![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2167036%2F1358d32c642b0750f6ba3e0c0db372a1%2FScreenshot%20(146).png?generation=1606671092011827&alt=media\")
While playing around with Images during EDA I come to notice this unusual behaviour. Please check if anyone is getting the same after switching to different accelerators and reading in images. Well maybe this isn't a big deal but I thought it would be better if everyone notices this.
\n
The shape of the images while using CPU seems different. Later images have a shape something like (1, 1, 2, H, W)
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2167036%2F1c6932617b6d6303a43ba5303a151833%2FScreenshot%20(145).png?generation=1606671194826515&alt=media\")
While using TPU the channel seems to appear first. Some later images have a shape such as (C, H, W)
", "messages": [ { "id": "1095539", "postDate": "11/29/2020 17:51:29", "content": "While playing around with Images during EDA I come to notice this unusual behaviour. Please check if anyone is getting the same after switching to different accelerators and reading in images. Well maybe this isn't a big deal but I thought it would be better if everyone notices this.
\n
The shape of the images while using CPU seems different. Later images have a shape something like (1, 1, 2, H, W)
\n
While using TPU the channel seems to appear first. Some later images have a shape such as (C, H, W)
", "rawMarkdown": "While playing around with Images during EDA I come to notice this unusual behaviour. Please check if anyone is getting the same after switching to different accelerators and reading in images. Well maybe this isn't a big deal but I thought it would be better if everyone notices this. \n\n#### SORRY FOR THE IMAGE QUALITY. \n\n\n## ON CPU\n\n\n.png?generation=1606671092011827&alt=media)\n\nThe shape of the images while using CPU seems different. Later images have a shape something like (1, 1, 2, H, W)\n\n## ON TPU\n\n.png?generation=1606671194826515&alt=media)\n\nWhile using TPU the channel seems to appear first. Some later images have a shape such as (C, H, W)", "votes": null }, { "id": "1095569", "postDate": "11/29/2020 18:13:43", "content": "The two extra dimensions of size 1 can be got rid of by applying np.squeeze() or a similar function. They are meaningless.
\nI am guessing their appearance in one run and not the other is due to the tifffile package version difference between the two instances (TPU-enabled, and CPU-only).
\nYou need to watch out for channel-first and channel-last - some input images have the former, some the latter.
\nAlso, I don't think you will be able to use TPU in your submission notebook, as TPU requires internet access and the submission notebooks are not allowed internet access.
use of TPUs is allowed in the competition and the example notebooks demonstrates them as well. Hope this is helpful.
", "rawMarkdown": "use of TPUs is allowed in the competition and the example notebooks demonstrates them as well. Hope this is helpful.", "votes": null }, { "id": "1095632", "postDate": "11/29/2020 19:48:54", "content": "Hey @budsims ,
\nWhat about the use of TPU for the inference?
\nThere are no clear instructions around this, on this page.
TPU submission is allowed as long as the runtime of the notebook <= 3 hours.
", "rawMarkdown": "TPU submission is allowed as long as the runtime of the notebook <= 3 hours.", "votes": null }, { "id": "1095758", "postDate": "11/29/2020 23:11:03", "content": "@aroraaman See here. For some reason, TPU isn't allowed during submission.
", "rawMarkdown": "aroraaman See [here](https://www.kaggle.com/c/hubmap-kidney-segmentation/discussion/199750). For some reason, TPU isn't allowed during submission.", "votes": null }, { "id": "1095972", "postDate": "11/30/2020 05:30:07", "content": "Also, when I submitted with TPU, runtime was less that 3 hr.
", "rawMarkdown": "Also, when I submitted with TPU, runtime was less that 3 hr.", "votes": null }, { "id": "1096657", "postDate": "11/30/2020 17:09:40", "content": "@joshi98kishan -- we published the information here, but the short answer is train with TPU/Internet and submit with GPU/private https://www.kaggle.com/marcosnovaes/hubmap-unet-keras-model-fit-with-tpu
", "rawMarkdown": "joshi98kishan -- we published the information here, but the short answer is train with TPU/Internet and submit with GPU/private [https://www.kaggle.com/marcosnovaes/hubmap-unet-keras-model-fit-with-tpu](https://www.kaggle.com/marcosnovaes/hubmap-unet-keras-model-fit-with-tpu)", "votes": null }, { "id": "1096664", "postDate": "11/30/2020 17:17:16", "content": "Okay, thank you. @budsims
", "rawMarkdown": "Okay, thank you. @budsims", "votes": null } ], "comments": [ { "id": 1095569, "author_name": "igor14497", "author_url": "", "post_date": "11/29/2020 18:13:43", "content": "The two extra dimensions of size 1 can be got rid of by applying np.squeeze() or a similar function. They are meaningless.
\nI am guessing their appearance in one run and not the other is due to the tifffile package version difference between the two instances (TPU-enabled, and CPU-only).
\nYou need to watch out for channel-first and channel-last - some input images have the former, some the latter.
\nAlso, I don't think you will be able to use TPU in your submission notebook, as TPU requires internet access and the submission notebooks are not allowed internet access.
use of TPUs is allowed in the competition and the example notebooks demonstrates them as well. Hope this is helpful.
", "votes": null, "replies": [] }, { "id": 1095632, "author_name": "joshi98kishan", "author_url": "", "post_date": "11/29/2020 19:48:54", "content": "Hey @budsims ,
\nWhat about the use of TPU for the inference?
\nThere are no clear instructions around this, on this page.
TPU submission is allowed as long as the runtime of the notebook <= 3 hours.
", "votes": null, "replies": [] }, { "id": 1095758, "author_name": "matthewmasters", "author_url": "", "post_date": "11/29/2020 23:11:03", "content": "@aroraaman See here. For some reason, TPU isn't allowed during submission.
", "votes": null, "replies": [] }, { "id": 1095972, "author_name": "joshi98kishan", "author_url": "", "post_date": "11/30/2020 05:30:07", "content": "Also, when I submitted with TPU, runtime was less that 3 hr.
", "votes": null, "replies": [] }, { "id": 1096657, "author_name": "budsims", "author_url": "", "post_date": "11/30/2020 17:09:40", "content": "@joshi98kishan -- we published the information here, but the short answer is train with TPU/Internet and submit with GPU/private https://www.kaggle.com/marcosnovaes/hubmap-unet-keras-model-fit-with-tpu
", "votes": null, "replies": [] }, { "id": 1096664, "author_name": "joshi98kishan", "author_url": "", "post_date": "11/30/2020 17:17:16", "content": "Okay, thank you. @budsims
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1095539": "While playing around with Images during EDA I come to notice this unusual behaviour. Please check if anyone is getting the same after switching to different accelerators and reading in images. Well maybe this isn't a big deal but I thought it would be better if everyone notices this. \n\n#### SORRY FOR THE IMAGE QUALITY. \n\n\n## ON CPU\n\n\n.png?generation=1606671092011827&alt=media)\n\nThe shape of the images while using CPU seems different. Later images have a shape something like (1, 1, 2, H, W)\n\n## ON TPU\n\n.png?generation=1606671194826515&alt=media)\n\nWhile using TPU the channel seems to appear first. Some later images have a shape such as (C, H, W)", "1095569": "The two extra dimensions of size 1 can be got rid of by applying np.squeeze() or a similar function. They are meaningless.\nI am guessing their appearance in one run and not the other is due to the tifffile package version difference between the two instances (TPU-enabled, and CPU-only).\nYou need to watch out for channel-first and channel-last - some input images have the former, some the latter.\nAlso, I don't think you will be able to use TPU in your submission notebook, as TPU requires internet access and the submission notebooks are not allowed internet access.", "1095597": "use of TPUs is allowed in the competition and the example notebooks demonstrates them as well. Hope this is helpful.", "1095632": "Hey @budsims ,\nWhat about the use of TPU for the inference? \nThere are no clear instructions around this, on this [page](https://www.kaggle.com/c/hubmap-kidney-segmentation/overview/code-requirements).", "1095755": "TPU submission is allowed as long as the runtime of the notebook <= 3 hours.", "1095758": "aroraaman See [here](https://www.kaggle.com/c/hubmap-kidney-segmentation/discussion/199750). For some reason, TPU isn't allowed during submission.", "1095972": "Also, when I submitted with TPU, runtime was less that 3 hr.", "1096657": "joshi98kishan -- we published the information here, but the short answer is train with TPU/Internet and submit with GPU/private [https://www.kaggle.com/marcosnovaes/hubmap-unet-keras-model-fit-with-tpu](https://www.kaggle.com/marcosnovaes/hubmap-unet-keras-model-fit-with-tpu)", "1096664": "Okay, thank you. @budsims" }, "source": "meta" }