Thank you for running such a fun competition. I'm enjoying it!
\nI have tried fine-tuning various models and came across this topics ref1.
\nYou stated that it is acceptable to use pre-trained models if they can be reproduced, but the rules of the competition(8-c) ref2 state that in no event limits commercial use of such code or model.
\nTherefore, is it correct to understand that we can only use pre-trained models that are both reproducible and available for commercial use?
As an example, there is the huggingface nvidia segformer ref3, which is currently not available for commercial use ref4 (See 3.3)
\n, but in a past competition, it was allowed as an exception because the license was changed during the competition ref5. However, for this competition, this model cannot be used, correct?
→ I have made the correction. It is also believed that segformer (Mix vision transformer) within segmentation-model-pytorch cannot be used for commercial purposes. Although segmentation-model-pytorch is licensed under the MIT license, there is a license for NVIDIA within the Mix Vision Transformer in it. ref8
\n※ It was stated in the comments. Thank you very much !!
To summarize, is the following correct?
\nTo avoid confusion after the competition, I would appreciate it if you could confirm my understanding. Thank you.
", "messages": [ { "id": 2247605, "postDate": "2023-05-06T06:35:21.623Z", "content": "\nThank you for running such a fun competition. I'm enjoying it!
\nI have tried fine-tuning various models and came across this topics ref1.
\nYou stated that it is acceptable to use pre-trained models if they can be reproduced, but the rules of the competition(8-c) ref2 state that in no event limits commercial use of such code or model.
\nTherefore, is it correct to understand that we can only use pre-trained models that are both reproducible and available for commercial use?
As an example, there is the huggingface nvidia segformer ref3, which is currently not available for commercial use ref4 (See 3.3)
\n, but in a past competition, it was allowed as an exception because the license was changed during the competition ref5. However, for this competition, this model cannot be used, correct?
→ I have made the correction. It is also believed that segformer (Mix vision transformer) within segmentation-model-pytorch cannot be used for commercial purposes. Although segmentation-model-pytorch is licensed under the MIT license, there is a license for NVIDIA within the Mix Vision Transformer in it. ref8
\n※ It was stated in the comments. Thank you very much !!
To summarize, is the following correct?
\nTo avoid confusion after the competition, I would appreciate it if you could confirm my understanding. Thank you.
", "rawMarkdown": "@jpposma \n\nThank you for running such a fun competition. I'm enjoying it!\n\nI have tried fine-tuning various models and came across this topics [ref1](https://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/discussion/399140).\n\nYou stated that it is acceptable to use pre-trained models if they can be reproduced, but the rules of the competition(8-c) [ref2](https://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/rules) state that in no event limits commercial use of such code or model.\nTherefore, is it correct to understand that we can only use pre-trained models that are both reproducible and available for commercial use?\n\n\n\nAs an example, there is the huggingface nvidia segformer [ref3](https://huggingface.co/docs/transformers/model_doc/segformer), which is currently not available for commercial use [ref4](https://github.com/NVlabs/SegFormer/blob/master/LICENSE) (See 3.3)\n, but in a past competition, it was allowed as an exception because the license was changed during the competition [ref5](https://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/356201). However, for this competition, this model cannot be used, correct?\n\n~~On the other hand, the imagenet pre-trained segformer from the segmentation-model-pytorch(see Mix Vision Transformer) is believed under the MIT license [ref6](https://github.com/qubvel/segmentation_models.pytorch), so it should be available for use.\nHowever, what I don't understand is that Hugging Face also has pre-trained models for NVIDIA's ImageNet SegFormer [ref7](https://huggingface.co/nvidia/mit-b1), which like before, cannot be used for commercial purposes. I am not sure if these are the same models. Perhaps, since the license for segmentation-model-pytorch is MIT, it should be available for use.~~\n\n→ I have made the correction. It is also believed that segformer (Mix vision transformer) within segmentation-model-pytorch cannot be used for commercial purposes. Although segmentation-model-pytorch is licensed under the MIT license, there is a license for NVIDIA within the Mix Vision Transformer in it. [ref8] (https://github.com/qubvel/segmentation_models.pytorch/blob/master/segmentation_models_pytorch/encoders/mix_transformer.py)\n※ It was stated in the comments. Thank you very much !!\n\nTo summarize, is the following correct?\n\n1. The huggingface nvidia segformer cannot be used. (even if it can be reproduced)\n2. ~~The segformer from the segmentation-model-pytorch can be used. (maybe this is MIT license)~~\n→ I have made the correction. Segformer(Mix Vision Transformer) also cannot be used from segmentation-model-pytorch.\n\nTo avoid confusion after the competition, I would appreciate it if you could confirm my understanding. Thank you.\n", "votes": 22 }, { "id": 2247787, "postDate": "2023-05-06T09:19:29.773Z", "content": "if you are using seformer, pvt2 is a good replacement
\nhttps://arxiv.org/abs/2106.13797
\nPVT v2: Improved Baselines with Pyramid Vision Transformer
if you check their papers, backbone MIT and PVTv2 are 98% the same. (mainly the convolutional postional encoding)
", "rawMarkdown": "if you are using seformer, pvt2 is a good replacement\n\nhttps://arxiv.org/abs/2106.13797\nPVT v2: Improved Baselines with Pyramid Vision Transformer\n\nif you check their papers, backbone MIT and PVTv2 are 98% the same. (mainly the convolutional postional encoding)", "votes": 10, "replies": [ { "id": 2248290, "postDate": "2023-05-06T17:26:53.330Z", "content": "@hengck23
\nThank you very much! Pyramid Vision Transformer looks great. I've never heard of it before! I'll give it a try.
Sorry that it took a while but we have an answer for you. You ARE allowed to use libraries that have a non-commercial clause in their license. https://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/discussion/410906
", "rawMarkdown": "Sorry that it took a while but we have an answer for you. You ARE allowed to use libraries that have a non-commercial clause in their license. https://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/discussion/410906", "votes": 7, "replies": [ { "id": 2262634, "postDate": "2023-05-17T04:11:34.587Z", "content": "@jpposma
\nThank you for making it clear! With this, there should be no dispute over this matter even after the competition is over. Moreover, this competition will become even more enjoyable.
Sorry for the delay, I totally missed this. I'm looking into this with @ryanholbrook and @wcukierski.
", "rawMarkdown": "Sorry for the delay, I totally missed this. I'm looking into this with @ryanholbrook and @wcukierski.", "votes": 3, "replies": [ { "id": 2257504, "postDate": "2023-05-13T12:07:30.430Z", "content": "@jpposma
\nThank you for your response! I feel relieved. Yes, please consult with the staff at Kaggle. I think it may also affect everyone's strategy. Additionally, we would like the competition to end peacefully and without any trouble after it's over.
@wcukierski , @ryanholbrook
\nThank you always. Could you please check this and give us the comment? Thank you in advance.
seems so clear, thnx for view
", "rawMarkdown": "seems so clear, thnx for view", "votes": 1 }, { "id": 2250539, "postDate": "2023-05-08T15:58:46.160Z", "content": "@jpposma any updates on the topic? do you have any final decision?
", "rawMarkdown": "@jpposma any updates on the topic? do you have any final decision?", "votes": 1 }, { "id": 2249539, "postDate": "2023-05-07T20:16:49.653Z", "content": "You should be able to replicate the MLP decoder they use in segformer but attach it to the Swin implementation found in Pytorch. I've found that the performance is the same even though this implementation isn't pretrained on a segmentation task. This also wouldn't violate the license as the code isn't copied directly.
", "rawMarkdown": "You should be able to replicate the MLP decoder they use in segformer but attach it to the Swin implementation found in Pytorch. I've found that the performance is the same even though this implementation isn't pretrained on a segmentation task. This also wouldn't violate the license as the code isn't copied directly. ", "votes": 2, "replies": [ { "id": 2250481, "postDate": "2023-05-08T15:23:28.693Z", "content": "@petersk20 Thank you for sharing your experience. If there are no issues with the license of the MLP decoder itself, that is a good idea, approach and results ! I also want to try challenging myself with something like that.
", "rawMarkdown": "@petersk20 Thank you for sharing your experience. If there are no issues with the license of the MLP decoder itself, that is a good idea, approach and results ! I also want to try challenging myself with something like that." } ] }, { "id": 2247910, "postDate": "2023-05-06T11:08:57.210Z", "content": "Although segmentation-model-pytorch is under MIT License, Mix Vision Transformer has the custom NVIDIA License.
\nhttps://github.com/qubvel/segmentation_models.pytorch/blob/master/segmentation_models_pytorch/encoders/mix_transformer.py
\nSo this cannot be used if you are aiming for the prizes.
As mentioned by @hengck23, pvt is a better alternative with Apache-2 License.
@samfc10
\nThank you very much. You are right. I should have confirmed that earlier. I fixed the topic!
\nHowever, it seems easy to misuse this, as discussed here.
@jpposma , staff
\nJust to be sure, could you also provide a comment from the host? (to avoid any trouble after the competition ends)
It seems Mix Vision Transformer in SMP can only be used with 3 input channels
", "rawMarkdown": "It seems Mix Vision Transformer in SMP can only be used with 3 input channels", "votes": 1 } ] } ] }, { "id": 2248934, "postDate": "2023-05-07T11:02:01.287Z", "rawMarkdown": "", "votes": 1, "isDeleted": true } ], "comments": [ { "id": 2247787, "author_name": "hengck23", "author_url": "", "post_date": "2023-05-06T09:19:29.773000", "content": "if you are using seformer, pvt2 is a good replacement
\nhttps://arxiv.org/abs/2106.13797
\nPVT v2: Improved Baselines with Pyramid Vision Transformer
if you check their papers, backbone MIT and PVTv2 are 98% the same. (mainly the convolutional postional encoding)
", "votes": 10, "replies": [ { "id": 2248290, "author_name": "chumajin", "author_url": "", "post_date": "2023-05-06T17:26:53.330000", "content": "@hengck23
\nThank you very much! Pyramid Vision Transformer looks great. I've never heard of it before! I'll give it a try.
Sorry that it took a while but we have an answer for you. You ARE allowed to use libraries that have a non-commercial clause in their license. https://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/discussion/410906
", "votes": 7, "replies": [ { "id": 2262634, "author_name": "chumajin", "author_url": "", "post_date": "2023-05-17T04:11:34.587000", "content": "@jpposma
\nThank you for making it clear! With this, there should be no dispute over this matter even after the competition is over. Moreover, this competition will become even more enjoyable.
Sorry for the delay, I totally missed this. I'm looking into this with @ryanholbrook and @wcukierski.
", "votes": 3, "replies": [ { "id": 2257504, "author_name": "chumajin", "author_url": "", "post_date": "2023-05-13T12:07:30.430000", "content": "@jpposma
\nThank you for your response! I feel relieved. Yes, please consult with the staff at Kaggle. I think it may also affect everyone's strategy. Additionally, we would like the competition to end peacefully and without any trouble after it's over.
@wcukierski , @ryanholbrook
\nThank you always. Could you please check this and give us the comment? Thank you in advance.
seems so clear, thnx for view
", "votes": 1, "replies": [] }, { "id": 2250539, "author_name": "Kostiantyn Maksymov", "author_url": "", "post_date": "2023-05-08T15:58:46.160000", "content": "@jpposma any updates on the topic? do you have any final decision?
", "votes": 1, "replies": [] }, { "id": 2249539, "author_name": "Kyle Peters", "author_url": "", "post_date": "2023-05-07T20:16:49.653000", "content": "You should be able to replicate the MLP decoder they use in segformer but attach it to the Swin implementation found in Pytorch. I've found that the performance is the same even though this implementation isn't pretrained on a segmentation task. This also wouldn't violate the license as the code isn't copied directly.
", "votes": 2, "replies": [ { "id": 2250481, "author_name": "chumajin", "author_url": "", "post_date": "2023-05-08T15:23:28.693000", "content": "@petersk20 Thank you for sharing your experience. If there are no issues with the license of the MLP decoder itself, that is a good idea, approach and results ! I also want to try challenging myself with something like that.
", "votes": 0, "replies": [] } ] }, { "id": 2247910, "author_name": "Jebastin Nadar", "author_url": "", "post_date": "2023-05-06T11:08:57.210000", "content": "Although segmentation-model-pytorch is under MIT License, Mix Vision Transformer has the custom NVIDIA License.
\nhttps://github.com/qubvel/segmentation_models.pytorch/blob/master/segmentation_models_pytorch/encoders/mix_transformer.py
\nSo this cannot be used if you are aiming for the prizes.
As mentioned by @hengck23, pvt is a better alternative with Apache-2 License.
@samfc10
\nThank you very much. You are right. I should have confirmed that earlier. I fixed the topic!
\nHowever, it seems easy to misuse this, as discussed here.
@jpposma , staff
\nJust to be sure, could you also provide a comment from the host? (to avoid any trouble after the competition ends)
It seems Mix Vision Transformer in SMP can only be used with 3 input channels
", "votes": 1, "replies": [] } ] } ] }, { "id": 2248934, "author_name": "", "author_url": "", "post_date": "2023-05-07T11:02:01.287000", "content": "", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "2247605": "@jpposma \n\nThank you for running such a fun competition. I'm enjoying it!\n\nI have tried fine-tuning various models and came across this topics [ref1](https://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/discussion/399140).\n\nYou stated that it is acceptable to use pre-trained models if they can be reproduced, but the rules of the competition(8-c) [ref2](https://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/rules) state that in no event limits commercial use of such code or model.\nTherefore, is it correct to understand that we can only use pre-trained models that are both reproducible and available for commercial use?\n\n\n\nAs an example, there is the huggingface nvidia segformer [ref3](https://huggingface.co/docs/transformers/model_doc/segformer), which is currently not available for commercial use [ref4](https://github.com/NVlabs/SegFormer/blob/master/LICENSE) (See 3.3)\n, but in a past competition, it was allowed as an exception because the license was changed during the competition [ref5](https://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/356201). However, for this competition, this model cannot be used, correct?\n\n~~On the other hand, the imagenet pre-trained segformer from the segmentation-model-pytorch(see Mix Vision Transformer) is believed under the MIT license [ref6](https://github.com/qubvel/segmentation_models.pytorch), so it should be available for use.\nHowever, what I don't understand is that Hugging Face also has pre-trained models for NVIDIA's ImageNet SegFormer [ref7](https://huggingface.co/nvidia/mit-b1), which like before, cannot be used for commercial purposes. I am not sure if these are the same models. Perhaps, since the license for segmentation-model-pytorch is MIT, it should be available for use.~~\n\n→ I have made the correction. It is also believed that segformer (Mix vision transformer) within segmentation-model-pytorch cannot be used for commercial purposes. Although segmentation-model-pytorch is licensed under the MIT license, there is a license for NVIDIA within the Mix Vision Transformer in it. [ref8] (https://github.com/qubvel/segmentation_models.pytorch/blob/master/segmentation_models_pytorch/encoders/mix_transformer.py)\n※ It was stated in the comments. Thank you very much !!\n\nTo summarize, is the following correct?\n\n1. The huggingface nvidia segformer cannot be used. (even if it can be reproduced)\n2. ~~The segformer from the segmentation-model-pytorch can be used. (maybe this is MIT license)~~\n→ I have made the correction. Segformer(Mix Vision Transformer) also cannot be used from segmentation-model-pytorch.\n\nTo avoid confusion after the competition, I would appreciate it if you could confirm my understanding. Thank you.\n", "2247787": "if you are using seformer, pvt2 is a good replacement\n\nhttps://arxiv.org/abs/2106.13797\nPVT v2: Improved Baselines with Pyramid Vision Transformer\n\nif you check their papers, backbone MIT and PVTv2 are 98% the same. (mainly the convolutional postional encoding)", "2262413": "Sorry that it took a while but we have an answer for you. You ARE allowed to use libraries that have a non-commercial clause in their license. https://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/discussion/410906", "2256817": "Sorry for the delay, I totally missed this. I'm looking into this with @ryanholbrook and @wcukierski.", "2257851": "seems so clear, thnx for view", "2250539": "@jpposma any updates on the topic? do you have any final decision?", "2249539": "You should be able to replicate the MLP decoder they use in segformer but attach it to the Swin implementation found in Pytorch. I've found that the performance is the same even though this implementation isn't pretrained on a segmentation task. This also wouldn't violate the license as the code isn't copied directly. ", "2247910": "Although `segmentation-model-pytorch` is under MIT License, Mix Vision Transformer has the custom NVIDIA License.\nhttps://github.com/qubvel/segmentation_models.pytorch/blob/master/segmentation_models_pytorch/encoders/mix_transformer.py\nSo this cannot be used if you are aiming for the prizes.\n\nAs mentioned by @hengck23, `pvt` is a better alternative with Apache-2 License.", "2248934": "" } }