![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3641623%2Fbb6efccfdef259343c3ccd975bfc4fe4%2Fseg_lice.png?generation=1660206981318165&alt=media\")
actually the code is not important. it is the pretrain model that they provide.
\ni do not want to spend time and compute making imagenet pretrain model.
@hengck23 yes I meant to say all pt weights are under a non-commercial license now
", "rawMarkdown": "@hengck23 yes I meant to say all pt weights are under a non-commercial license now ", "votes": 1 }, { "id": 1906600, "postDate": "2022-08-20T03:50:07.883Z", "rawMarkdown": "", "isDeleted": true }, { "id": 1906603, "postDate": "2022-08-20T03:52:34.657Z", "content": "i think segformer cannot be used if you want to claim for prize (both ranking and judges prize).
", "rawMarkdown": "i think segformer cannot be used if you want to claim for prize (both ranking and judges prize).", "votes": 2 }, { "id": 1906609, "postDate": "2022-08-20T04:10:27.500Z", "content": "true . . . I deleted comment . . seemed to be a stupid question by me 😂, thanks for the quick reply.
\n\n", "rawMarkdown": "true . . . I deleted comment . . seemed to be a stupid question by me 😂, thanks for the quick reply.\n> (for context of the answer) > I asked if I could use segformer with training some images and making it public for commercial use? \n", "votes": 1 } ] }, { "id": 1894293, "postDate": "2022-08-11T11:38:06.833Z", "content": "(for context of the answer) > I asked if I could use segformer with training some images and making it public for commercial use?
\n
I already imagined this would happen. That's why I'm not using
", "rawMarkdown": "I already imagined this would happen. That's why I'm not using", "votes": 1 }, { "id": 1894139, "postDate": "2022-08-11T09:46:57.610Z", "content": "any convolution positional encoding (using conv2d) will have similar performers as mix-transformer Mit.
\ni have tried a couple of them. i will provide a list of substitution and results later
\nas for the decoder, just concate + aspp (or conv).
", "rawMarkdown": "any convolution positional encoding (using conv2d) will have similar performers as mix-transformer Mit. \n\ni have tried a couple of them. i will provide a list of substitution and results later\n\nas for the decoder, just concate + aspp (or conv).\n\n", "votes": 1, "replies": [ { "id": 1894537, "postDate": "2022-08-11T15:04:14.930Z", "rawMarkdown": "", "isDeleted": true }, { "id": 1894547, "postDate": "2022-08-11T15:08:07.737Z", "content": "Are there any models with pretrained weights that use a segformer like mix-ffn to replace positional encodings?
\nOr are you suggesting to replace PE in other transformer models with mix-ffn?
recommned replacement
\n![\"https://i.ibb.co/3s93QwN/Selection-075.png\"](\"https://i.ibb.co/3s93QwN/Selection-075.png\")
\nfor model, refer to https://www.kaggle.com/code/hengck23/lb-0-78-coat-and-simple-3x3-conv-fusion
\nweights/training are not shared
PVT v2 also works the same as segformer. i gave smiliar results on local CV
\n![\"Selection-071\"](\"https://i.ibb.co/2t52NLc/Selection-071.png\")
![\"Selection-072\"](\"https://i.ibb.co/L1bDwBP/Selection-072.png\")
\n\n",
"votes": 2
},
{
"id": 1906602,
"postDate": "2022-08-20T03:52:27.517Z",
"content": "Thanks!! , this is cool and will use CoAT one instead the Segformer.
", "rawMarkdown": "Thanks!! , this is cool and will use CoAT one instead the Segformer." }, { "id": 1906604, "postDate": "2022-08-20T03:55:53.043Z", "content": "if you want to to use COAT, i suggest the parallel branch version
\nhttps://www.kaggle.com/code/hengck23/lb-0-80-coat-small-parallel-at-1024
if you want to stick to segformer or daformer head, then try PVT v2
\nhttps://www.kaggle.com/code/hengck23/lb-0-78-coat-with-no-decoder
I used ensemble of segformer models and it worked well.
\nDoes segformer in mmsegmentation also have license issue?
\nThere was little time to switch to another model when I notice there is a possibility of prize contender☹️
I used following pretrained models:
\nhttps://github.com/open-mmlab/mmsegmentation/tree/master/configs/segformer
unfortunately yes.
\nmmsegmentation has updated segformer license too.
\nbut you can discuss with the competition host or NVIDIA
\nhttps://github.com/open-mmlab/mmsegmentation/pull/1699
\n![\"Selection-319\"](\"https://i.ibb.co/HFFjbvt/Selection-319.png\")
on a side note, if segformer works well, it means that position encoding in vision transformer is the a drawback when:
\nthis is unlike other competition like COC, Kitti where objects are placed on the ground. someone can do an experiment on shifted COCO images on released public models)
", "rawMarkdown": "unfortunately yes.\n\nmmsegmentation has updated segformer license too.\n\nbut you can discuss with the competition host or NVIDIA\n\nhttps://github.com/open-mmlab/mmsegmentation/pull/1699\n\n\n---\n\non a side note, if segformer works well, it means that position encoding in vision transformer is the a drawback when:\n1. there is little data\n2. absolute position has little/confusing information in this competition. \n\nthis is unlike other competition like COC, Kitti where objects are placed on the ground. someone can do an experiment on shifted COCO images on released public models)",
"votes": 3
},
{
"id": 1951285,
"postDate": "2022-09-23T01:13:12.797Z",
"content": "Hmm…
\nThen, is it nothing to do with the rank of this competition?
\ni.e. possibility of derank
i think it will have to depends on the decision of kaggle and competition host.
\nthey will contact you.
most likely you will keep your medals and rank but not the prize. but it depends on how you present and explain your case, e.g.
\nwhy you did not check licensing …?
LESSON LEARNED
\nlife is a miracle if you prepared for it
", "rawMarkdown": "i think it will have to depends on the decision of kaggle and competition host.\nthey will contact you.\n\nmost likely you will keep your medals and rank but not the prize. but it depends on how you present and explain your case, e.g.\nwhy you did not check licensing ...? \n- e.g. because you are unaware of recent license change? etc ...\n- e.g. because you blindly followed some public notebook.\n\n\nLESSON LEARNED\n\nlife is a miracle if you prepared for it", "votes": 2 }, { "id": 1951292, "postDate": "2022-09-23T01:27:07.643Z", "content": "thanks for your explanation
\nwell, it would be a hard task…
@opusen in your defense -
\nHugging face changed the liscence half way through this competition .. so you could argue that when the competition started it was under apache 2.0 .
@hengck23 has a screenshot of this somewhere on his discussion page
", "rawMarkdown": "@hengck23 has a screenshot of this somewhere on his discussion page ", "votes": 1 }, { "id": 1952680, "postDate": "2022-09-23T20:18:53.063Z", "content": "just tell the truth and hope for the best.
\nhonesty is the best policy.
there is a chance that \"HuBMAP + HPA\" is open sourcing the solution and maybe it is not considered as 'commercial'.
\n\"HuBMAP + HPA\" are non profit organization.
@hengck23 @rarun2596 Thank you very much. It's really encouraging.
\nI will discuss with the competition host and kaggle :)
actually the code is not important. it is the pretrain model that they provide.
\ni do not want to spend time and compute making imagenet pretrain model.
@hengck23 yes I meant to say all pt weights are under a non-commercial license now
", "votes": 1, "replies": [] }, { "id": 1906600, "author_name": "", "author_url": "", "post_date": "2022-08-20T03:50:07.883000", "content": "", "votes": 0, "replies": [] }, { "id": 1906603, "author_name": "hengck23", "author_url": "", "post_date": "2022-08-20T03:52:34.657000", "content": "i think segformer cannot be used if you want to claim for prize (both ranking and judges prize).
", "votes": 2, "replies": [] }, { "id": 1906609, "author_name": "Bibhabasu Mohapatra", "author_url": "", "post_date": "2022-08-20T04:10:27.500000", "content": "true . . . I deleted comment . . seemed to be a stupid question by me 😂, thanks for the quick reply.
\n\n", "votes": 1, "replies": [] } ] }, { "id": 1894293, "author_name": "Robson", "author_url": "", "post_date": "2022-08-11T11:38:06.833000", "content": "(for context of the answer) > I asked if I could use segformer with training some images and making it public for commercial use?
\n
I already imagined this would happen. That's why I'm not using
", "votes": 1, "replies": [] }, { "id": 1894139, "author_name": "hengck23", "author_url": "", "post_date": "2022-08-11T09:46:57.610000", "content": "any convolution positional encoding (using conv2d) will have similar performers as mix-transformer Mit.
\ni have tried a couple of them. i will provide a list of substitution and results later
\nas for the decoder, just concate + aspp (or conv).
", "votes": 1, "replies": [ { "id": 1894537, "author_name": "", "author_url": "", "post_date": "2022-08-11T15:04:14.930000", "content": "", "votes": 0, "replies": [] }, { "id": 1894547, "author_name": "_CA℟L_", "author_url": "", "post_date": "2022-08-11T15:08:07.737000", "content": "Are there any models with pretrained weights that use a segformer like mix-ffn to replace positional encodings?
\nOr are you suggesting to replace PE in other transformer models with mix-ffn?
recommned replacement
\n
\nfor model, refer to https://www.kaggle.com/code/hengck23/lb-0-78-coat-and-simple-3x3-conv-fusion
\nweights/training are not shared
PVT v2 also works the same as segformer. i gave smiliar results on local CV
\n
Thanks!! , this is cool and will use CoAT one instead the Segformer.
", "votes": 0, "replies": [] }, { "id": 1906604, "author_name": "hengck23", "author_url": "", "post_date": "2022-08-20T03:55:53.043000", "content": "if you want to to use COAT, i suggest the parallel branch version
\nhttps://www.kaggle.com/code/hengck23/lb-0-80-coat-small-parallel-at-1024
if you want to stick to segformer or daformer head, then try PVT v2
\nhttps://www.kaggle.com/code/hengck23/lb-0-78-coat-with-no-decoder
I used ensemble of segformer models and it worked well.
\nDoes segformer in mmsegmentation also have license issue?
\nThere was little time to switch to another model when I notice there is a possibility of prize contender☹️
I used following pretrained models:
\nhttps://github.com/open-mmlab/mmsegmentation/tree/master/configs/segformer
unfortunately yes.
\nmmsegmentation has updated segformer license too.
\nbut you can discuss with the competition host or NVIDIA
\nhttps://github.com/open-mmlab/mmsegmentation/pull/1699
\n
on a side note, if segformer works well, it means that position encoding in vision transformer is the a drawback when:
\nthis is unlike other competition like COC, Kitti where objects are placed on the ground. someone can do an experiment on shifted COCO images on released public models)
", "votes": 3, "replies": [] }, { "id": 1951285, "author_name": "opusen", "author_url": "", "post_date": "2022-09-23T01:13:12.797000", "content": "Hmm…
\nThen, is it nothing to do with the rank of this competition?
\ni.e. possibility of derank
i think it will have to depends on the decision of kaggle and competition host.
\nthey will contact you.
most likely you will keep your medals and rank but not the prize. but it depends on how you present and explain your case, e.g.
\nwhy you did not check licensing …?
LESSON LEARNED
\nlife is a miracle if you prepared for it
", "votes": 2, "replies": [] }, { "id": 1951292, "author_name": "opusen", "author_url": "", "post_date": "2022-09-23T01:27:07.643000", "content": "thanks for your explanation
\nwell, it would be a hard task…
@opusen in your defense -
\nHugging face changed the liscence half way through this competition .. so you could argue that when the competition started it was under apache 2.0 .
@hengck23 has a screenshot of this somewhere on his discussion page
", "votes": 1, "replies": [] }, { "id": 1952680, "author_name": "hengck23", "author_url": "", "post_date": "2022-09-23T20:18:53.063000", "content": "just tell the truth and hope for the best.
\nhonesty is the best policy.
there is a chance that \"HuBMAP + HPA\" is open sourcing the solution and maybe it is not considered as 'commercial'.
\n\"HuBMAP + HPA\" are non profit organization.
@hengck23 @rarun2596 Thank you very much. It's really encouraging.
\nI will discuss with the competition host and kaggle :)