{ "id": 341041, "title": "EfficientNet Transfer Learning + Data Augmentation", "url": "/competitions/mayo-clinic-strip-ai/discussion/341041", "author_name": "Minhajul Hoque", "post_date": "2022-08-01T03:08:06.974000", "votes": 10, "comment_count": 8, "views": 0, "content": "

Hi friends :),

\n

I have been working on a notebook that uses Tensorflow and transfer learning. I really hope some of you can take inspiration from the notebook to work on your own.

\n

Excited to see the feedback I get from you bright minds.
\nNotebook: Mayo Clinic: EfficientNet Transfer Learning

", "messages": [ { "id": 1879345, "postDate": "2022-08-01T03:08:06.973Z", "content": "

Hi friends :),

\n

I have been working on a notebook that uses Tensorflow and transfer learning. I really hope some of you can take inspiration from the notebook to work on your own.

\n

Excited to see the feedback I get from you bright minds.
\nNotebook: Mayo Clinic: EfficientNet Transfer Learning

", "rawMarkdown": "Hi friends :),\n\nI have been working on a notebook that uses Tensorflow and transfer learning. I really hope some of you can take inspiration from the notebook to work on your own.\n\nExcited to see the feedback I get from you bright minds. \nNotebook: [Mayo Clinic: EfficientNet Transfer Learning](https://www.kaggle.com/minhajulhoque/mayo-clinic-efficientnet-transfer-learning)", "votes": 9 }, { "id": 1888758, "postDate": "2022-08-07T19:00:05.220Z", "content": "

Excellent work on this notebook! I found it very clear, well-organized, and helpful. I particularly appreciate the detail you went into on the data augmentation.

\n

The Devastator.

", "rawMarkdown": "Excellent work on this notebook! I found it very clear, well-organized, and helpful. I particularly appreciate the detail you went into on the data augmentation.\n\n\nThe Devastator.\n", "votes": 1, "replies": [ { "id": 1891946, "postDate": "2022-08-09T18:51:15.173Z", "content": "

Thanks for the feedback! Appreciate it coming from you ✊

", "rawMarkdown": "Thanks for the feedback! Appreciate it coming from you ✊" } ] }, { "id": 1883549, "postDate": "2022-08-03T22:50:42.140Z", "content": "

Very nice solution!

", "rawMarkdown": "Very nice solution!", "votes": 1, "replies": [ { "id": 1883695, "postDate": "2022-08-04T03:21:26.253Z", "content": "

Thank you! Glad you liked it

", "rawMarkdown": "Thank you! Glad you liked it" } ] }, { "id": 1883470, "postDate": "2022-08-03T20:51:12.623Z", "content": "

Thanks for your sharing. What is the classification report? It seems that the model might predict everything as CE.

", "rawMarkdown": "Thanks for your sharing. What is the classification report? It seems that the model might predict everything as CE. ", "votes": 1, "replies": [ { "id": 1883694, "postDate": "2022-08-04T03:21:06.090Z", "content": "

For the four test images, it does seem that way. I'll have to validate if those four are actually CE

", "rawMarkdown": "For the four test images, it does seem that way. I'll have to validate if those four are actually CE" } ] }, { "id": 1879427, "postDate": "2022-08-01T04:20:11.450Z", "content": "

Great work! It seems like MIL is not the only way to achieve a good performance on this task.
\nBy the way, how did you figure out the 2 images (\"b894f4_0\" and \"6baf51_0\") cause issues for the training (at least, they are not totally blurry like some other pictures)? I wonder whether this can be an issue for all other types of model.

", "rawMarkdown": "Great work! It seems like MIL is not the only way to achieve a good performance on this task.\nBy the way, how did you figure out the 2 images (\"b894f4_0\" and \"6baf51_0\") cause issues for the training (at least, they are not totally blurry like some other pictures)? I wonder whether this can be an issue for all other types of model.", "votes": 1, "replies": [ { "id": 1882066, "postDate": "2022-08-03T04:04:41.733Z", "content": "

Thanks for the feedback and the reply! Those two samples were removed due to the fact that they did not make it to the downscaled dataset.

\n

Although I believe flow_from_directory would ignore those paths that point to none existing files, I decided to remove them just in case.

", "rawMarkdown": "Thanks for the feedback and the reply! Those two samples were removed due to the fact that they did not make it to the downscaled dataset. \n\nAlthough I believe `flow_from_directory` would ignore those paths that point to none existing files, I decided to remove them just in case.", "votes": 1 } ] } ], "comments": [ { "id": 1888758, "author_name": "The Devastator", "author_url": "", "post_date": "2022-08-07T19:00:05.220000", "content": "

Excellent work on this notebook! I found it very clear, well-organized, and helpful. I particularly appreciate the detail you went into on the data augmentation.

\n

The Devastator.

", "votes": 1, "replies": [ { "id": 1891946, "author_name": "Minhajul Hoque", "author_url": "", "post_date": "2022-08-09T18:51:15.173000", "content": "

Thanks for the feedback! Appreciate it coming from you ✊

", "votes": 0, "replies": [] } ] }, { "id": 1883549, "author_name": "Stirkoff", "author_url": "", "post_date": "2022-08-03T22:50:42.140000", "content": "

Very nice solution!

", "votes": 1, "replies": [ { "id": 1883695, "author_name": "Minhajul Hoque", "author_url": "", "post_date": "2022-08-04T03:21:26.253000", "content": "

Thank you! Glad you liked it

", "votes": 0, "replies": [] } ] }, { "id": 1883470, "author_name": "Tiger", "author_url": "", "post_date": "2022-08-03T20:51:12.623000", "content": "

Thanks for your sharing. What is the classification report? It seems that the model might predict everything as CE.

", "votes": 1, "replies": [ { "id": 1883694, "author_name": "Minhajul Hoque", "author_url": "", "post_date": "2022-08-04T03:21:06.090000", "content": "

For the four test images, it does seem that way. I'll have to validate if those four are actually CE

", "votes": 0, "replies": [] } ] }, { "id": 1879427, "author_name": "hjunlee941", "author_url": "", "post_date": "2022-08-01T04:20:11.450000", "content": "

Great work! It seems like MIL is not the only way to achieve a good performance on this task.
\nBy the way, how did you figure out the 2 images (\"b894f4_0\" and \"6baf51_0\") cause issues for the training (at least, they are not totally blurry like some other pictures)? I wonder whether this can be an issue for all other types of model.

", "votes": 1, "replies": [ { "id": 1882066, "author_name": "Minhajul Hoque", "author_url": "", "post_date": "2022-08-03T04:04:41.733000", "content": "

Thanks for the feedback and the reply! Those two samples were removed due to the fact that they did not make it to the downscaled dataset.

\n

Although I believe flow_from_directory would ignore those paths that point to none existing files, I decided to remove them just in case.

", "votes": 1, "replies": [] } ] } ], "raw_markdown_by_id": { "1879345": "Hi friends :),\n\nI have been working on a notebook that uses Tensorflow and transfer learning. I really hope some of you can take inspiration from the notebook to work on your own.\n\nExcited to see the feedback I get from you bright minds. \nNotebook: [Mayo Clinic: EfficientNet Transfer Learning](https://www.kaggle.com/minhajulhoque/mayo-clinic-efficientnet-transfer-learning)", "1888758": "Excellent work on this notebook! I found it very clear, well-organized, and helpful. I particularly appreciate the detail you went into on the data augmentation.\n\n\nThe Devastator.\n", "1883549": "Very nice solution!", "1883470": "Thanks for your sharing. What is the classification report? It seems that the model might predict everything as CE. ", "1879427": "Great work! It seems like MIL is not the only way to achieve a good performance on this task.\nBy the way, how did you figure out the 2 images (\"b894f4_0\" and \"6baf51_0\") cause issues for the training (at least, they are not totally blurry like some other pictures)? I wonder whether this can be an issue for all other types of model." } }