terms of settlement ：
 uninstall h5py3.1.0 edition , install h5py2.10.0 edition 
pip install h5py==2.10.0

Q： Training stuck

 The reason for the error ：
 use parser_voc.py Generated train_image.txt The image path in is wrong 

 terms of settlement ：
\ The backslash 
 Change to 
/ Slash 

Q： The lack of libcst

 The reason for the error ：
 The lack of libcst rely on 

 terms of settlement :
 install libcst
pip3 install libcst

Q： The lack of X2MindSpore Tools

 The reason for the error ：
 Missing in server X2MindSpore Tools 

 terms of settlement ：
 install X2MindSpore Tools