According to most kernels, people are choosing resnet as their model choice. I googled about resnet and found out that it uses residual learning and that’s why it is very good. Is it the best model right now for image classification?
", "messages": [ { "id": 573341, "postDate": "2019-07-12T06:20:18.983Z", "content": "IMHO it is considered default by today's standards. Resnet 18 and 34 are small and fast so you can experiment quickly and translate gained knowledge to bigger models such as resnet 101. Other reason is that resnets are well studied in academia and are continuously improving ( for example wide resnets, resnext, seresnext etc.).
", "rawMarkdown": "IMHO it is considered default by today's standards. Resnet 18 and 34 are small and fast so you can experiment quickly and translate gained knowledge to bigger models such as resnet 101. Other reason is that resnets are well studied in academia and are continuously improving ( for example wide resnets, resnext, seresnext etc.).", "votes": 8, "replies": [ { "id": 573350, "postDate": "2019-07-12T06:42:32.643Z", "content": "Thanks!
", "rawMarkdown": "Thanks!" } ] }, { "id": 573336, "postDate": "2019-07-12T06:07:32.753Z", "content": "According to most kernels, people are choosing resnet as their model choice. I googled about resnet and found out that it uses residual learning and that’s why it is very good. Is it the best model right now for image classification?
", "rawMarkdown": "According to most kernels, people are choosing resnet as their model choice. I googled about resnet and found out that it uses residual learning and that’s why it is very good. Is it the best model right now for image classification?", "votes": 5 }, { "id": 574767, "postDate": "2019-07-14T13:08:46.333Z", "content": "ResNet is something like a first choice model for image classification. But it seems it may change in the nearest future, have you looked at EfficientNet? It looks promising for this competition at least, see graph in this kernel:
\n\nhttps://www.kaggle.com/hmendonca/efficientnetb0-fastai-blindness-detection
", "rawMarkdown": "ResNet is something like a first choice model for image classification. But it seems it may change in the nearest future, have you looked at EfficientNet? It looks promising for this competition at least, see graph in this kernel:\n\nhttps://www.kaggle.com/hmendonca/efficientnetb0-fastai-blindness-detection", "votes": 4, "replies": [ { "id": 580971, "postDate": "2019-07-21T07:18:28.343Z", "content": "Thanks for that link. It was very useful!
", "rawMarkdown": "Thanks for that link. It was very useful!" } ] }, { "id": 573873, "postDate": "2019-07-12T22:03:25.677Z", "content": "The Google paper on this type of dataset and problem also used resnet. I think it had some reasoning for doing so as well.
", "rawMarkdown": "The Google paper on this type of dataset and problem also used resnet. I think it had some reasoning for doing so as well.", "votes": 1 }, { "id": 573910, "postDate": "2019-07-13T00:04:39.097Z", "content": "I'm using Inception v3 results for me were better than my ResNet-50 but I'm going to try a ResNet-101
", "rawMarkdown": "I'm using Inception v3 results for me were better than my ResNet-50 but I'm going to try a ResNet-101", "replies": [ { "id": 574624, "postDate": "2019-07-14T07:52:55.547Z", "content": "I think that ResNet-101 should be better than Inception
", "rawMarkdown": "I think that ResNet-101 should be better than Inception ", "votes": 1 } ] }, { "id": 573600, "postDate": "2019-07-12T13:29:24.730Z", "content": "So,what's your model?
", "rawMarkdown": "So,what's your model?", "replies": [ { "id": 573676, "postDate": "2019-07-12T15:45:09.803Z", "content": "Resnet
", "rawMarkdown": "Resnet", "votes": 1 }, { "id": 573695, "postDate": "2019-07-12T16:12:54.670Z", "content": "What about you?
", "rawMarkdown": "What about you?" }, { "id": 573701, "postDate": "2019-07-12T16:18:30.187Z", "content": "The same as you
", "rawMarkdown": "The same as you", "votes": 1 } ] }, { "id": 581074, "postDate": "2019-07-21T11:18:03.943Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 573341, "author_name": "Miroslav Valan", "author_url": "", "post_date": "2019-07-12T06:20:18.983000", "content": "IMHO it is considered default by today's standards. Resnet 18 and 34 are small and fast so you can experiment quickly and translate gained knowledge to bigger models such as resnet 101. Other reason is that resnets are well studied in academia and are continuously improving ( for example wide resnets, resnext, seresnext etc.).
", "votes": 8, "replies": [ { "id": 573350, "author_name": "[ods.ai] Kyrylo", "author_url": "", "post_date": "2019-07-12T06:42:32.643000", "content": "Thanks!
", "votes": 0, "replies": [] } ] }, { "id": 574767, "author_name": "Anna Novikova", "author_url": "", "post_date": "2019-07-14T13:08:46.333000", "content": "ResNet is something like a first choice model for image classification. But it seems it may change in the nearest future, have you looked at EfficientNet? It looks promising for this competition at least, see graph in this kernel:
\n\nhttps://www.kaggle.com/hmendonca/efficientnetb0-fastai-blindness-detection
", "votes": 4, "replies": [ { "id": 580971, "author_name": "[ods.ai] Kyrylo", "author_url": "", "post_date": "2019-07-21T07:18:28.343000", "content": "Thanks for that link. It was very useful!
", "votes": 0, "replies": [] } ] }, { "id": 573873, "author_name": "averagemn", "author_url": "", "post_date": "2019-07-12T22:03:25.677000", "content": "The Google paper on this type of dataset and problem also used resnet. I think it had some reasoning for doing so as well.
", "votes": 1, "replies": [] }, { "id": 573910, "author_name": "C4rl05/V", "author_url": "", "post_date": "2019-07-13T00:04:39.097000", "content": "I'm using Inception v3 results for me were better than my ResNet-50 but I'm going to try a ResNet-101
", "votes": 0, "replies": [ { "id": 574624, "author_name": "[ods.ai] Kyrylo", "author_url": "", "post_date": "2019-07-14T07:52:55.547000", "content": "I think that ResNet-101 should be better than Inception
", "votes": 1, "replies": [] } ] }, { "id": 573600, "author_name": "哈尔的移动城堡", "author_url": "", "post_date": "2019-07-12T13:29:24.730000", "content": "So,what's your model?
", "votes": 0, "replies": [ { "id": 573676, "author_name": "[ods.ai] Kyrylo", "author_url": "", "post_date": "2019-07-12T15:45:09.803000", "content": "Resnet
", "votes": 1, "replies": [] }, { "id": 573695, "author_name": "[ods.ai] Kyrylo", "author_url": "", "post_date": "2019-07-12T16:12:54.670000", "content": "What about you?
", "votes": 0, "replies": [] }, { "id": 573701, "author_name": "哈尔的移动城堡", "author_url": "", "post_date": "2019-07-12T16:18:30.187000", "content": "The same as you
", "votes": 1, "replies": [] } ] }, { "id": 581074, "author_name": "", "author_url": "", "post_date": "2019-07-21T11:18:03.943000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "573341": "IMHO it is considered default by today's standards. Resnet 18 and 34 are small and fast so you can experiment quickly and translate gained knowledge to bigger models such as resnet 101. Other reason is that resnets are well studied in academia and are continuously improving ( for example wide resnets, resnext, seresnext etc.).", "573336": "According to most kernels, people are choosing resnet as their model choice. I googled about resnet and found out that it uses residual learning and that’s why it is very good. Is it the best model right now for image classification?", "574767": "ResNet is something like a first choice model for image classification. But it seems it may change in the nearest future, have you looked at EfficientNet? It looks promising for this competition at least, see graph in this kernel:\n\nhttps://www.kaggle.com/hmendonca/efficientnetb0-fastai-blindness-detection", "573873": "The Google paper on this type of dataset and problem also used resnet. I think it had some reasoning for doing so as well.", "573910": "I'm using Inception v3 results for me were better than my ResNet-50 but I'm going to try a ResNet-101", "573600": "So,what's your model?", "581074": "" } }