I'm trying to use TF Swin tile-based model (using image preprocessing pipeline by @jirkaborovec), but I'm getting Notebook Threw Exception errors at the same time (approx. ~45 min after submitting). I wonder what's the reason for it? Here's my submission notebook:
\nhttps://www.kaggle.com/code/atamazian/ubc-ocean-tf-swin-large-inference-debug
I'm trying to use TF Swin tile-based model (using image preprocessing pipeline by @jirkaborovec), but I'm getting Notebook Threw Exception errors at the same time (approx. ~45 min after submitting). I wonder what's the reason for it? Here's my submission notebook:
\nhttps://www.kaggle.com/code/atamazian/ubc-ocean-tf-swin-large-inference-debug
Could you share the exception as the linked notebook does not show any outputs?
", "rawMarkdown": "Could you share the exception as the linked notebook does not show any outputs?", "votes": 1, "replies": [ { "id": 2554160, "postDate": "2023-12-08T20:48:37.260Z", "content": "The notebook successfully runs on the public test set, but crashes on the private test set. Therefore, I'm unable to provide any details (that's the problem).
", "rawMarkdown": "The notebook successfully runs on the public test set, but crashes on the private test set. Therefore, I'm unable to provide any details (that's the problem)." }, { "id": 2555148, "postDate": "2023-12-09T17:57:53.043Z", "content": "I managed to solve this issue by reducing slicing size from 2048 to 1024.
", "rawMarkdown": "I managed to solve this issue by reducing slicing size from 2048 to 1024.", "votes": 1, "replies": [ { "id": 2557236, "postDate": "2023-12-11T09:58:08.840Z", "content": "Do you think it's helpful? does it change the prediction after reducing the slice size?
", "rawMarkdown": "Do you think it's helpful? does it change the prediction after reducing the slice size?" }, { "id": 2558847, "postDate": "2023-12-12T11:47:19.717Z", "content": "In fact it does - my model scored terrible since I trained it with slice size of 2048
", "rawMarkdown": "In fact it does - my model scored terrible since I trained it with slice size of 2048" }, { "id": 2558862, "postDate": "2023-12-12T11:56:02.987Z", "content": "So you still kept the scale factor 0.25 also for your 1024px tiles?
", "rawMarkdown": "So you still kept the scale factor 0.25 also for your 1024px tiles?" }, { "id": 2558949, "postDate": "2023-12-12T13:26:44.370Z", "content": "No, TF Swin has weights only for image sizes of 224 and 384. I've modified your code a little, replacing scale parameter with direct definition of final image size (in my case it's 384). Scale parameter of 0.25 corresponds to image size of 512 (for slice size of 2048), and 256 (for slice size of 1024).
The notebook successfully runs
", "rawMarkdown": "The notebook successfully runs" } ], "comments": [ { "id": 2554153, "author_name": "Jirka", "author_url": "", "post_date": "2023-12-08T20:26:26.687000", "content": "Could you share the exception as the linked notebook does not show any outputs?
", "votes": 1, "replies": [ { "id": 2554160, "author_name": "Araik Tamazian", "author_url": "", "post_date": "2023-12-08T20:48:37.260000", "content": "The notebook successfully runs on the public test set, but crashes on the private test set. Therefore, I'm unable to provide any details (that's the problem).
", "votes": 0, "replies": [] }, { "id": 2555148, "author_name": "Araik Tamazian", "author_url": "", "post_date": "2023-12-09T17:57:53.043000", "content": "I managed to solve this issue by reducing slicing size from 2048 to 1024.
", "votes": 1, "replies": [ { "id": 2557236, "author_name": "Batool Alghamdi", "author_url": "", "post_date": "2023-12-11T09:58:08.840000", "content": "Do you think it's helpful? does it change the prediction after reducing the slice size?
", "votes": 0, "replies": [] }, { "id": 2558847, "author_name": "Araik Tamazian", "author_url": "", "post_date": "2023-12-12T11:47:19.717000", "content": "In fact it does - my model scored terrible since I trained it with slice size of 2048
", "votes": 0, "replies": [] }, { "id": 2558862, "author_name": "Jirka", "author_url": "", "post_date": "2023-12-12T11:56:02.987000", "content": "So you still kept the scale factor 0.25 also for your 1024px tiles?
", "votes": 0, "replies": [] }, { "id": 2558949, "author_name": "Araik Tamazian", "author_url": "", "post_date": "2023-12-12T13:26:44.370000", "content": "No, TF Swin has weights only for image sizes of 224 and 384. I've modified your code a little, replacing scale parameter with direct definition of final image size (in my case it's 384). Scale parameter of 0.25 corresponds to image size of 512 (for slice size of 2048), and 256 (for slice size of 1024).
The notebook successfully runs
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "2553985": "I'm trying to use TF Swin tile-based model (using image preprocessing pipeline by @jirkaborovec), but I'm getting `Notebook Threw Exception ` errors at the same time (approx. ~45 min after submitting). I wonder what's the reason for it? Here's my submission notebook:\nhttps://www.kaggle.com/code/atamazian/ubc-ocean-tf-swin-large-inference-debug", "2554153": "Could you share the exception as the linked notebook does not show any outputs?", "2555953": "The notebook successfully runs" } }