I tried PolyNet, SEnet154, InceptionResNetV2, PNASNet-5-Large. Their effects are very poor. The PolyNet score is only 0.74!!!!!\nThey all use the same data as before and have a similar network structure.\nMy highest score was given to me by EfficientNet, which puzzled me.\nIsn’t it the larger the model, the better the effect?
", "messages": [ { "id": 951401, "postDate": "2020-07-30T06:00:38.280Z", "content": "See ImageNet accuracy\n- All top spots are taken by FixEfficientNet and EfficientNet!\n- Rank 10: ResNeXt-101 32x48d\n- Rank 52: TResNet-L\n- Rank 53: FixPNASNet-5\n- Rank 69: PNASNet-5\n- Rank 73: NASNet-A Large\n- Rank 82: ResNet-50\n- Rank 94: SENet-154\n- Rank 109: PolyNet\n- Rank 110: SE-ResNeXt-101 32x4d \n- Rank 121: Inception ResNet V2\n- Rank 134: SE-ResNet-101b\n- Rank 138: Inception V4\n- Rank 353: MobileNet V3-Large 1.0\n- Rank 391: VGG-19
\n\nYou could always train more models and ensemble them, to improve the generalizability of your overall model.
", "rawMarkdown": "See [ImageNet accuracy](https://sotabench.com/benchmarks/image-classification-on-imagenet)\n- All top spots are taken by FixEfficientNet and EfficientNet!\n- Rank 10: ResNeXt-101 32x48d\n- Rank 52: TResNet-L\n- Rank 53: FixPNASNet-5\n- Rank 69: PNASNet-5\n- Rank 73: NASNet-A Large\n- Rank 82: ResNet-50\n- Rank 94: SENet-154\n- Rank 109: PolyNet\n- Rank 110: SE-ResNeXt-101 32x4d \n- Rank 121: Inception ResNet V2\n- Rank 134: SE-ResNet-101b\n- Rank 138: Inception V4\n- Rank 353: MobileNet V3-Large 1.0\n- Rank 391: VGG-19\n\nYou could always train more models and ensemble them, to improve the generalizability of your overall model.", "votes": 3, "replies": [ { "id": 951468, "postDate": "2020-07-30T06:54:53.823Z", "content": "I quite agree with you, and I did the same.\nBut models with too low scores will only lower my score, and I have no confidence in adding these models. Do you have any good options for model selection?
", "rawMarkdown": "I quite agree with you, and I did the same.\nBut models with too low scores will only lower my score, and I have no confidence in adding these models. Do you have any good options for model selection?" }, { "id": 951987, "postDate": "2020-07-30T14:36:21.170Z", "rawMarkdown": "", "isDeleted": true, "replies": [ { "id": 952070, "postDate": "2020-07-30T15:34:40.243Z", "content": "@prateekis1 Please read the reports of 2019 winners!
\n\nPS: You can get a high score just using EfficientNet at different image sizes (#1 winner last year).
", "rawMarkdown": "@prateekis1 Please read the reports of [2019 winners](https://challenge2019.isic-archive.com/leaderboard.html)!\n\nPS: You can get a high score just using EfficientNet at different image sizes (#1 winner last year)." } ] } ] }, { "id": 951336, "postDate": "2020-07-30T04:36:42.920Z", "content": "I tried PolyNet, SEnet154, InceptionResNetV2, PNASNet-5-Large. Their effects are very poor. The PolyNet score is only 0.74!!!!!\nThey all use the same data as before and have a similar network structure.\nMy highest score was given to me by EfficientNet, which puzzled me.\nIsn’t it the larger the model, the better the effect?
", "rawMarkdown": "I tried PolyNet, SEnet154, InceptionResNetV2, PNASNet-5-Large. Their effects are very poor. The PolyNet score is only 0.74!!!!!\nThey all use the same data as before and have a similar network structure.\nMy highest score was given to me by EfficientNet, which puzzled me.\nIsn’t it the larger the model, the better the effect?", "votes": 3 }, { "id": 951538, "postDate": "2020-07-30T07:46:26.633Z", "content": "I have a non Efficientnet model scoring 0.9510 at LB (5 folds) . \nSo yes , you can give a try to others backbones ( I am using Pytorch though, not sure they are all available on Tensorflow).
\n\nAnd not necessarily those huge tweaked models dominating Imagenet LB.
", "rawMarkdown": "I have a non Efficientnet model scoring 0.9510 at LB (5 folds) . \nSo yes , you can give a try to others backbones ( I am using Pytorch though, not sure they are all available on Tensorflow). \n\nAnd not necessarily those huge tweaked models dominating Imagenet LB. \n\n", "votes": 1 }, { "id": 951371, "postDate": "2020-07-30T05:21:47.017Z", "content": "+1 for this. Although, I tried only two models so far, but SENet 154 scores around 0.86 and Resnext50 scores around 0.89 for the same configuration where EffNet B0 scores 0.914 on CV and 0.934 on LB (Img Size: 128*128). The interesting part is these models have much lower CV-LB gap than EffNets.
", "rawMarkdown": "+1 for this. Although, I tried only two models so far, but SENet 154 scores around 0.86 and Resnext50 scores around 0.89 for the same configuration where EffNet B0 scores 0.914 on CV and 0.934 on LB (Img Size: 128*128). The interesting part is these models have much lower CV-LB gap than EffNets.", "votes": 1, "replies": [ { "id": 951465, "postDate": "2020-07-30T06:53:31.667Z", "content": "Yes, I just listed a few models.\nFor this I have tried many models, but their effects are all very poor.\nbtw, can the use of 128 also have such good results? This is something I did not expect.
", "rawMarkdown": "Yes, I just listed a few models.\nFor this I have tried many models, but their effects are all very poor.\nbtw, can the use of 128 also have such good results? This is something I did not expect." } ] }, { "id": 951357, "postDate": "2020-07-30T04:56:20.240Z", "rawMarkdown": "", "isDeleted": true, "replies": [ { "id": 951460, "postDate": "2020-07-30T06:51:50.607Z", "content": "I trained on Titan XP GPU*8 for 2 days. I think it should fit enough.
", "rawMarkdown": "I trained on Titan XP GPU*8 for 2 days. I think it should fit enough." } ] } ], "comments": [ { "id": 951401, "author_name": "Sirish Somanchi", "author_url": "", "post_date": "2020-07-30T06:00:38.280000", "content": "See ImageNet accuracy\n- All top spots are taken by FixEfficientNet and EfficientNet!\n- Rank 10: ResNeXt-101 32x48d\n- Rank 52: TResNet-L\n- Rank 53: FixPNASNet-5\n- Rank 69: PNASNet-5\n- Rank 73: NASNet-A Large\n- Rank 82: ResNet-50\n- Rank 94: SENet-154\n- Rank 109: PolyNet\n- Rank 110: SE-ResNeXt-101 32x4d \n- Rank 121: Inception ResNet V2\n- Rank 134: SE-ResNet-101b\n- Rank 138: Inception V4\n- Rank 353: MobileNet V3-Large 1.0\n- Rank 391: VGG-19
\n\nYou could always train more models and ensemble them, to improve the generalizability of your overall model.
", "votes": 3, "replies": [ { "id": 951468, "author_name": "Mingjie Wang", "author_url": "", "post_date": "2020-07-30T06:54:53.823000", "content": "I quite agree with you, and I did the same.\nBut models with too low scores will only lower my score, and I have no confidence in adding these models. Do you have any good options for model selection?
", "votes": 0, "replies": [] }, { "id": 951987, "author_name": "", "author_url": "", "post_date": "2020-07-30T14:36:21.170000", "content": "", "votes": 0, "replies": [ { "id": 952070, "author_name": "Sirish Somanchi", "author_url": "", "post_date": "2020-07-30T15:34:40.243000", "content": "@prateekis1 Please read the reports of 2019 winners!
\n\nPS: You can get a high score just using EfficientNet at different image sizes (#1 winner last year).
", "votes": 0, "replies": [] } ] } ] }, { "id": 951538, "author_name": "Serigne ", "author_url": "", "post_date": "2020-07-30T07:46:26.633000", "content": "I have a non Efficientnet model scoring 0.9510 at LB (5 folds) . \nSo yes , you can give a try to others backbones ( I am using Pytorch though, not sure they are all available on Tensorflow).
\n\nAnd not necessarily those huge tweaked models dominating Imagenet LB.
", "votes": 1, "replies": [] }, { "id": 951371, "author_name": "Kumar Shubham", "author_url": "", "post_date": "2020-07-30T05:21:47.017000", "content": "+1 for this. Although, I tried only two models so far, but SENet 154 scores around 0.86 and Resnext50 scores around 0.89 for the same configuration where EffNet B0 scores 0.914 on CV and 0.934 on LB (Img Size: 128*128). The interesting part is these models have much lower CV-LB gap than EffNets.
", "votes": 1, "replies": [ { "id": 951465, "author_name": "Mingjie Wang", "author_url": "", "post_date": "2020-07-30T06:53:31.667000", "content": "Yes, I just listed a few models.\nFor this I have tried many models, but their effects are all very poor.\nbtw, can the use of 128 also have such good results? This is something I did not expect.
", "votes": 0, "replies": [] } ] }, { "id": 951357, "author_name": "", "author_url": "", "post_date": "2020-07-30T04:56:20.240000", "content": "", "votes": 0, "replies": [ { "id": 951460, "author_name": "Mingjie Wang", "author_url": "", "post_date": "2020-07-30T06:51:50.607000", "content": "I trained on Titan XP GPU*8 for 2 days. I think it should fit enough.
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "951401": "See [ImageNet accuracy](https://sotabench.com/benchmarks/image-classification-on-imagenet)\n- All top spots are taken by FixEfficientNet and EfficientNet!\n- Rank 10: ResNeXt-101 32x48d\n- Rank 52: TResNet-L\n- Rank 53: FixPNASNet-5\n- Rank 69: PNASNet-5\n- Rank 73: NASNet-A Large\n- Rank 82: ResNet-50\n- Rank 94: SENet-154\n- Rank 109: PolyNet\n- Rank 110: SE-ResNeXt-101 32x4d \n- Rank 121: Inception ResNet V2\n- Rank 134: SE-ResNet-101b\n- Rank 138: Inception V4\n- Rank 353: MobileNet V3-Large 1.0\n- Rank 391: VGG-19\n\nYou could always train more models and ensemble them, to improve the generalizability of your overall model.", "951336": "I tried PolyNet, SEnet154, InceptionResNetV2, PNASNet-5-Large. Their effects are very poor. The PolyNet score is only 0.74!!!!!\nThey all use the same data as before and have a similar network structure.\nMy highest score was given to me by EfficientNet, which puzzled me.\nIsn’t it the larger the model, the better the effect?", "951538": "I have a non Efficientnet model scoring 0.9510 at LB (5 folds) . \nSo yes , you can give a try to others backbones ( I am using Pytorch though, not sure they are all available on Tensorflow). \n\nAnd not necessarily those huge tweaked models dominating Imagenet LB. \n\n", "951371": "+1 for this. Although, I tried only two models so far, but SENet 154 scores around 0.86 and Resnext50 scores around 0.89 for the same configuration where EffNet B0 scores 0.914 on CV and 0.934 on LB (Img Size: 128*128). The interesting part is these models have much lower CV-LB gap than EffNets.", "951357": "" } }