I saw that people have been getting good results from ResNet50 in this competition, however when I tried it myself on Tensorflow, the results were pretty bad. The validation accuracy and LB could never go past 0.89, compared to EfficientNet which achieved it easily. Is there a fundamental difference between Tensorflow and Pytorch's implementation?
", "messages": [ { "id": "1167120", "postDate": "01/24/2021 04:17:44", "content": "I saw that people have been getting good results from ResNet50 in this competition, however when I tried it myself on Tensorflow, the results were pretty bad. The validation accuracy and LB could never go past 0.89, compared to EfficientNet which achieved it easily. Is there a fundamental difference between Tensorflow and Pytorch's implementation?
", "rawMarkdown": "I saw that people have been getting good results from ResNet50 in this competition, however when I tried it myself on Tensorflow, the results were pretty bad. The validation accuracy and LB could never go past 0.89, compared to EfficientNet which achieved it easily. Is there a fundamental difference between Tensorflow and Pytorch's implementation?", "votes": null }, { "id": "1167139", "postDate": "01/24/2021 04:50:30", "content": "Yes, I wondered the same when i saw people getting lb 90.1 with Resnet but When I Tried the result where Poor,
", "rawMarkdown": "Yes, I wondered the same when i saw people getting lb 90.1 with Resnet but When I Tried the result where Poor,", "votes": null }, { "id": "1167407", "postDate": "01/24/2021 08:33:41", "content": "Apart from using ResNet model, there are many other things also that plays the role.
\nLike learning rate , feed-forward network, optimizer , the way you fin tune model , TTA Steps , etc…
\nSo , there might be a case that the ones how are using resnet and getting score over 0.9 have experimented and used the best parameters. And this is the only reason that they are getting such good score.
Try Experimenting , it can be of some help surely.
", "rawMarkdown": "Apart from using ResNet model, there are many other things also that plays the role.\nLike learning rate , feed-forward network, optimizer , the way you fin tune model , TTA Steps , etc...\nSo , there might be a case that the ones how are using resnet and getting score over 0.9 have experimented and used the best parameters. And this is the only reason that they are getting such good score.\n\nTry Experimenting , it can be of some help surely.", "votes": null }, { "id": "1167564", "postDate": "01/24/2021 11:05:01", "content": "The original Resnet is great but it has been further tweaked, improved, retrained in too many ways. That's why you have many variants now that people may still call Resnet. But they can give very different results.
\nFor instance, (se)resnext50_32x4d is a good compromise that people use regularly because it's as fast as resnet50 while giving better results in many tasks.
\nI don't know if all these variants are available in TF, and/or their correctly pretrained weights.
", "rawMarkdown": "The original Resnet is great but it has been further tweaked, improved, retrained in too many ways. That's why you have many variants now that people may still call Resnet. But they can give very different results. \n\nFor instance, (se)resnext50_32x4d is a good compromise that people use regularly because it's as fast as resnet50 while giving better results in many tasks. \n\nI don't know if all these variants are available in TF, and/or their correctly pretrained weights.", "votes": null } ], "comments": [ { "id": 1167139, "author_name": "mithilsalunkhe", "author_url": "", "post_date": "01/24/2021 04:50:30", "content": "Yes, I wondered the same when i saw people getting lb 90.1 with Resnet but When I Tried the result where Poor,
", "votes": null, "replies": [] }, { "id": 1167407, "author_name": "prashantarorat", "author_url": "", "post_date": "01/24/2021 08:33:41", "content": "Apart from using ResNet model, there are many other things also that plays the role.
\nLike learning rate , feed-forward network, optimizer , the way you fin tune model , TTA Steps , etc…
\nSo , there might be a case that the ones how are using resnet and getting score over 0.9 have experimented and used the best parameters. And this is the only reason that they are getting such good score.
Try Experimenting , it can be of some help surely.
", "votes": null, "replies": [] }, { "id": 1167564, "author_name": "serigne", "author_url": "", "post_date": "01/24/2021 11:05:01", "content": "The original Resnet is great but it has been further tweaked, improved, retrained in too many ways. That's why you have many variants now that people may still call Resnet. But they can give very different results.
\nFor instance, (se)resnext50_32x4d is a good compromise that people use regularly because it's as fast as resnet50 while giving better results in many tasks.
\nI don't know if all these variants are available in TF, and/or their correctly pretrained weights.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1167120": "I saw that people have been getting good results from ResNet50 in this competition, however when I tried it myself on Tensorflow, the results were pretty bad. The validation accuracy and LB could never go past 0.89, compared to EfficientNet which achieved it easily. Is there a fundamental difference between Tensorflow and Pytorch's implementation?", "1167139": "Yes, I wondered the same when i saw people getting lb 90.1 with Resnet but When I Tried the result where Poor,", "1167407": "Apart from using ResNet model, there are many other things also that plays the role.\nLike learning rate , feed-forward network, optimizer , the way you fin tune model , TTA Steps , etc...\nSo , there might be a case that the ones how are using resnet and getting score over 0.9 have experimented and used the best parameters. And this is the only reason that they are getting such good score.\n\nTry Experimenting , it can be of some help surely.", "1167564": "The original Resnet is great but it has been further tweaked, improved, retrained in too many ways. That's why you have many variants now that people may still call Resnet. But they can give very different results. \n\nFor instance, (se)resnext50_32x4d is a good compromise that people use regularly because it's as fast as resnet50 while giving better results in many tasks. \n\nI don't know if all these variants are available in TF, and/or their correctly pretrained weights." }, "source": "meta" }