Training Asccl on Ascend server

This has enabled the training of asccl on Ascend NPU.

Requirements

• Install PyTorch (pytorch.org).
• pip install -r requirements.txt
• Download the MEAD dataset from (here).
• Download the pre-trained weights (here) (" backbone.pth ") and place it under "./pretrain/backbone.pth"

Environmental configuration

Before running the code, it is necessary to configure the environment through 'env. sh':

source env.sh

Preprocessing

The obtained MEAD dataset is first preprocessed with 'dataloader/align_face.py':

python ./dataloader/align_face.py

Training

To train the model, run './trainer/train_asccl.py' with the preprocessed dataset path configured:

python ./trainer/train_asccl.py

After approximately 50 epochs, you can obtain a checkpoint file. This checkpoint can be used to supervise the training of the SPFEM model.