I read many great discussions and beautiful kernels and see most kagglers used the models below. PS: I add some corresponding LB scores which I saw from the discussions, I am not sure whether the score is the best or not.
\n\nHope this helps.
", "messages": [ { "id": 651018, "postDate": "2019-10-17T01:01:06.380Z", "content": "I read many great discussions and beautiful kernels and see most kagglers used the models below. PS: I add some corresponding LB scores which I saw from the discussions, I am not sure whether the score is the best or not.
\n\nHope this helps.
", "rawMarkdown": "I read many great discussions and beautiful kernels and see most kagglers used the models below. PS: I add some corresponding LB scores which I saw from the discussions, I am not sure whether the score is the best or not.\n\n- Densenet\n- Resnext101_32x8d_wsl(0.084) \n- Seresnext50,se_resnext50_32x4d:LB 0.074\n- EfficientNet B0,B1(256x256 --> 0.077,0.073(B0, 224x224.))\n- VGG19 - 224x224 - 0.073\n- 256x256, model: Resnet50 0.089\n- resnet34, image 256x256, single fold -> 0.078 lb\n- ResNeXt-101 32x16d : 0.086\n- size 224x224, model: inceptionV3, single fold --> LB 0.079\n- ImagNet pre-trained with modifications\n\nHope this helps.\n\n\n", "votes": 11 }, { "id": 651136, "postDate": "2019-10-17T04:53:29.047Z", "content": "Very Helpful.. Thanks @longyin2
", "rawMarkdown": "Very Helpful.. Thanks @longyin2 ", "votes": 1, "replies": [ { "id": 651163, "postDate": "2019-10-17T06:08:57.137Z", "content": "Welcome
", "rawMarkdown": "Welcome" } ] }, { "id": 655576, "postDate": "2019-10-23T07:56:53.683Z", "content": "Thanks for sharing.\nBefore kagglers try them,they should know that:\nmany of these models are useless if you don't have good preprocess function.
", "rawMarkdown": "Thanks for sharing.\nBefore kagglers try them,they should know that:\nmany of these models are useless if you don't have good preprocess function." }, { "id": 651978, "postDate": "2019-10-18T07:15:07.717Z", "content": "Hi ,thanks.I've tried resnext-101-32x8d(224*224),efficientnet-b0(224*224),I'm going to try it. You said VGG19, and do stacking of these three models.
", "rawMarkdown": "Hi ,thanks.I've tried resnext-101-32x8d(224*224),efficientnet-b0(224*224),I'm going to try it. You said VGG19, and do stacking of these three models." }, { "id": 651516, "postDate": "2019-10-17T14:51:21.837Z", "content": "Hi @longyin2 thanks ! Could you share some details on your pre-processing please, I have been stucked for two weeks, I do multiple windows (brain, subdural, bone) followed by normalization with the train set mean and standard deviation. Have you noticed a better strategy for normalization (mean centering, reduction) ? thanks
", "rawMarkdown": "Hi @longyin2 thanks ! Could you share some details on your pre-processing please, I have been stucked for two weeks, I do multiple windows (brain, subdural, bone) followed by normalization with the train set mean and standard deviation. Have you noticed a better strategy for normalization (mean centering, reduction) ? thanks ", "replies": [ { "id": 651801, "postDate": "2019-10-18T00:53:11.230Z", "content": "My pre-processing is quite easy, just resize the picture to 224*224 and all three channel are the same.
", "rawMarkdown": "My pre-processing is quite easy, just resize the picture to 224*224 and all three channel are the same." }, { "id": 651962, "postDate": "2019-10-18T06:36:51.050Z", "content": "Thanks for your answer ! do you use any kind of normalization like min-max scaling or standard scaling ? I have mixed feelings about both, I don't see a clear winner.
", "rawMarkdown": "Thanks for your answer ! do you use any kind of normalization like min-max scaling or standard scaling ? I have mixed feelings about both, I don't see a clear winner." }, { "id": 651983, "postDate": "2019-10-18T07:20:32.253Z", "content": "Not yet.
", "rawMarkdown": "Not yet." } ] }, { "id": 651170, "postDate": "2019-10-17T06:19:40.560Z", "content": "Very interested to see what model heads (eg creative ways to use the any class) and different ways of using metadata people come up with.
\n\nDo you think these scores are achievable with just simple classifiers and basic tricks?
", "rawMarkdown": "Very interested to see what model heads (eg creative ways to use the any class) and different ways of using metadata people come up with.\n\nDo you think these scores are achievable with just simple classifiers and basic tricks?" } ], "comments": [ { "id": 651136, "author_name": "Ailurophile", "author_url": "", "post_date": "2019-10-17T04:53:29.047000", "content": "Very Helpful.. Thanks @longyin2
", "votes": 1, "replies": [ { "id": 651163, "author_name": "LongYin/杰少", "author_url": "", "post_date": "2019-10-17T06:08:57.137000", "content": "Welcome
", "votes": 0, "replies": [] } ] }, { "id": 655576, "author_name": "Tian Bingyang", "author_url": "", "post_date": "2019-10-23T07:56:53.683000", "content": "Thanks for sharing.\nBefore kagglers try them,they should know that:\nmany of these models are useless if you don't have good preprocess function.
", "votes": 0, "replies": [] }, { "id": 651978, "author_name": "genius0182", "author_url": "", "post_date": "2019-10-18T07:15:07.717000", "content": "Hi ,thanks.I've tried resnext-101-32x8d(224*224),efficientnet-b0(224*224),I'm going to try it. You said VGG19, and do stacking of these three models.
", "votes": 0, "replies": [] }, { "id": 651516, "author_name": "pantoine", "author_url": "", "post_date": "2019-10-17T14:51:21.837000", "content": "Hi @longyin2 thanks ! Could you share some details on your pre-processing please, I have been stucked for two weeks, I do multiple windows (brain, subdural, bone) followed by normalization with the train set mean and standard deviation. Have you noticed a better strategy for normalization (mean centering, reduction) ? thanks
", "votes": 0, "replies": [ { "id": 651801, "author_name": "LongYin/杰少", "author_url": "", "post_date": "2019-10-18T00:53:11.230000", "content": "My pre-processing is quite easy, just resize the picture to 224*224 and all three channel are the same.
", "votes": 0, "replies": [] }, { "id": 651962, "author_name": "pantoine", "author_url": "", "post_date": "2019-10-18T06:36:51.050000", "content": "Thanks for your answer ! do you use any kind of normalization like min-max scaling or standard scaling ? I have mixed feelings about both, I don't see a clear winner.
", "votes": 0, "replies": [] }, { "id": 651983, "author_name": "LongYin/杰少", "author_url": "", "post_date": "2019-10-18T07:20:32.253000", "content": "Not yet.
", "votes": 0, "replies": [] } ] }, { "id": 651170, "author_name": "cherring", "author_url": "", "post_date": "2019-10-17T06:19:40.560000", "content": "Very interested to see what model heads (eg creative ways to use the any class) and different ways of using metadata people come up with.
\n\nDo you think these scores are achievable with just simple classifiers and basic tricks?
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "651018": "I read many great discussions and beautiful kernels and see most kagglers used the models below. PS: I add some corresponding LB scores which I saw from the discussions, I am not sure whether the score is the best or not.\n\n- Densenet\n- Resnext101_32x8d_wsl(0.084) \n- Seresnext50,se_resnext50_32x4d:LB 0.074\n- EfficientNet B0,B1(256x256 --> 0.077,0.073(B0, 224x224.))\n- VGG19 - 224x224 - 0.073\n- 256x256, model: Resnet50 0.089\n- resnet34, image 256x256, single fold -> 0.078 lb\n- ResNeXt-101 32x16d : 0.086\n- size 224x224, model: inceptionV3, single fold --> LB 0.079\n- ImagNet pre-trained with modifications\n\nHope this helps.\n\n\n", "651136": "Very Helpful.. Thanks @longyin2 ", "655576": "Thanks for sharing.\nBefore kagglers try them,they should know that:\nmany of these models are useless if you don't have good preprocess function.", "651978": "Hi ,thanks.I've tried resnext-101-32x8d(224*224),efficientnet-b0(224*224),I'm going to try it. You said VGG19, and do stacking of these three models.", "651516": "Hi @longyin2 thanks ! Could you share some details on your pre-processing please, I have been stucked for two weeks, I do multiple windows (brain, subdural, bone) followed by normalization with the train set mean and standard deviation. Have you noticed a better strategy for normalization (mean centering, reduction) ? thanks ", "651170": "Very interested to see what model heads (eg creative ways to use the any class) and different ways of using metadata people come up with.\n\nDo you think these scores are achievable with just simple classifiers and basic tricks?" } }