search

syuqings / video-paragraph

Codes for paper "Towards Diverse Paragraph Captioning for Untrimmed Videos". CVPR 2021
MIT License
64 stars 3 forks source link

RL not working #3

Open Kashu7100 opened 3 years ago

Kashu7100 commented 3 years ago

[issue] The fine-tuning step doesn't increase the scores (it even decreases the score). Please refer to the green line in the chart below.

image

[How to reproduce] I have trained the provided code on ActivityNet dataset. I have followed the data preparation and training instructions of the README.

[questions] Is this the same behavior observed during your training?

syuqings commented 3 years ago

Hi, in my training, I found the RL cannot improve the pragraph generation as significantly as traditional video captioning. It will increase the scores in the first several epochs and then decrease. The following is my re-training log. log.txt

Kashu7100 commented 3 years ago

Thank you for the reply. Do you think this is due to the model architecture (compared to the ordinary two stage models) or some implementation? I will also take a look at the rl part of the code.