![\"\"](\"https://media-exp1.licdn.com/dms/image/C4D22AQFuJbKRh_Oolw/feedshare-shrink_800/0?e=1597881600&v=beta&t=HbtQHwjUvETwM5CG1lmmXGlONzY_qP5r4oi3vU0UFVE\")
found 1 interesting paper today
\n\nAbstract : \nIn general, sufficient data is essential for the better perfor-mance and generalization of deep-learning models. However, lots of lim-itations(cost, resources, etc.) of data collection leads to lack of enoughdata in most of the areas. In addition, various domains of each datasources and licenses also lead to difficulties in collection of sufficientdata. This situation makes us hard to utilize not only the pre-trainedmodel, but also the external knowledge. Therefore, it is important toleverage small dataset effectively for achieving the better performance.We applied some techniques in three aspects: data, loss function, andprediction to enable training from scratch with less data. With thesemethods, we obtain high accuracy by leveraging ImageNet data whichconsist of only 50 images per class. Furthermore, our model is ranked 4thin Visual Inductive Printers for Data-Effective Computer Vision Chal-lenge.
\n\nalso check figure 2 of that paper :)
\n\npaper link : https://arxiv.org/pdf/2007.07805.pdf
", "messages": [ { "id": "934567", "postDate": "07/18/2020 14:45:51", "content": "found 1 interesting paper today
\n\nAbstract : \nIn general, sufficient data is essential for the better perfor-mance and generalization of deep-learning models. However, lots of lim-itations(cost, resources, etc.) of data collection leads to lack of enoughdata in most of the areas. In addition, various domains of each datasources and licenses also lead to difficulties in collection of sufficientdata. This situation makes us hard to utilize not only the pre-trainedmodel, but also the external knowledge. Therefore, it is important toleverage small dataset effectively for achieving the better performance.We applied some techniques in three aspects: data, loss function, andprediction to enable training from scratch with less data. With thesemethods, we obtain high accuracy by leveraging ImageNet data whichconsist of only 50 images per class. Furthermore, our model is ranked 4thin Visual Inductive Printers for Data-Effective Computer Vision Chal-lenge.
\n\nalso check figure 2 of that paper :)
\n\npaper link : https://arxiv.org/pdf/2007.07805.pdf
", "rawMarkdown": "**found 1 interesting paper today**\n\nAbstract : \nIn general, sufficient data is essential for the better perfor-mance and generalization of deep-learning models. However, lots of lim-itations(cost, resources, etc.) of data collection leads to lack of enoughdata in most of the areas. In addition, various domains of each datasources and licenses also lead to difficulties in collection of sufficientdata. This situation makes us hard to utilize not only the pre-trainedmodel, but also the external knowledge. Therefore, it is important toleverage small dataset effectively for achieving the better performance.We applied some techniques in three aspects: data, loss function, andprediction to enable training from scratch with less data. With thesemethods, we obtain high accuracy by leveraging ImageNet data whichconsist of only 50 images per class. Furthermore, our model is ranked 4thin Visual Inductive Printers for Data-Effective Computer Vision Chal-lenge.\n\n\n\nalso check figure 2 of that paper :)\n\npaper link : https://arxiv.org/pdf/2007.07805.pdf", "votes": null }, { "id": "934617", "postDate": "07/18/2020 15:32:44", "content": "Is there any snippet for focal cosine loss in tensorflow??
", "rawMarkdown": "Is there any snippet for focal cosine loss in tensorflow??", "votes": null }, { "id": "934641", "postDate": "07/18/2020 15:52:15", "content": "@msharuk589 sorry, i found pytorch implementation but no tensorflow implementation of cosine focal loss
", "rawMarkdown": "msharuk589 sorry, i found pytorch implementation but no tensorflow implementation of cosine focal loss", "votes": null }, { "id": "934671", "postDate": "07/18/2020 16:13:43", "content": "Could you post that here?Have you experimented on that? @mobassir
", "rawMarkdown": "Could you post that here?Have you experimented on that? @mobassir", "votes": null }, { "id": "934684", "postDate": "07/18/2020 16:30:40", "content": "no i haven't tried that yet,here is pytorch implementation : https://github.com/byeongjokim/VIPriors-Image-Classification-Challenge/blob/332e04fd3e82b20d312128bad302a9081f5c37ce/timm/loss/cosine.py
", "rawMarkdown": "no i haven't tried that yet,here is pytorch implementation : https://github.com/byeongjokim/VIPriors-Image-Classification-Challenge/blob/332e04fd3e82b20d312128bad302a9081f5c37ce/timm/loss/cosine.py", "votes": null } ], "comments": [ { "id": 934617, "author_name": "msharuk589", "author_url": "", "post_date": "07/18/2020 15:32:44", "content": "Is there any snippet for focal cosine loss in tensorflow??
", "votes": null, "replies": [ { "id": 934641, "author_name": "mobassir", "author_url": "", "post_date": "07/18/2020 15:52:15", "content": "@msharuk589 sorry, i found pytorch implementation but no tensorflow implementation of cosine focal loss
", "votes": null, "replies": [] }, { "id": 934671, "author_name": "msharuk589", "author_url": "", "post_date": "07/18/2020 16:13:43", "content": "Could you post that here?Have you experimented on that? @mobassir
", "votes": null, "replies": [] }, { "id": 934684, "author_name": "mobassir", "author_url": "", "post_date": "07/18/2020 16:30:40", "content": "no i haven't tried that yet,here is pytorch implementation : https://github.com/byeongjokim/VIPriors-Image-Classification-Challenge/blob/332e04fd3e82b20d312128bad302a9081f5c37ce/timm/loss/cosine.py
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "934567": "**found 1 interesting paper today**\n\nAbstract : \nIn general, sufficient data is essential for the better perfor-mance and generalization of deep-learning models. However, lots of lim-itations(cost, resources, etc.) of data collection leads to lack of enoughdata in most of the areas. In addition, various domains of each datasources and licenses also lead to difficulties in collection of sufficientdata. This situation makes us hard to utilize not only the pre-trainedmodel, but also the external knowledge. Therefore, it is important toleverage small dataset effectively for achieving the better performance.We applied some techniques in three aspects: data, loss function, andprediction to enable training from scratch with less data. With thesemethods, we obtain high accuracy by leveraging ImageNet data whichconsist of only 50 images per class. Furthermore, our model is ranked 4thin Visual Inductive Printers for Data-Effective Computer Vision Chal-lenge.\n\n\n\nalso check figure 2 of that paper :)\n\npaper link : https://arxiv.org/pdf/2007.07805.pdf", "934617": "Is there any snippet for focal cosine loss in tensorflow??", "934641": "msharuk589 sorry, i found pytorch implementation but no tensorflow implementation of cosine focal loss", "934671": "Could you post that here?Have you experimented on that? @mobassir", "934684": "no i haven't tried that yet,here is pytorch implementation : https://github.com/byeongjokim/VIPriors-Image-Classification-Challenge/blob/332e04fd3e82b20d312128bad302a9081f5c37ce/timm/loss/cosine.py" }, "source": "meta" }