mindspore-mindinsight/mindinsight/mindconverter

• Introduction
• Installation
• Commandline Usage
• • Use example:
• Unsupported Situation #1
• Unsupported Situation #2
• Unsupported Situation #3

Introduction

MindConverter is a tool that converting PyTorch scripts to MindSpore scripts. With minial manual editing and the guidance from conversion reports, users may easily migrate their model from PyTorch framework to MindSpore.

Installation

This tool is part of MindInsight and accessible to users after installing MindInsight, no extra installation is needed.

Commandline Usage

Set the model scripts directory as the PYTHONPATH environment variable first:

export PYTHONPATH=<model scripts dir>

mindconverter commandline usage:

mindconverter [-h] [--version] --in_file IN_FILE [--output OUTPUT]
              [--report REPORT]

optional arguments:
  -h, --help         show this help message and exit
  --version          show program's version number and exit
  --in_file IN_FILE  Specify path for script file.
  --output OUTPUT    Specify path for converted script file directory. Default
                     is output directory in the current working directory.
  --report REPORT    Specify report directory. Default is the current working
                     directorys

Use example:

We have a collection of PyTorch model scripts

~$ ls
models
~$ ls models
lenet.py resnet.py vgg.py

Then we set the PYTHONPATH environment variable and convert alexnet.py

~$ export PYTHONPATH=~/models
~$ mindconverter --in_file models/lenet.py

Then we will see a conversion report and the output MindSpore script

~$ ls
lenet_report.txt models output
~$ ls output
lenet.py

Please checkout models/lenet.py and output/lenet.py as the input/output example representatively.
Since the conversion is not 100% flawless, we encourage users to checkout the report when fixing issues of the converted script.

Unsupported Situation #1

Classes and functions that can't be converted:

• The use of shape, ndim and dtype member of torch.Tensor.
• torch.nn.AdaptiveXXXPoolXd and torch.nn.functional.adaptive_XXX_poolXd()
• torch.nn.functional.Dropout
• torch.unsqueeze() and torch.Tensor.unsqueeze()
• torch.chunk() and torch.Tensor.chunk()

Unsupported Situation #2

Subclassing from the subclasses of nn.Module

e.g. (code snip from torchvision,models.mobilenet)

from torch import nn

class ConvBNReLU(nn.Sequential):
    def __init__(self, in_planes, out_planes, kernel_size=3, stride=1, groups=1):
        padding = (kernel_size - 1) // 2
        super(ConvBNReLU, self).__init__(
            nn.Conv2d(in_planes, out_planes, kernel_size, stride, padding, groups=groups, bias=False),
            nn.BatchNorm2d(out_planes),
            nn.ReLU6(inplace=True)
        )

Unsupported Situation #3

Unconventional import naming

e.g.

import torch.nn as mm