1、补充trainCAN.py配置

config_defaults = {
    "epochs": 30,
    "train_batch_size": 8,
    "valid_batch_size": 12,
    "gpu_nums": 1,
    "optimizer": "adam",
    "learning_rate": 2e-4,
    "weight_decay": 0,
    "rand_seed": 888,
    "accumulation_steps": 2,
    "path": ""   # 预训练模型路径
}

2、修改listener.py第15行

cmd = 'CUDA_VISIBLE_DEVICES="GPU" python -m torch.distributed.launch --nproc_per_node GNUM --master_port 2515  trainCAN.py >>./logs/data.log 2>&1 &'

3、视情况修改GPU数量、指定GPU编号

NUM = 2

GPU_list=[6,7]

4、激活环境，运行命令

python listener.py

5、更多相关内容

见论文或Figure文件夹

论文: Where to Focus: Central Attention-Based Face Forgery Detection

Link

6、公开仓库 Link