{ "id": 312024, "title": "(UPDATED)5 Reduced Resolution Image Datasets (128, 224, 512, 1024, 2048)", "url": "/competitions/ultra-mnist/discussion/312024", "author_name": "", "post_date": "2022-03-10T02:59:15.096473700Z", "votes": 10, "comment_count": 2, "views": 0, "content": "

Hello everyone!

\n

I have created 5 different datasets for this competition. I have taken the original images and scaled them down to 5 different lower resolutions (128 x 128, 224 x 224, 512 x 512, 1024x1024 and\n1024 x1024 ). From previous competitions, we have seen that higher resolutions generally have better model performance but these provide a great starting point.

\n

All 5 datasets also include the train.csv and submission.csv files in each of these datasets for convenience.

\n

These datasets are also a lot smaller in scale compared to the original dataset which should make it easier to download locally or even use in a Kaggle kernel. The size of the dataset with 224x224 images is only 1.16 GB as compared to the original dataset which has a size of 34.5 GB.
\nPlease let me know if you have any feedback ✨

\n

Here are the dataset links:

\n

(128 x 128 dataset) - https://www.kaggle.com/odins0n/ultramnist128x128 (593 MB)
\n(224 x 224 dataset) - https://www.kaggle.com/odins0n/ultramnist224x224 (1.16 GB)
\n(512 x 512 dataset) - https://www.kaggle.com/odins0n/ultramnist512x512 (2.97 GB)
\n(1024 x 1024 dataset) - https://www.kaggle.com/odins0n/ultramnist1024x1024 (4.85 GB)
\n(2048 x 2048 dataset) - https://www.kaggle.com/odins0n/ultramnist2048x2048 (15.46 GB)

\n

If you want are interested in knowing how the data was compressed and which interpolation method was used, check out this notebook - https://www.kaggle.com/odins0n/ultramnist-image-resizer-512x512

", "messages": [ { "id": "1717545", "postDate": "03/10/2022 02:59:15", "content": "

Hello everyone!

\n

I have created 5 different datasets for this competition. I have taken the original images and scaled them down to 5 different lower resolutions (128 x 128, 224 x 224, 512 x 512, 1024x1024 and\n1024 x1024 ). From previous competitions, we have seen that higher resolutions generally have better model performance but these provide a great starting point.

\n

All 5 datasets also include the train.csv and submission.csv files in each of these datasets for convenience.

\n

These datasets are also a lot smaller in scale compared to the original dataset which should make it easier to download locally or even use in a Kaggle kernel. The size of the dataset with 224x224 images is only 1.16 GB as compared to the original dataset which has a size of 34.5 GB.
\nPlease let me know if you have any feedback ✨

\n

Here are the dataset links:

\n

(128 x 128 dataset) - https://www.kaggle.com/odins0n/ultramnist128x128 (593 MB)
\n(224 x 224 dataset) - https://www.kaggle.com/odins0n/ultramnist224x224 (1.16 GB)
\n(512 x 512 dataset) - https://www.kaggle.com/odins0n/ultramnist512x512 (2.97 GB)
\n(1024 x 1024 dataset) - https://www.kaggle.com/odins0n/ultramnist1024x1024 (4.85 GB)
\n(2048 x 2048 dataset) - https://www.kaggle.com/odins0n/ultramnist2048x2048 (15.46 GB)

\n

If you want are interested in knowing how the data was compressed and which interpolation method was used, check out this notebook - https://www.kaggle.com/odins0n/ultramnist-image-resizer-512x512

", "rawMarkdown": "Hello everyone!\n\nI have created 5 different datasets for this competition. I have taken the original images and scaled them down to 5 different lower resolutions (128 x 128, 224 x 224, 512 x 512, 1024x1024 and\n1024 x1024 ). From previous competitions, we have seen that higher resolutions generally have better model performance but these provide a great starting point.\n\nAll 5 datasets also include the train.csv and submission.csv files in each of these datasets for convenience.\n\nThese datasets are also a lot smaller in scale compared to the original dataset which should make it easier to download locally or even use in a Kaggle kernel. The size of the dataset with 224x224 images is only 1.16 GB as compared to the original dataset which has a size of 34.5 GB.\nPlease let me know if you have any feedback ✨\n\nHere are the dataset links:\n\n(128 x 128 dataset) - https://www.kaggle.com/odins0n/ultramnist128x128 (593 MB)\n(224 x 224 dataset) - https://www.kaggle.com/odins0n/ultramnist224x224 (1.16 GB)\n(512 x 512 dataset) - https://www.kaggle.com/odins0n/ultramnist512x512 (2.97 GB)\n(1024 x 1024 dataset) - https://www.kaggle.com/odins0n/ultramnist1024x1024 (4.85 GB)\n(2048 x 2048 dataset) - https://www.kaggle.com/odins0n/ultramnist2048x2048 (15.46 GB)\n\nIf you want are interested in knowing how the data was compressed and which interpolation method was used, check out this notebook - https://www.kaggle.com/odins0n/ultramnist-image-resizer-512x512", "votes": null }, { "id": "1717711", "postDate": "03/10/2022 06:41:04", "content": "

Hey Sanskar,

\n

Can you clarify the note? I am a bit new to competitions and the wording is unclear to me. I thought we can use the competition data in any way (including resizing it) without any data from other sources. In my assumption, \"no external data\" means we can still use datasets like yours or ones that have more drastic changes (like the cleaned images from Remek).

", "rawMarkdown": "Hey Sanskar,\n\nCan you clarify the note? I am a bit new to competitions and the wording is unclear to me. I thought we can use the competition data in any way (including resizing it) without any data from other sources. In my assumption, \"no external data\" means we can still use datasets like yours or ones that have more drastic changes (like the cleaned images from Remek).", "votes": null }, { "id": "1719941", "postDate": "03/12/2022 09:35:45", "content": "

This is a simple resize. So, this can be used for final submission. Any dataset with pre-processing of images is also allowed for GPU track but not for Innovation track. Further clarification is here: https://www.kaggle.com/c/ultra-mnist/discussion/312470

", "rawMarkdown": "This is a simple resize. So, this can be used for final submission. Any dataset with pre-processing of images is also allowed for GPU track but not for Innovation track. Further clarification is here: https://www.kaggle.com/c/ultra-mnist/discussion/312470", "votes": null } ], "comments": [ { "id": 1717711, "author_name": "outwrest", "author_url": "", "post_date": "03/10/2022 06:41:04", "content": "

Hey Sanskar,

\n

Can you clarify the note? I am a bit new to competitions and the wording is unclear to me. I thought we can use the competition data in any way (including resizing it) without any data from other sources. In my assumption, \"no external data\" means we can still use datasets like yours or ones that have more drastic changes (like the cleaned images from Remek).

", "votes": null, "replies": [ { "id": 1719941, "author_name": "abhishek", "author_url": "", "post_date": "03/12/2022 09:35:45", "content": "

This is a simple resize. So, this can be used for final submission. Any dataset with pre-processing of images is also allowed for GPU track but not for Innovation track. Further clarification is here: https://www.kaggle.com/c/ultra-mnist/discussion/312470

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1717545": "Hello everyone!\n\nI have created 5 different datasets for this competition. I have taken the original images and scaled them down to 5 different lower resolutions (128 x 128, 224 x 224, 512 x 512, 1024x1024 and\n1024 x1024 ). From previous competitions, we have seen that higher resolutions generally have better model performance but these provide a great starting point.\n\nAll 5 datasets also include the train.csv and submission.csv files in each of these datasets for convenience.\n\nThese datasets are also a lot smaller in scale compared to the original dataset which should make it easier to download locally or even use in a Kaggle kernel. The size of the dataset with 224x224 images is only 1.16 GB as compared to the original dataset which has a size of 34.5 GB.\nPlease let me know if you have any feedback ✨\n\nHere are the dataset links:\n\n(128 x 128 dataset) - https://www.kaggle.com/odins0n/ultramnist128x128 (593 MB)\n(224 x 224 dataset) - https://www.kaggle.com/odins0n/ultramnist224x224 (1.16 GB)\n(512 x 512 dataset) - https://www.kaggle.com/odins0n/ultramnist512x512 (2.97 GB)\n(1024 x 1024 dataset) - https://www.kaggle.com/odins0n/ultramnist1024x1024 (4.85 GB)\n(2048 x 2048 dataset) - https://www.kaggle.com/odins0n/ultramnist2048x2048 (15.46 GB)\n\nIf you want are interested in knowing how the data was compressed and which interpolation method was used, check out this notebook - https://www.kaggle.com/odins0n/ultramnist-image-resizer-512x512", "1717711": "Hey Sanskar,\n\nCan you clarify the note? I am a bit new to competitions and the wording is unclear to me. I thought we can use the competition data in any way (including resizing it) without any data from other sources. In my assumption, \"no external data\" means we can still use datasets like yours or ones that have more drastic changes (like the cleaned images from Remek).", "1719941": "This is a simple resize. So, this can be used for final submission. Any dataset with pre-processing of images is also allowed for GPU track but not for Innovation track. Further clarification is here: https://www.kaggle.com/c/ultra-mnist/discussion/312470" }, "source": "meta" }