![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F15620263%2F031a2aa184a9511af30dd036dd97dcb4%2Flosses.png?generation=1745717780490145&alt=media\")
\nHave you ever tried fine-tuning Protenix with training data?
\nI'm currently attempting it, but my experiment is encountering an unsolvable error.
\nAs you can see in the image, the gap between the training set's MSE loss and the validation set's (CASP15) MSE loss is extremely large.
\nFor anyone who has experience fine-tuning Protenix, how did you address this issue?\"
\nHave you ever tried fine-tuning Protenix with training data?
\nI'm currently attempting it, but my experiment is encountering an unsolvable error.
\nAs you can see in the image, the gap between the training set's MSE loss and the validation set's (CASP15) MSE loss is extremely large.
\nFor anyone who has experience fine-tuning Protenix, how did you address this issue?\"
*It's fine-tuned with no-MSA
", "rawMarkdown": "*It's fine-tuned with no-MSA", "votes": null }, { "id": "3188348", "postDate": "04/27/2025 13:02:46", "content": "Same as yours, but I don't think it is a problem. I think the MSE loss is computed from the diffusion steps. In training and inference, the diffusion is different. In inference, the diffusion steps are 200, and each step will generate an MSE loss and accumulate to be a final MSE loss. But during training, I think it's much less for efficiency, maybe just randomly sample several steps. I am still trying to understand diffusion. Correct me if I am wrong.
\nSo far my fine-tuned Protenix model fails to outperform the original checkpoint. I am trying to understand why. If you have any thoughts, it would be great to hear from you.
", "rawMarkdown": "Same as yours, but I don't think it is a problem. I think the MSE loss is computed from the diffusion steps. In training and inference, the diffusion is different. In inference, the diffusion steps are 200, and each step will generate an MSE loss and accumulate to be a final MSE loss. But during training, I think it's much less for efficiency, maybe just randomly sample several steps. I am still trying to understand diffusion. Correct me if I am wrong. \n\nSo far my fine-tuned Protenix model fails to outperform the original checkpoint. I am trying to understand why. If you have any thoughts, it would be great to hear from you.", "votes": null }, { "id": "3188599", "postDate": "04/27/2025 23:15:38", "content": "I managed to improve the ptm scores in the output file’s confidence scores for CASP15 targets, but the actual TM-scores calculated on Kaggle using C1 atoms did not improve significantly. It might be similar to how relaxation in RhoFold+ doesn’t necessarily enhance the TM-score, even though it performs energy minimization.
", "rawMarkdown": "I managed to improve the ptm scores in the output file’s confidence scores for CASP15 targets, but the actual TM-scores calculated on Kaggle using C1 atoms did not improve significantly. It might be similar to how relaxation in RhoFold+ doesn’t necessarily enhance the TM-score, even though it performs energy minimization.", "votes": null }, { "id": "3191637", "postDate": "05/02/2025 00:07:34", "content": "Protenix is for proteins, which are nothing to do with this task. It will unlikely help a lot for the RNAs, AFAIK.
", "rawMarkdown": "Protenix is for proteins, which are nothing to do with this task. It will unlikely help a lot for the RNAs, AFAIK.", "votes": null }, { "id": "3194066", "postDate": "05/05/2025 11:31:54", "content": "The finetuned Protenix model which failed did you use MSA or without MSA?
", "rawMarkdown": "The finetuned Protenix model which failed did you use MSA or without MSA?", "votes": null }, { "id": "3194069", "postDate": "05/05/2025 11:34:31", "content": "Without MSA
", "rawMarkdown": "Without MSA", "votes": null }, { "id": "3203870", "postDate": "05/17/2025 12:17:56", "content": "Hi, bro! Thank you for sharing!
\nI'm confused about how to fine-tune Protenix.
\nWhen you fine-tuning, what did you refer to or base your approach on?
\nCould you please tell me?
I just substitute protein's data for Kaggle's dataset, and run almost same code (finetune_demo.sh)
", "rawMarkdown": "I just substitute protein's data for Kaggle's dataset, and run almost same code (finetune_demo.sh)", "votes": null }, { "id": "3204207", "postDate": "05/17/2025 23:35:37", "content": "Thank you! finetune_demo.sh is from Github?
", "rawMarkdown": "Thank you! finetune_demo.sh is from Github?", "votes": null }, { "id": "3204584", "postDate": "05/18/2025 14:04:39", "content": "I've tried to do that, but I faced C-level memory errors while running prepare_training_data.py and assertion errors while I was trying to finetune. I was using CIF files for prepare_training_data.py. Did you do the same?
", "rawMarkdown": "I've tried to do that, but I faced C-level memory errors while running prepare_training_data.py and assertion errors while I was trying to finetune. I was using CIF files for prepare_training_data.py. Did you do the same?", "votes": null }, { "id": "3204588", "postDate": "05/18/2025 14:10:04", "content": "Yes, I used PDB_RNA datas that competition hosts gave
", "rawMarkdown": "Yes, I used PDB_RNA datas that competition hosts gave", "votes": null }, { "id": "3204590", "postDate": "05/18/2025 14:10:34", "content": "Yep. It's from their GitHub
", "rawMarkdown": "Yep. It's from their GitHub", "votes": null }, { "id": "3207668", "postDate": "05/23/2025 05:39:43", "content": "Combining the pr and tr is useful maybe
", "rawMarkdown": "Combining the pr and tr is useful maybe", "votes": null } ], "comments": [ { "id": 3188058, "author_name": "doheon114", "author_url": "", "post_date": "04/27/2025 01:49:22", "content": "*It's fine-tuned with no-MSA
", "votes": null, "replies": [] }, { "id": 3188348, "author_name": "zoushuxian", "author_url": "", "post_date": "04/27/2025 13:02:46", "content": "Same as yours, but I don't think it is a problem. I think the MSE loss is computed from the diffusion steps. In training and inference, the diffusion is different. In inference, the diffusion steps are 200, and each step will generate an MSE loss and accumulate to be a final MSE loss. But during training, I think it's much less for efficiency, maybe just randomly sample several steps. I am still trying to understand diffusion. Correct me if I am wrong.
\nSo far my fine-tuned Protenix model fails to outperform the original checkpoint. I am trying to understand why. If you have any thoughts, it would be great to hear from you.
", "votes": null, "replies": [ { "id": 3188599, "author_name": "doheon114", "author_url": "", "post_date": "04/27/2025 23:15:38", "content": "I managed to improve the ptm scores in the output file’s confidence scores for CASP15 targets, but the actual TM-scores calculated on Kaggle using C1 atoms did not improve significantly. It might be similar to how relaxation in RhoFold+ doesn’t necessarily enhance the TM-score, even though it performs energy minimization.
", "votes": null, "replies": [] }, { "id": 3194066, "author_name": "siddhantoon", "author_url": "", "post_date": "05/05/2025 11:31:54", "content": "The finetuned Protenix model which failed did you use MSA or without MSA?
", "votes": null, "replies": [ { "id": 3194069, "author_name": "doheon114", "author_url": "", "post_date": "05/05/2025 11:34:31", "content": "Without MSA
", "votes": null, "replies": [] } ] } ] }, { "id": 3191637, "author_name": "ilyakupchenko", "author_url": "", "post_date": "05/02/2025 00:07:34", "content": "Protenix is for proteins, which are nothing to do with this task. It will unlikely help a lot for the RNAs, AFAIK.
", "votes": null, "replies": [] }, { "id": 3203870, "author_name": "taiki2", "author_url": "", "post_date": "05/17/2025 12:17:56", "content": "Hi, bro! Thank you for sharing!
\nI'm confused about how to fine-tune Protenix.
\nWhen you fine-tuning, what did you refer to or base your approach on?
\nCould you please tell me?
I just substitute protein's data for Kaggle's dataset, and run almost same code (finetune_demo.sh)
", "votes": null, "replies": [ { "id": 3204207, "author_name": "taiki2", "author_url": "", "post_date": "05/17/2025 23:35:37", "content": "Thank you! finetune_demo.sh is from Github?
", "votes": null, "replies": [ { "id": 3204590, "author_name": "doheon114", "author_url": "", "post_date": "05/18/2025 14:10:34", "content": "Yep. It's from their GitHub
", "votes": null, "replies": [] } ] }, { "id": 3204584, "author_name": "ambisinistra", "author_url": "", "post_date": "05/18/2025 14:04:39", "content": "I've tried to do that, but I faced C-level memory errors while running prepare_training_data.py and assertion errors while I was trying to finetune. I was using CIF files for prepare_training_data.py. Did you do the same?
", "votes": null, "replies": [ { "id": 3204588, "author_name": "doheon114", "author_url": "", "post_date": "05/18/2025 14:10:04", "content": "Yes, I used PDB_RNA datas that competition hosts gave
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 3207668, "author_name": "", "author_url": "", "post_date": "05/23/2025 05:39:43", "content": "Combining the pr and tr is useful maybe
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3188053": "\nHave you ever tried fine-tuning Protenix with training data?\nI'm currently attempting it, but my experiment is encountering an unsolvable error.\nAs you can see in the image, the gap between the training set's MSE loss and the validation set's (CASP15) MSE loss is extremely large.\nFor anyone who has experience fine-tuning Protenix, how did you address this issue?\"", "3188058": "*It's fine-tuned with no-MSA", "3188348": "Same as yours, but I don't think it is a problem. I think the MSE loss is computed from the diffusion steps. In training and inference, the diffusion is different. In inference, the diffusion steps are 200, and each step will generate an MSE loss and accumulate to be a final MSE loss. But during training, I think it's much less for efficiency, maybe just randomly sample several steps. I am still trying to understand diffusion. Correct me if I am wrong. \n\nSo far my fine-tuned Protenix model fails to outperform the original checkpoint. I am trying to understand why. If you have any thoughts, it would be great to hear from you.", "3188599": "I managed to improve the ptm scores in the output file’s confidence scores for CASP15 targets, but the actual TM-scores calculated on Kaggle using C1 atoms did not improve significantly. It might be similar to how relaxation in RhoFold+ doesn’t necessarily enhance the TM-score, even though it performs energy minimization.", "3191637": "Protenix is for proteins, which are nothing to do with this task. It will unlikely help a lot for the RNAs, AFAIK.", "3194066": "The finetuned Protenix model which failed did you use MSA or without MSA?", "3194069": "Without MSA", "3203870": "Hi, bro! Thank you for sharing!\nI'm confused about how to fine-tune Protenix. \nWhen you fine-tuning, what did you refer to or base your approach on? \nCould you please tell me?", "3203910": "I just substitute protein's data for Kaggle's dataset, and run almost same code (finetune_demo.sh)", "3204207": "Thank you! finetune_demo.sh is from Github?", "3204584": "I've tried to do that, but I faced C-level memory errors while running prepare_training_data.py and assertion errors while I was trying to finetune. I was using CIF files for prepare_training_data.py. Did you do the same?", "3204588": "Yes, I used PDB_RNA datas that competition hosts gave", "3204590": "Yep. It's from their GitHub", "3207668": "Combining the pr and tr is useful maybe" }, "source": "meta" }