A confusing question about the inference

[Lecxxx]

Hi, Thanks for your contribution,but I'm confused about the inference. I must run the Pvqvae inference first,

python scripts/inference.py --name OUTPUT/pvqvae_exp_name/checkpoint/last.pth --func inference_reconstruction --gpu 0 --batch_size 8

then run the transformer inference second

python scripts/inference.py --name OUTPUT/transformer_exp_name/checkpoint/last.pth --func inference_complet_sample_in_feature_for_evaluation --gpu 0 --batch_size 1

Is it right?

[liuqk3]

No. Just

python scripts/inference.py --name OUTPUT/transformer_exp_name/checkpoint/last.pth --func inference_complet_sample_in_feature_for_evaluation --gpu 0 --batch_size 1

is enough.

[Lecxxx]

@liuqk3 Thanks for the reply. But after I run python scripts/inference.py --name OUTPUT/transformer_naturalscene/checkpoint/000199e_495399iter.pth --func inference_complet_sample_in_feature_for_evaluation --gpu 1 --batch_size 1 the completed rusult is terrible.

[liuqk3]

@Lecxxx ,

It seems that the model is not load correctly. You can have a look at this.

[Lecxxx]

@liuqk3 Thanks for the reply. In the prepare_data.md,I found that you only used some pictures of val in Places2,Could you tell me if the Inference part only uses 800 photos in places2?

[liuqk3]

@Lecxxx , Yes. We follow ICT, the testing images on FFHQ, Places2 and ImageNet are exactly the same with ICT. Only 800 images are used for inference to get the metrics.

[Hgy12345]

@liuqk3 Thanks for the reply. But after I run the completed rusult is terrible.python scripts/inference.py --name OUTPUT/transformer_naturalscene/checkpoint/000199e_495399iter.pth --func inference_complet_sample_in_feature_for_evaluation --gpu 1 --batch_size 1

Have you solved the problem? I had a similar situation