Skip to content

Latest commit

 

History

History
79 lines (69 loc) · 2.66 KB

README.md

File metadata and controls

79 lines (69 loc) · 2.66 KB
Raw

Deployment tool

You can run inference of models using different backends. Currently, only Ascend backend is supported. To proceed, follow next steps:

1. Create "om" file from the Mindspore checkpoint. For it, ASCEND_TOOLKIT_HOME environment variable should be set:

export ASCEND_TOOLKIT_HOME=/usr/local/Ascend/ascend-toolkit/<ascend_toolkit_version>

By default, it's set to /usr/local/Ascend/ascend-toolkit/latest.

Then run the script:

python export.py --config_path <path_to_models_config_file> --target_format om --shape <N, T, C, H, W> --output_name <name_of_file> --output_type <UINT8/FP16/FP32> --fusion_switch_file mindediting/deploy/utils/fusion_switch.cfg

Important: Do not forget to set fusion file, it highly impacts the performance.

Finally, you'll get a file <name_of_file>.om

2. Prepare the configuration file for deployment:

You can run different tasks in a pipeline, for that create own section in the configuration file.

task_name: task_type: <image/video> input_file: <input file / folder> output_file: windows_size: <portion of frames to read, better to set to 1> fixed_width: <width to interpolate the original resolution before the inference, better to use original width> fixed_height: <height to interpolate the original resolution before the inference, better to use original height> up_scale: color_space: <rgb/bgr> tasks:

  • task_name: task_type: super_resolution data_io: <basic/ipt_sr/srdiff_sr/rgb_to_bgr/bgr_to_rgb> backend: ascend dtype: <float16/float32> model_file: "<path_to_model>.om" once_process_frames: <number of frames to run inference on, may be all video size> frame_overlap: patch_overlap: up_scale: device_id: tiling: <default/simple_image/None>
  • task_name: ... ...

The examples of task configuraton files are placed at LLVT/configs directory.

3. Add path to mindediting package to the PYTHONPATH environment variable:

export PYTHONPATH={path_to_LLVT}:$PYTHONPATH

And run the script using the created configuration file:

python -m mindediting.deploy.deploy -pf <config.yaml>

4. You can also validate the "om" model and get metrics using val.py script.

For that, set the following lines to the model configuration file:

model:
  name: "om"
  load_path: <path to "om" model>
validator:
  name: "tiling" or "om_default"
  temporal_overlap: <>
  spatial_overlap: <>
  temporal_size: <>
  spatial_size: <>
  scale: <>
  input_tensor_type: numpy
  dtype: <>

Then run:

python val.py --config_path <configuration file>