What is your best single-model score?
\n\nMy current best model is resnet152, one model (one fold with 4x scale TTA) gives 0.926 on public LB and 0.9349 on local validation.\nCurious to know your score and model. Has anyone tried EfficientNet?
", "messages": [ { "id": "634619", "postDate": "09/26/2019 14:08:00", "content": "What is your best single-model score?
\n\nMy current best model is resnet152, one model (one fold with 4x scale TTA) gives 0.926 on public LB and 0.9349 on local validation.\nCurious to know your score and model. Has anyone tried EfficientNet?
", "rawMarkdown": "What is your best single-model score? \n\nMy current best model is resnet152, one model (one fold with 4x scale TTA) gives 0.926 on public LB and 0.9349 on local validation.\nCurious to know your score and model. Has anyone tried EfficientNet?", "votes": null }, { "id": "635476", "postDate": "09/27/2019 16:19:04", "content": "Hi, I'm new to DL. Could you please tell me which direction you are using? I'm using the Center-net to detect the centre and the boundary first and then do the classification. I'm not sure if it's the right way.
", "rawMarkdown": "Hi, I'm new to DL. Could you please tell me which direction you are using? I'm using the Center-net to detect the centre and the boundary first and then do the classification. I'm not sure if it's the right way.", "votes": null }, { "id": "635505", "postDate": "09/27/2019 18:03:06", "content": "I'm using UNet with a ResNet backbone to segment the characters. Interested in what others are doing. @lopuhin have you tried either of these methods?
", "rawMarkdown": "I'm using UNet with a ResNet backbone to segment the characters. Interested in what others are doing. @lopuhin have you tried either of these methods?", "votes": null }, { "id": "635512", "postDate": "09/27/2019 18:16:27", "content": "I'm also doing detection and classification separately - it's possible to do the same with one model, but here I feel that classification requires much more attention and tuning, while detection is relatively straightforward, and it's easier to tune just classification model.\nFor detection I'm using faster-rcnn: it's usually the best scoring approach for detection tasks, although this task is quite unusual, so other approaches may be better.\nFor classification, no domain-specific things that I tried worked for me (and my domain knowledge is quite poor), but general neural network tuning goes a long way.
", "rawMarkdown": "I'm also doing detection and classification separately - it's possible to do the same with one model, but here I feel that classification requires much more attention and tuning, while detection is relatively straightforward, and it's easier to tune just classification model.\nFor detection I'm using faster-rcnn: it's usually the best scoring approach for detection tasks, although this task is quite unusual, so other approaches may be better.\nFor classification, no domain-specific things that I tried worked for me (and my domain knowledge is quite poor), but general neural network tuning goes a long way.", "votes": null }, { "id": "635584", "postDate": "09/27/2019 21:50:24", "content": "Interesting! Are you using a Faster-RCNN+backbone model and throwing away the labels? It doesn't seem like there are implementations for bounding box-only
\n\nEdit: I see, using one class would be a reasonable way to go about it
", "rawMarkdown": "Interesting! Are you using a Faster-RCNN+backbone model and throwing away the labels? It doesn't seem like there are implementations for bounding box-only\n\nEdit: I see, using one class would be a reasonable way to go about it", "votes": null }, { "id": "635691", "postDate": "09/28/2019 03:30:44", "content": "Ok, thank you for sharing. It is a long way to fine-tuning the model or experiment on a new structure.
", "rawMarkdown": "Ok, thank you for sharing. It is a long way to fine-tuning the model or experiment on a new structure.", "votes": null }, { "id": "635781", "postDate": "09/28/2019 07:42:20", "content": "\n\n\nI see, using one class would be a reasonable way to go about it
\n
Yes, that's what I'm doing.
", "rawMarkdown": "> I see, using one class would be a reasonable way to go about it\n\nYes, that's what I'm doing.", "votes": null }, { "id": "635953", "postDate": "09/28/2019 14:08:48", "content": "\n\n\nYes, that's what I'm doing.
\n
Hi, did you augment images also? since there's only one image of some classes from train images
", "rawMarkdown": "> Yes, that's what I'm doing.\n\nHi, did you augment images also? since there's only one image of some classes from train images", "votes": null }, { "id": "636317", "postDate": "09/29/2019 09:33:11", "content": "\n\n\nHi, did you augment images also? since there's only one image of some classes from train images
\n
Hi, I'm using standard augmentations from https://github.com/albu/albumentations/ library, including adjusting colors that helps simulate different paper styles. Also it seems that preventing overfitting is quite important here. I'm not doing anything special with rare classes yet.
", "rawMarkdown": "> Hi, did you augment images also? since there's only one image of some classes from train images\n\nHi, I'm using standard augmentations from https://github.com/albu/albumentations/ library, including adjusting colors that helps simulate different paper styles. Also it seems that preventing overfitting is quite important here. I'm not doing anything special with rare classes yet.", "votes": null }, { "id": "637027", "postDate": "09/30/2019 14:51:23", "content": "0.826 - detection plus classification, without validation. Using augment images (albumentations https://github.com/albu/albumentations/ ) without TTA yet.
", "rawMarkdown": "0.826 - detection plus classification, without validation. Using augment images (albumentations https://github.com/albu/albumentations/ ) without TTA yet.", "votes": null }, { "id": "646552", "postDate": "10/11/2019 12:18:50", "content": "@lopuhin did you use flips in augmentation. Intuitively, flips are not good set of augmentations for these character recognition and please can u enlighten on the set of possible augmentation to use for TTA??
", "rawMarkdown": "lopuhin did you use flips in augmentation. Intuitively, flips are not good set of augmentations for these character recognition and please can u enlighten on the set of possible augmentation to use for TTA??", "votes": null }, { "id": "646564", "postDate": "10/11/2019 12:36:47", "content": "Right, I didn't even try flips, I use only re-scaling, which makes sense as glyphs could be bigger or smaller without changing their meaning. So I average predictions from 4 different scales, from mean scale used during training up to 1.3x larger scale.
", "rawMarkdown": "Right, I didn't even try flips, I use only re-scaling, which makes sense as glyphs could be bigger or smaller without changing their meaning. So I average predictions from 4 different scales, from mean scale used during training up to 1.3x larger scale.", "votes": null } ], "comments": [ { "id": 635476, "author_name": "moriatyyuan", "author_url": "", "post_date": "09/27/2019 16:19:04", "content": "Hi, I'm new to DL. Could you please tell me which direction you are using? I'm using the Center-net to detect the centre and the boundary first and then do the classification. I'm not sure if it's the right way.
", "votes": null, "replies": [ { "id": 635505, "author_name": "usmannkhan", "author_url": "", "post_date": "09/27/2019 18:03:06", "content": "I'm using UNet with a ResNet backbone to segment the characters. Interested in what others are doing. @lopuhin have you tried either of these methods?
", "votes": null, "replies": [] }, { "id": 635512, "author_name": "lopuhin", "author_url": "", "post_date": "09/27/2019 18:16:27", "content": "I'm also doing detection and classification separately - it's possible to do the same with one model, but here I feel that classification requires much more attention and tuning, while detection is relatively straightforward, and it's easier to tune just classification model.\nFor detection I'm using faster-rcnn: it's usually the best scoring approach for detection tasks, although this task is quite unusual, so other approaches may be better.\nFor classification, no domain-specific things that I tried worked for me (and my domain knowledge is quite poor), but general neural network tuning goes a long way.
", "votes": null, "replies": [] }, { "id": 635584, "author_name": "usmannkhan", "author_url": "", "post_date": "09/27/2019 21:50:24", "content": "Interesting! Are you using a Faster-RCNN+backbone model and throwing away the labels? It doesn't seem like there are implementations for bounding box-only
\n\nEdit: I see, using one class would be a reasonable way to go about it
", "votes": null, "replies": [] }, { "id": 635691, "author_name": "moriatyyuan", "author_url": "", "post_date": "09/28/2019 03:30:44", "content": "Ok, thank you for sharing. It is a long way to fine-tuning the model or experiment on a new structure.
", "votes": null, "replies": [] }, { "id": 635781, "author_name": "lopuhin", "author_url": "", "post_date": "09/28/2019 07:42:20", "content": "\n\n\nI see, using one class would be a reasonable way to go about it
\n
Yes, that's what I'm doing.
", "votes": null, "replies": [] }, { "id": 635953, "author_name": "moximo13", "author_url": "", "post_date": "09/28/2019 14:08:48", "content": "\n\n\nYes, that's what I'm doing.
\n
Hi, did you augment images also? since there's only one image of some classes from train images
", "votes": null, "replies": [] }, { "id": 636317, "author_name": "lopuhin", "author_url": "", "post_date": "09/29/2019 09:33:11", "content": "\n\n\nHi, did you augment images also? since there's only one image of some classes from train images
\n
Hi, I'm using standard augmentations from https://github.com/albu/albumentations/ library, including adjusting colors that helps simulate different paper styles. Also it seems that preventing overfitting is quite important here. I'm not doing anything special with rare classes yet.
", "votes": null, "replies": [] }, { "id": 646552, "author_name": "sabbiracoustic1006", "author_url": "", "post_date": "10/11/2019 12:18:50", "content": "@lopuhin did you use flips in augmentation. Intuitively, flips are not good set of augmentations for these character recognition and please can u enlighten on the set of possible augmentation to use for TTA??
", "votes": null, "replies": [] }, { "id": 646564, "author_name": "lopuhin", "author_url": "", "post_date": "10/11/2019 12:36:47", "content": "Right, I didn't even try flips, I use only re-scaling, which makes sense as glyphs could be bigger or smaller without changing their meaning. So I average predictions from 4 different scales, from mean scale used during training up to 1.3x larger scale.
", "votes": null, "replies": [] } ] }, { "id": 637027, "author_name": "sorokinv", "author_url": "", "post_date": "09/30/2019 14:51:23", "content": "0.826 - detection plus classification, without validation. Using augment images (albumentations https://github.com/albu/albumentations/ ) without TTA yet.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "634619": "What is your best single-model score? \n\nMy current best model is resnet152, one model (one fold with 4x scale TTA) gives 0.926 on public LB and 0.9349 on local validation.\nCurious to know your score and model. Has anyone tried EfficientNet?", "635476": "Hi, I'm new to DL. Could you please tell me which direction you are using? I'm using the Center-net to detect the centre and the boundary first and then do the classification. I'm not sure if it's the right way.", "635505": "I'm using UNet with a ResNet backbone to segment the characters. Interested in what others are doing. @lopuhin have you tried either of these methods?", "635512": "I'm also doing detection and classification separately - it's possible to do the same with one model, but here I feel that classification requires much more attention and tuning, while detection is relatively straightforward, and it's easier to tune just classification model.\nFor detection I'm using faster-rcnn: it's usually the best scoring approach for detection tasks, although this task is quite unusual, so other approaches may be better.\nFor classification, no domain-specific things that I tried worked for me (and my domain knowledge is quite poor), but general neural network tuning goes a long way.", "635584": "Interesting! Are you using a Faster-RCNN+backbone model and throwing away the labels? It doesn't seem like there are implementations for bounding box-only\n\nEdit: I see, using one class would be a reasonable way to go about it", "635691": "Ok, thank you for sharing. It is a long way to fine-tuning the model or experiment on a new structure.", "635781": "> I see, using one class would be a reasonable way to go about it\n\nYes, that's what I'm doing.", "635953": "> Yes, that's what I'm doing.\n\nHi, did you augment images also? since there's only one image of some classes from train images", "636317": "> Hi, did you augment images also? since there's only one image of some classes from train images\n\nHi, I'm using standard augmentations from https://github.com/albu/albumentations/ library, including adjusting colors that helps simulate different paper styles. Also it seems that preventing overfitting is quite important here. I'm not doing anything special with rare classes yet.", "637027": "0.826 - detection plus classification, without validation. Using augment images (albumentations https://github.com/albu/albumentations/ ) without TTA yet.", "646552": "lopuhin did you use flips in augmentation. Intuitively, flips are not good set of augmentations for these character recognition and please can u enlighten on the set of possible augmentation to use for TTA??", "646564": "Right, I didn't even try flips, I use only re-scaling, which makes sense as glyphs could be bigger or smaller without changing their meaning. So I average predictions from 4 different scales, from mean scale used during training up to 1.3x larger scale." }, "source": "meta" }