About Fine-tuning

[nyh3174680732]

Hi author, the original upload link you provided has expired. I downloaded a file shared by someone in the issues, and the file structure is shown in the first image below. I trained for iters: 15000 on top of pretrained/ETP/mlm.sap_r2r/ckpts/model_step_82500.pt, but the results were very poor, with SR only reaching 0.36. In a closed issue, I noticed that someone seemed to have downloaded your original uploaded file, and its structure is shown in the second image below. Could you please check which file structure is the one you originally uploaded?

[MarSaKi]

Sorry for the inconvenience, the data and model weights have been uploaded to https://modelscope.cn/models/admagic/ETPNav/files, and the README has been updated.
Please have a try. Thanks!

[nyh3174680732]

Sorry for the inconvenience, the data and model weights have been uploaded to https://modelscope.cn/models/admagic/ETPNav/files, and the README has been updated. Please have a try. Thanks!

Thank you for re-uploading the relevant models and datasets. However, I have a question: when you fine-tuned the pre-trained model model_step_82500.pt on the R2R dataset, which model from the ddppo-models (shown in the picture) did you use? I used gibson-2plus-resnet50.pth for fine-tuning on R2R, and after training for 15,000 iterations, the model performs poorly on the R2R val_unseen dataset, with an OS score of only around 0.37. Even with the same iter12000, the performance is still very poor. I have no idea what the problem is. I noticed that your model already achieves good performance at ckpt.iter12000.pth. Could you clarify which ddppo model you used?

[MarSaKi]

We use data/ddppo-models/gibson-2plus-resnet50.pth. The performance drop may be resulted by the machine (as discussed in other issues), could you change to another machine using RTX series GPUs?

[nyh3174680732]

We use data/ddppo-models/gibson-2plus-resnet50.pth. The performance drop may be resulted by the machine (as discussed in other issues), could you change to another machine using RTX series GPUs?

I'm currently training on 2 NVIDIA A800 GPUs. Could you tell me which RTX series GPU (and which specific model) was used for training the original model in the paper?

[MarSaKi]

I used 4 RTX 3090 gpus to finetune the model.
I think A series GPUs are not stable for Habitat rendering, and may resulted poor performance, as I tried on A100 gpus previously.

[nyh3174680732]

I used 4 RTX 3090 gpus to finetune the model. I think A series GPUs are not stable for Habitat rendering, and may resulted poor performance, as I tried on A100 gpus previously.

thanks for your reply，I try to use RTX series GPU to finetune the model.

[Kinginlife]

I used 4 RTX 3090 gpus to finetune the model. I think A series GPUs are not stable for Habitat rendering, and may resulted poor performance, as I tried on A100 gpus previously.

thanks for your reply，I try to use RTX series GPU to finetune the model.

Hi, I am also experiencing a similar issue during reproduction. I am currently training on 2x A800 GPUs, but the accuracy is only reaching 0.37 at 12,000 iterations and 0.39 at 15,000 iterations, which seems abnormally low. However, I verified my environment by running the evaluation script with the official fine-tuned checkpoints provided in the Releases. The evaluation results match the paper.
I was wondering if you have already resolved this issue? If so, could you please share your solution or any specific configurations that helped? Any insights would be greatly appreciated!