{ "id": 448005, "title": "Using Thumbnails or Real images dataset ", "url": "/competitions/UBC-OCEAN/discussion/448005", "author_name": "Raheem Nasirudeen", "post_date": "2023-10-18T05:22:00.599000", "votes": 4, "comment_count": 8, "views": 0, "content": "

Which one is the best to use?
\nBest regards, everyone.

", "messages": [ { "id": 2486676, "postDate": "2023-10-18T05:22:00.600Z", "content": "

Which one is the best to use?
\nBest regards, everyone.

", "rawMarkdown": "Which one is the best to use?\nBest regards, everyone. ", "votes": 4 }, { "id": 2500847, "postDate": "2023-10-27T03:57:37.710Z", "content": "

Using thumbnail images I am able to get LB score of 0.28 so far

", "rawMarkdown": "Using thumbnail images I am able to get LB score of 0.28 so far", "votes": 1, "replies": [ { "id": 2509573, "postDate": "2023-11-02T13:03:09.570Z", "content": "

My best scoring model thus far on thumbnails (EfficientNetB5) scored 0.33.

", "rawMarkdown": "My best scoring model thus far on thumbnails (EfficientNetB5) scored 0.33." } ] }, { "id": 2486718, "postDate": "2023-10-18T06:05:55.710Z", "content": "

as for now, I was not able to process the test_images without the OOM error on Kaggle.
\nWill try to explore the inference using the test_thumbnails folder.

", "rawMarkdown": "as for now, I was not able to process the test_images without the OOM error on Kaggle.\nWill try to explore the inference using the test_thumbnails folder.", "votes": 1, "replies": [ { "id": 2486845, "postDate": "2023-10-18T07:28:36.740Z", "content": "

Thanks a lot .

", "rawMarkdown": "Thanks a lot ." } ] }, { "id": 2508749, "postDate": "2023-11-02T01:18:46.397Z", "content": "

Perhaps we need to use real images and resize them to obtain data with the same magnification, but this will result in a testing time of nearly 8 hours on the test set.

", "rawMarkdown": "Perhaps we need to use real images and resize them to obtain data with the same magnification, but this will result in a testing time of nearly 8 hours on the test set." }, { "id": 2508555, "postDate": "2023-11-01T19:34:43.813Z", "content": "

I think you may decide between thumbnail and tiles training as there is not much way you can load the image and also almost all CNNs have limited receptive field, so in most cases up to 512px

", "rawMarkdown": "I think you may decide between [thumbnail](https://www.kaggle.com/code/jirkaborovec/cancer-subtype-baseline-with-lightning-timm) and [tiles](https://www.kaggle.com/code/jirkaborovec/cancer-subtype-tiles-w-lightning-timm-models) training as there is not much way you can load the image and also almost all CNNs have limited receptive field, so in most cases up to 512px", "replies": [ { "id": 2511021, "postDate": "2023-11-03T12:44:19.740Z", "content": "

Are you getting better accuracy using tile, thumbnail is full size image image and ideally model should perform better on thumbnail right?

", "rawMarkdown": "Are you getting better accuracy using tile, thumbnail is full size image image and ideally model should perform better on thumbnail right?", "replies": [ { "id": 2511054, "postDate": "2023-11-03T13:13:10.700Z", "content": "

Yes, you can compare these two notebooks:

\n", "rawMarkdown": "Yes, you can compare these two notebooks:\n- [CanceršŸ”¬Subtype: tilesšŸ–½ w/ Lightningāš”TIMM models](https://kaggle.com/code/jirkaborovec/cancer-subtype-tiles-w-lightning-timm-models) and its [inference](https://kaggle.com/code/jirkaborovec/cancer-subtype-lightning-torch-inference-tiles)\n- [CanceršŸ”¬Subtype: baseline with Lightningāš”TIMM](https://kaggle.com/code/jirkaborovec/cancer-subtype-baseline-with-lightning-timm) and its [inference](https://kaggle.com/code/jirkaborovec/cancer-subtype-lit-torch-inference-thumbnails)", "votes": 2 } ] } ] } ], "comments": [ { "id": 2500847, "author_name": "Sonu Jha", "author_url": "", "post_date": "2023-10-27T03:57:37.710000", "content": "

Using thumbnail images I am able to get LB score of 0.28 so far

", "votes": 1, "replies": [ { "id": 2509573, "author_name": "lmadsen", "author_url": "", "post_date": "2023-11-02T13:03:09.570000", "content": "

My best scoring model thus far on thumbnails (EfficientNetB5) scored 0.33.

", "votes": 0, "replies": [] } ] }, { "id": 2486718, "author_name": "Igor Krashenyi", "author_url": "", "post_date": "2023-10-18T06:05:55.710000", "content": "

as for now, I was not able to process the test_images without the OOM error on Kaggle.
\nWill try to explore the inference using the test_thumbnails folder.

", "votes": 1, "replies": [ { "id": 2486845, "author_name": "Raheem Nasirudeen", "author_url": "", "post_date": "2023-10-18T07:28:36.740000", "content": "

Thanks a lot .

", "votes": 0, "replies": [] } ] }, { "id": 2508749, "author_name": "Dean Gale", "author_url": "", "post_date": "2023-11-02T01:18:46.397000", "content": "

Perhaps we need to use real images and resize them to obtain data with the same magnification, but this will result in a testing time of nearly 8 hours on the test set.

", "votes": 0, "replies": [] }, { "id": 2508555, "author_name": "Jirka", "author_url": "", "post_date": "2023-11-01T19:34:43.813000", "content": "

I think you may decide between thumbnail and tiles training as there is not much way you can load the image and also almost all CNNs have limited receptive field, so in most cases up to 512px

", "votes": 0, "replies": [ { "id": 2511021, "author_name": "Sonu Jha", "author_url": "", "post_date": "2023-11-03T12:44:19.740000", "content": "

Are you getting better accuracy using tile, thumbnail is full size image image and ideally model should perform better on thumbnail right?

", "votes": 0, "replies": [ { "id": 2511054, "author_name": "Jirka", "author_url": "", "post_date": "2023-11-03T13:13:10.700000", "content": "

Yes, you can compare these two notebooks:

\n", "votes": 2, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2486676": "Which one is the best to use?\nBest regards, everyone. ", "2500847": "Using thumbnail images I am able to get LB score of 0.28 so far", "2486718": "as for now, I was not able to process the test_images without the OOM error on Kaggle.\nWill try to explore the inference using the test_thumbnails folder.", "2508749": "Perhaps we need to use real images and resize them to obtain data with the same magnification, but this will result in a testing time of nearly 8 hours on the test set.", "2508555": "I think you may decide between [thumbnail](https://www.kaggle.com/code/jirkaborovec/cancer-subtype-baseline-with-lightning-timm) and [tiles](https://www.kaggle.com/code/jirkaborovec/cancer-subtype-tiles-w-lightning-timm-models) training as there is not much way you can load the image and also almost all CNNs have limited receptive field, so in most cases up to 512px" } }