celeba_spoof train out model vs the provided trained antispoofing model

[woxiaohan]

I tested the provided trained antispoofing model "MN3_antispoof.pth.tar" on my video, and the spoof faces results is better than live faces. so i used entire celeba_spoof plus self-built data to train.
During the training, I verified that the optimal model obtained by training with the original celeba_spoof dataset and default parameters(only change datasets = dict(Celeba_root='./CelebA_Spoof') and test_dataset = dict(type='celeba_spoof') in config.py), but i did not achieve the corresponding effect of your provided model "MN3_antispoof.pth.tar" on my video.
Is the model you provided trained and verified only with the celeba_spoof dataset? or combined with LCC_FASDcropped and CASIA?