tranz 0.5.2

Point-embedding knowledge graph models: TransE, RotatE, ComplEx, DistMult. GPU training via candle.
Documentation
1
2
3
4
5
6
timestamp	model	dataset	epochs	dim	lr	optimizer	1n	recip	n3	l2	swa	rel_pred	mrr	h@1	h@3	h@10	head_mrr	tail_mrr	train_s	commit	command	notes
2026-04-06T09:00	complex	WN18RR	10	50	0.001	adamw	yes	yes	0	0	0	0	0.3964	0.3653	0.4126	0.4505	-	-	2615	4136898	tranz train --data data/WN18RR/ --model complex --dim 50 --1n --label-smoothing 0.1 --reciprocals --epochs 10 --lr 0.001 --log-interval 5 --eval --output /tmp/tranz-quick	quick validation (dim=50, 10ep)
2026-04-06T12:00	complex	WN18RR	50	100	0.001	adamw	yes	yes	0	0	0	0	0.4334	0.4062	0.4435	0.4866	0.4049	0.4618	11286	9e1aadd	tranz train --data data/WN18RR/ --model complex --dim 100 --1n --label-smoothing 0.1 --reciprocals --epochs 50 --lr 0.001 --eval	baseline Adam 1-N 50ep
2026-04-06T12:00	complex	WN18RR	50	100	0.001	adamw	yes	yes	0	0	25	0	0.4299	0.4022	0.4405	0.4876	0.4021	0.4577	11286	9e1aadd	tranz train --data data/WN18RR/ --model complex --dim 100 --1n --label-smoothing 0.1 --reciprocals --epochs 50 --lr 0.001 --swa 25 --eval	+SWA start=25 (SWA helps more at longer training)
2026-04-06T22:00	complex	WN18RR	50	100	0.001	adamw	yes	yes	0	0	0	0.1	0.4099	0.3896	0.4177	0.4488	0.3831	0.4366	10224	9b9aabf	tranz train --data data/WN18RR/ --model complex --dim 100 --1n --label-smoothing 0.1 --reciprocals --epochs 50 --lr 0.001 --rel-pred 0.1 --eval	+rel-pred hurts on WN18RR (11 relations too few)
2026-04-07T00:00	complex	WN18RR	100	100	0.1	adagrad	yes	yes	0.1	0	0	0	0.4275	0.3960	0.4373	0.4915	0.4050	0.4499	17840	da944d2	tranz train --data data/WN18RR/ --model complex --dim 100 --1n --reciprocals --optimizer adagrad --init-scale 1e-3 --n3 0.1 --lr 0.1 --epochs 100 --eval	Adagrad+N3 (README recipe, no label-smoothing)