Environment.md
Inference - Environment Installation¶
MindOCR supports inference for Ascend310/Ascend310P device.
Please make sure that the Ascend AI processor software package is correctly installed on your system. If it is not installed, please refer to the section Installing Ascend AI processor software package to install it.
The MindOCR backend supports two types of inference: ACL and MindSpore Lite. Before inference using ACL mode, you need to use ATC tool to convert the model to om format, or to use converter_lite tool to convert the model to MindIR format, the specific differences are as follows:
| ACL | Mindspore Lite | |
|---|---|---|
| Conversion Tool | ATC | converter_lite |
| Inference Model Format | om | MindIR |
1. ACL inference¶
For the ACL inference of MindOCR, it currently relies on the Python API interface by MindX, which currently only supports Python 3.9.
| package | version |
|---|---|
| Python | 3.9 |
| MindX | 3.0.0 |
On the basis of the Python 3.9 environment, download the mxVision SDK installation package for MindX and refer to the tutorial for installation. The main steps are as follows:
# add executable permissions
chmod +x Ascend-mindxsdk-mxvision_{version}_linux-{arch}.run
# execute the installation command
# if prompted to specify the path to CANN, add parameters such as: --cann-path=/usr/local/Ascend/latest
./Ascend-mindxsdk-mxvision_{version}_linux-{arch}.run --install
# set environment variable
source mxVision/set_env.sh
If use python interface, after installation, test whether mindx can be imported normally：python -c "import mindx"
If prompted that mindx cannot be found, go to the mxVision/Python directory and install the corresponding Whl package:
cd mxVision/python
pip install *.whl
2. MindSpore Lite inference¶
For the MindSpore Lite inference of MindOCR, It requires the version 2.2.0 or higher of the MindSpore Lite cloud-side inference toolkit.
Download the Ascend version of the cloud-side inference toolkit tar.gz file, as well as the Python interface Wheel package. For example, when using Ascend，system on Linux-x86_64，Python 3.7，and mindspore-lite 2.2.0，the following package should be downloaded and installed:
The download address provides the Python package for version 3.7. If you need other versions, please refer to the compilation tutorial.
Just decompress the inference toolkit, and set environment variables:
export LITE_HOME=/your_path_to/mindspore-lite
export LD_LIBRARY_PATH=$LITE_HOME/runtime/lib::$LITE_HOME/runtime/third_party/dnnl:$LITE_HOME/tools/converter/lib:$LD_LIBRARY_PATH
export PATH=$LITE_HOME/tools/converter/converter:$LITE_HOME/tools/benchmark:$PATH
If using python interface, install the required .whl package using pip:
pip install mindspore_lite-{version}-{python_version}-linux_{arch}.whl