search

ReneeYe / ConST

code for paper "Cross-modal Contrastive Learning for Speech Translation" (NAACL 2022)
MIT License
62 stars 6 forks source link

How to run the code without external MT data? #8

Open XqZeppelinhead0702 opened 1 week ago

XqZeppelinhead0702 commented 1 week ago

Hello, recently I try to train the ConST model without external MT data just to test its performance in this situation and I'm confused about how to regulate the parameters to achieve this. I think the '--task, --external-parallel-mt-data, --criterion' should be changed but I fail to find the correct candidate for them due to so many choices. I would be extremely grateful if you could give me some guidance. Looking forward to your early reply.

ReneeYe commented 2 days ago

Hello, recently I try to train the ConST model without external MT data just to test its performance in this situation and I'm confused about how to regulate the parameters to achieve this. I think the '--task, --external-parallel-mt-data, --criterion' should be changed but I fail to find the correct candidate for them due to so many choices. I would be extremely grateful if you could give me some guidance. Looking forward to your early reply.

It's been a long time since I last ran the code. If I remember correctly, the code in script/train_en2x.sh also supports WITHOUT external MT dataset, --task=speech_to_text_triplet_with_extra_mt but leave --external-parallel-mt-data as none, for criterion, you may use label_smoothed_cross_entropy_with_constrastive.