search

JinhuiYE / SignCL

This is the official code repository for the paper 'Improving Gloss-free Sign Language Translation by Reducing Representation Density'.
13 stars 0 forks source link

Problem about reproducing GFSLT! #4

Open AndrewBens opened 1 week ago

AndrewBens commented 1 week ago

Hello🙂! Thanks for your excellent works!😊

Have you ever encountered this problem? I am trying to reproduce GFSLT. But when I try to run train_slt.py after completing 80 epochs of pre-training(VLP_2), I found that a problem occurs when loading the visual encoder on both the Phoenix and CSL-Daily datasets：

Load parameters for Visual Encoder...

Missing keys: mbart.final_logits_bias mbart.model.shared.weight mbart.lm_head.weight sign_emb.src_emb.weight sign_emb.src_emb.bias sign_emb.bn_ac.0.weight sign_emb.bn_ac.0.bias sign_emb.bn_ac.0.running_mean sign_emb.bn_ac.0.running_var

Do you have any suggestions? Thank you!

JinhuiYE commented 1 week ago

hi, do you load with the same tokenizer and config? examples/GFSLT-VLP/configs/config_gloss_free_csl.yaml examples/scripts.sh

JinhuiYE commented 1 week ago

I didn't meet with this

AndrewBens commented 1 week ago

@JinhuiYE Thank you a lot. I am using the mbart_char that I trimmed myself. There may be some bugs. I plan to try the pretrained model you shared in other issues.🙂

AndrewBens commented 1 week ago

@JinhuiYE May I ask you, how long will it take to pre-train and fine-tune CSL-Daily on eight A800s?🤝

JinhuiYE commented 1 week ago

Hi, we are happy to share. In our case, we use 8*A800 GPUs for the two training stages. The pretraining takes about 12 hours for 80 epochs. The finetuning stage takes approximately one day.

AndrewBens commented 1 week ago

@JinhuiYE Get it！Thank you！