search

ku21fan / STR-Fewer-Labels

Scene Text Recognition (STR) methods trained with fewer real labels (CVPR 2021)
MIT License
173 stars 27 forks source link

ablation experiment question #5

Closed daeing closed 2 years ago

daeing commented 2 years ago

Thanks for your excellent work, you have tried a lot of benchmark and ablation experiment, that's very nice of you. and I have a question, I was just wonder if you train the semi- and self supervised model in the same steps (200k)? Because I notice train moco or pseudo label in a larger steps may have a good result. so I was just wondering if you have contrast the result in 200k steps, 300k steps.

ku21fan commented 2 years ago

Hello,

Thank you for your interest in our work :)

if you train the semi- and self supervised model in the same steps (200k)?

Yes, I did.

if you have contrast the result in 200k steps, 300k steps.

Yes, I did. In our experiments, it was up to the experimental setting (using augmentation or semi/self-supervised learning). Sometimes 200K is better, and sometimes 300K is better. We thought that it does not make a significant difference that can change our conclusion. Thus we used 200K steps for faster training.

If the number of steps makes a significant difference and insight, and if someone can find some law or theory between iteration and performance, it would be interesting work! (we had tried to find some law or theory, but it was not easy to conclude)

Hope it helps :) best.

daeing commented 2 years ago

Hello,

Thank you for your interest in our work :)

if you train the semi- and self supervised model in the same steps (200k)?

Yes, I did.

if you have contrast the result in 200k steps, 300k steps.

Yes, I did. In our experiments, it was up to the experimental setting (using augmentation or semi/self-supervised learning). Sometimes 200K is better, and sometimes 300K is better. We thought that it does not make a significant difference that can change our conclusion. Thus we used 200K steps for faster training.

If the number of steps makes a significant difference and insight, and if someone can find some law or theory between iteration and performance, it would be interesting work! (we had tried to find some law or theory, but it was not easy to conclude)

Hope it helps :) best.

Many thanks to your reply, I got what you mean, have a nice day.