Sense-X / UniFormer

[ICLR2022] official implementation of UniFormer
Apache License 2.0
828 stars 111 forks source link

Training time for kinetics-400 #96

Closed AbdelrahmanShakerYousef closed 2 years ago

AbdelrahmanShakerYousef commented 2 years ago

Hello,

Thank you for sharing the codebase of your exciting work.

Could you please let me know the training time for pertaining & training on kinetics-400 and the resources you used?

Thank you!

Andy1621 commented 2 years ago

Thanks for your question. In my original experiments, all the models are run on a single V100 machine (8GPU) except the largest model. For detailed training time, you can find the provided logs! It may take ~2 days to train K400 for 110 epochs with 16 frames.

For your experiments, you can use fewer epochs (~55) and fewer frames (~8) to verify your idea fastly.

AbdelrahmanShakerYousef commented 2 years ago

Thank you for reply.

According to verifying the idea, shall I pretrain on ImageNet first and then train on K400? If this is the case, how many days your model need to be pretrained on ImageNet ?

Get Outlook for iOShttps://aka.ms/o0ukef


From: Kunchang Li @.> Sent: Thursday, November 17, 2022 7:06:13 AM To: Sense-X/UniFormer @.> Cc: Abdelrahman Mohamed Shaker Youssief @.>; Author @.> Subject: Re: [Sense-X/UniFormer] Training time for kinetics-400 (Issue #96)

Thanks for your question. In my original experiments, all the models are run on a single V100 machine (8GPU) except the largest model. For detailed training time, you can find the provided logs! It may take ~2 days to train K400 for 110 epochs with 16 frames.

For your experiments, you can use fewer epochs (~55) and fewer frames (~8) to verify your idea fastly.

— Reply to this email directly, view it on GitHubhttps://are01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2FSense-X%2FUniFormer%2Fissues%2F96%23issuecomment-1317996456&data=05%7C01%7Cabdelrahman.youssief%40mbzuai.ac.ae%7Cec151365ecad423f5b2808dac848b108%7Cc93272d31b074b3da3b619b34a973915%7C0%7C0%7C638042511780785787%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=61nHk4FTyYKkHd20Y9KryfsC5xzOgDHawvqhHlzz%2BEE%3D&reserved=0, or unsubscribehttps://are01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fnotifications%2Funsubscribe-auth%2FAMO773HSHAIYFXOANMPJTNDWIWOKLANCNFSM6AAAAAASCECWSA&data=05%7C01%7Cabdelrahman.youssief%40mbzuai.ac.ae%7Cec151365ecad423f5b2808dac848b108%7Cc93272d31b074b3da3b619b34a973915%7C0%7C0%7C638042511780785787%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=0uTc6kT04n1hc8XJPa0YmLLwdvGbVFvoChPyf%2F9wONU%3D&reserved=0. You are receiving this because you authored the thread.Message ID: @.***>

Andy1621 commented 2 years ago

It's a good question. As a new backbone, it needs to be pre-trained on ImageNet first, which will cost 1~2 day. Thus, it actually needs about 3~4 days to verify it on K400. However, the performance trend is similar. Those models work better on ImageNet often work better on K400.

AbdelrahmanShakerYousef commented 2 years ago

Thank you Andy for your reply. Is it 12 days or 34 days? I guess your text is unclear about that.

Andy1621 commented 2 years ago

1 to 2 days and 3 to 4 days.

The ~ automatically changes to delete.😊

Andy1621 commented 2 years ago

Hi! We have released UniFormerV2, which adopts pretrained ViT. It can save a lot of training resources. Welcome to try it!