Pre-trained models

The following table summarizes the performance of the pre-trained models and the URLs of checkpoints and configurations. The scores differ from the paper due to refactoring of the code (especially AR, which improves performance substantially). We set --seed_everything 0 for both training and evaluation. The training logs can be found at TensorBoard.dev here.

Name	Accuracy (RGB)	Accuracy (Alpha)	F-score (RGB)	F-score (Alpha)	BG-FCD	Text-FCD	Pixel-FCD	Contrast (% Pages)	Contrast (# Elements)	URL	Note
Stats (mode)	0.717	0.891	0.003	0.219	22.15	265.08	7.49	23.79	0.59	ckpt, config	Evaluate with --model.sampling false.
Stats (sampling)	0.620	0.821	0.004	0.206	0.71	83.55	172.94	93.35	4.07	ckpt, config	Evaluate with --model.sampling true.
AR (greedy)	0.758	0.916	0.064	0.612	2.14	23.91	12.63	65.69	1.96	ckpt, config	Evaluate with --model.top_p 0.0.
AR (top-p, p=0.8)	0.755	0.915	0.062	0.605	1.98	22.33	18.57	68.63	2.11	ckpt, config	Evaluate with --model.top_p 0.8.
AR (top-p, p=0.9)	0.753	0.914	0.061	0.604	1.84	21.00	21.33	69.78	2.19	ckpt, config	Evaluate with --model.top_p 0.9.
NAR	0.774	0.929	0.078	0.677	1.97	22.46	7.80	71.83	2.22	ckpt, config
CVAE	0.773	0.929	0.069	0.663	1.65	28.67	7.90	72.91	2.23	ckpt, config
Upsampler	-	-	-	-	-	-	-	-	-	ckpt, config	Used together with all the other models.