Usage Guide for Models Provided by ExecuTorch

This guide provides examples and instructions for deploying and executing models using the ExecuTorch runtime on Qualcomm platforms.

Model categories

The following models in two folders can be categorized based on their primary use cases.

1. Language Model:

   • mobilebert_fine_tune

2. Speech Model:

   • wav2letter

3. Vision Model:

   • edsr
   • inception_v3
   • inception_v4
   • mobilenet_v2
   • mobilenet_v3
   • torchvision_vit

Prerequisite

Please follow another README first to set up environment.

Model running

Some models require specific datasets. Please download them in advance and place them in the appropriate folders.

Detailed instructions for each model are provided below. If you want to export the model without running it, please add --compile_only to the command.

1. deeplab_v3:

   • Required Dataset : VOCSegmentation
     Will be downloaded automatically if -d is specified

     python deeplab_v3.py -m ${SOC_MODEL} -b path/to/build-android/ -s ${DEVICE_SERIAL} 
     
     

2. edsr:

   • Required Dataset : DIV2K

     Will be downloaded automatically if -d is specified. Alternatively, you can provide your own dataset using --hr_ref_dir and --lr_dir.

     pip install piq
     python edsr.py -m ${SOC_MODEL} -b path/to/build-android/ -s ${DEVICE_SERIAL} -d 
     
     

3. inception_v3, inception_v4, mobilenet_v2,mobilenet_v3, torchvision_vit:

   • Required Dataset : ImageNet

     Download dataset first, and place it in a valid folder.

     python SCRIPT_NAME.py -m ${SOC_MODEL} -b path/to/build-android/ -s ${DEVICE_SERIAL} -d path/to/ImageNet
     
     

4. mobilebert_fine_tune:

   • You can specify the pretrained weight using -p <path/to/pretrained_weight>, if no pretrained weights are provided, using --num_epochs to set number of epochs to train the model.

   • -F --use_fp16: If specified, the model will run in FP16 mode and the PTQ will be ignored.

     python mobilebert_fine_tune.py -m ${SOC_MODEL} -b path/to/build-android/ -s ${DEVICE_SERIAL} --num_epochs <number_of_epochs>
     
     

5. wav2letter:

   • Pretrained weight :

     for torchaudio.models.Wav2Letter version, please download at here, and place it in a valid folder.

     python wav2letter.py -m ${SOC_MODEL} -b path/to/build-android/ -s ${DEVICE_SERIAL} -p path/to/pretrained_weight