ML4TSP Pretrained Files 2024-02-02

This repository primarily stores the pretrained files for ML4TSP. All the files in this repository have a last update date prior to 2024-02-02.

1. Dataset

1.1 Supervised Learning Training Dataset

File Naming Convention: tsp{nodes_num}_{distribution}_{solver(params)}_{size}.txt

Problem Scale	Train File
TSP50	tsp50_uniform_lkh_5k_1.28m.txt
TSP100	tsp100_uniform_lkh_5k_1.28m.txt
TSP500	tsp500_uniform_lkh_50k_128k.txt
TSP1000	tsp1000_uniform_lkh_100k_64k.txt

1.2 Test Dataset (Uniform)

File Naming Convention: tsp{nodes_num}_{solver(params)}_{avg_length}.txt

Problem Scale	Test	Distribution	Size
TSP50	tsp50_concorde_5.68759.txt	uniform	1280
TSP100	tsp100_concorde_7.75585.txt	uniform	1280
TSP500	tsp500_concorde_16.54581.txt	uniform	128
TSP1000	tsp1000_concorde_23.11812.txt	uniform	128
TSP10000	tsp10000_concorde_large_71.84185.txt	uniform	16

2. Model Parameters

2.1 NAR Model Parameters

Pretrained File	Net Type	Layer	Embed	Hidden	Out	Epoch(select)
tsp50_diffusion.pt	gnn	12	128	256	2	100(?)
tsp50_dimes.pt	gnn	12	128	256	2	405/500step
tsp50_gnn.pt	gnn	12	128	256	2	100(96)
tsp50_gnn_wise.pt	gnn	12	128	256	2	100(76)
tsp50_gnn4reg.pt	gnn	12	128	256	2	100(62)
tsp50_us.pt	sag	3	64	64	50	100(3)
tsp100_diffusion.pt	gnn	12	128	256	2	50(?)
tsp100_dimes.pt	gnn	12	128	256	2	240/250step
tsp100_gnn.pt	gnn	12	128	256	2	50(50)
tsp100_gnn_wise.pt	gnn	12	128	256	2	50(48)
tsp100_gnn4reg.pt	gnn	12	128	256	2	50(18)
tsp100_us.pt	sag	3	64	64	50	50(3)
tsp500_diffusion.pt	gnn	12	128	256	2	50(?)
tsp500_dimes.pt	gnn	12	128	256	2	66/100step
tsp500_gnn.pt	gnn	12	128	256	2	50(22)
tsp500_gnn_wise.pt	gnn	12	128	256	2	50(14)
tsp1000_diffusion.pt	gnn	12	128	256	2	50(?)
tsp1000_gnn_wise.pt	gnn	12	128	256	2	50(44)

2.2 AR Model Parameters

Pretrained File	Net Type	Layer	Embed	Heads	Baseline	Epoch(select)
tsp50_am.pt	gat	3	128	8	rollout	360(360)
tsp50_pomo.pt	gat	3	128	8	shared	360(360)
tsp50_symnco.pt	gat	3	128	8	no	360(360)
tsp100_am.pt	gat	3	128	8	rollout	500(500)
tsp100_pomo.pt	gat	3	128	8	shared	100(100)
tsp100_symnco.pt	gat	3	128	8	no	330(329)

3. Training Details

3.1 NAR Model

• lr_scheduler: "cosine-decay" (torch.optim.lr_scheduler.CosineAnnealingLR)
• learning-rate: 0.003(initial)
• optimizer: "AdamW" (torch.optim.AdamW)

3.2 AR Model

• lr_scheduler: None
• learning-rate: 0.0001(fix)
• optimizer: "Adam" (torch.optim.Adam)