it seems hard to reproduce your results in paper.

[ictnlp]

it seems hard to reproduce your results in paper.

[patrick-tssn]

The best score on DSTC8-AVSD so far (LM + VASM)
CIDEr: 1.220
Bleu_4: 0.434
Bleu_3: 0.517
Bleu_2: 0.617
Bleu_1: 0.738
ROUGE_L: 0.589
METEOR: 0.284

[lizekang]

Hello, It's hard to find the checkpoint. I will help you to reproduce it. Which checkpoint do you select in your experiments?

[patrick-tssn]

thanks for your reply, the parameters of generate model are: ckpt:114888 beam_size: 4 min_len: 1 penalty: 0.3. I test it on dstc8avsd test set. training procedure is same with your released code.

[lizekang]

Which epoch checkpoint did you select? You should select the checkpoint at epoch 3 or 4. For the other hyper-parameters, you can keep same with the code.

[patrick-tssn]

I can obtain comparable score now, thank you for your reply.