![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2386017%2F1df6def2496017f9b736826b3acafb4c%2Fresult.jpg?generation=1656768939441397&alt=media\")
Initially, I thought there was an error with the RLEs but it turns out that quite a lot of the HPA images aren't (3000x3000).
\nThank you @urosjarc for catching my mistake and helping :)
\nI thought all the HPA images were 3000x3000 because in the data section the hosts say:
\n\n\n\"All HPA images are 3000 x 3000 pixels with a tissue area within the image around 2500 x 2500 pixels\"
\n
I didn't want to just delete this post after my realization in case anyone else makes the same mistake as me in assuming they are all 3000x3000 (a mistake I really shouldn't have made… but oh well). As such I renamed the post title to be more informative and will include the distribution of image sizes below:
\n3000 326\n2631 2\n2416 2\n2942 2\n2790 2\n2764 2\n2654 2\n2539 1\n2680 1\n2727 1\n2308 1\n2867 1\n2783 1\n2869 1\n2760 1\n2630 1\n2511 1\n2593 1\n2675 1\n3070 1\nInitially, I thought there was an error with the RLEs but it turns out that quite a lot of the HPA images aren't (3000x3000).
\nThank you @urosjarc for catching my mistake and helping :)
\nI thought all the HPA images were 3000x3000 because in the data section the hosts say:
\n\n\n\"All HPA images are 3000 x 3000 pixels with a tissue area within the image around 2500 x 2500 pixels\"
\n
I didn't want to just delete this post after my realization in case anyone else makes the same mistake as me in assuming they are all 3000x3000 (a mistake I really shouldn't have made… but oh well). As such I renamed the post title to be more informative and will include the distribution of image sizes below:
\n3000 326\n2631 2\n2416 2\n2942 2\n2790 2\n2764 2\n2654 2\n2539 1\n2680 1\n2727 1\n2308 1\n2867 1\n2783 1\n2869 1\n2760 1\n2630 1\n2511 1\n2593 1\n2675 1\n3070 1\nProbably you are using decoding with 3000x3000 and it gets corrupted since this image is 2630x2630.
\nTry to decode it with this function:
\ndef decode_rle(rle: str, shape=(2630, 2630)):\n rle = [int(i) for i in rle.split(' ')]\n pairs = list(zip(rle[0::2], rle[1::2]))\n\n p_loc = []\n\n for start, length in pairs:\n for p_pos in range(start, start + length):\n p_loc.append((p_pos % shape[1], p_pos // shape[0]))\n\n canvas = np.zeros(shape, dtype=np.uint8).T\n canvas[tuple(zip(*p_loc))] = 1.0\n return canvas\nHere are my results with overlay…
\n
You’re absolutely correct.
\nI was confused by the hosts specifically saying:
\n“All HPA images are 3000 x 3000 pixels”
\n—-
\nThanks for this catch, I’ll double check things when I get back to a computer but this makes a lot of sense.
", "rawMarkdown": "You’re absolutely correct.\n\nI was confused by the hosts specifically saying:\n\n“All HPA images are 3000 x 3000 pixels”\n\n—-\n\nThanks for this catch, I’ll double check things when I get back to a computer but this makes a lot of sense.", "votes": null } ], "comments": [ { "id": 1840682, "author_name": "urosjarc", "author_url": "", "post_date": "07/02/2022 13:39:15", "content": "Probably you are using decoding with 3000x3000 and it gets corrupted since this image is 2630x2630.
\nTry to decode it with this function:
\ndef decode_rle(rle: str, shape=(2630, 2630)):\n rle = [int(i) for i in rle.split(' ')]\n pairs = list(zip(rle[0::2], rle[1::2]))\n\n p_loc = []\n\n for start, length in pairs:\n for p_pos in range(start, start + length):\n p_loc.append((p_pos % shape[1], p_pos // shape[0]))\n\n canvas = np.zeros(shape, dtype=np.uint8).T\n canvas[tuple(zip(*p_loc))] = 1.0\n return canvas\nHere are my results with overlay…
\n
You’re absolutely correct.
\nI was confused by the hosts specifically saying:
\n“All HPA images are 3000 x 3000 pixels”
\n—-
\nThanks for this catch, I’ll double check things when I get back to a computer but this makes a lot of sense.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1840644": "Initially, I thought there was an error with the RLEs but it turns out that quite a lot of the HPA images aren't (3000x3000). \n\nThank you @urosjarc for catching my mistake and helping :)\n\n---\n\nI thought all the HPA images were 3000x3000 because in the [**data**](https://www.kaggle.com/competitions/hubmap-organ-segmentation/data) section the hosts say:\n\n> \"All HPA images are 3000 x 3000 pixels with a tissue area within the image around 2500 x 2500 pixels\"\n\nI didn't want to just delete this post after my realization in case anyone else makes the same mistake as me in assuming they are all 3000x3000 (a mistake I really shouldn't have made... but oh well). As such I renamed the post title to be more informative and will include the distribution of image sizes below:\n\n---\n\n```\n3000 326\n2631 2\n2416 2\n2942 2\n2790 2\n2764 2\n2654 2\n2539 1\n2680 1\n2727 1\n2308 1\n2867 1\n2783 1\n2869 1\n2760 1\n2630 1\n2511 1\n2593 1\n2675 1\n3070 1\n```", "1840682": "Probably you are using decoding with 3000x3000 and it gets corrupted since this image is 2630x2630.\n\nTry to decode it with this function:\n```\ndef decode_rle(rle: str, shape=(2630, 2630)):\n rle = [int(i) for i in rle.split(' ')]\n pairs = list(zip(rle[0::2], rle[1::2]))\n\n p_loc = []\n\n for start, length in pairs:\n for p_pos in range(start, start + length):\n p_loc.append((p_pos % shape[1], p_pos // shape[0]))\n\n canvas = np.zeros(shape, dtype=np.uint8).T\n canvas[tuple(zip(*p_loc))] = 1.0\n return canvas\n```\n\nHere are my results with overlay...\n\n", "1840690": "You’re absolutely correct.\n\nI was confused by the hosts specifically saying:\n\n“All HPA images are 3000 x 3000 pixels”\n\n—-\n\nThanks for this catch, I’ll double check things when I get back to a computer but this makes a lot of sense." }, "source": "meta" }