I've been @nickuzmenkov's kernel as a Template, with a heavier model architecture (Not made public yet). And have been facing some errors with the TPU time limit.
\nThe time limit is 120 minutes for a TPU kernel, and my Inference kernel takes 203 minutes (almost double). Any ideas from the community on how I can speed this up ?
I've been @nickuzmenkov's kernel as a Template, with a heavier model architecture (Not made public yet). And have been facing some errors with the TPU time limit.
\nThe time limit is 120 minutes for a TPU kernel, and my Inference kernel takes 203 minutes (almost double). Any ideas from the community on how I can speed this up ?
Hello!
\nDo you mean GPU (TPUs are not allowed for inference)? What backbone you're using? (e.g. the last version of my inference notebook takes 14 minutes to run with 5 EfficientNetB4 ensemble).
", "rawMarkdown": "Hello! \n\nDo you mean GPU (TPUs are not allowed for inference)? What backbone you're using? (e.g. the last version of my inference notebook takes 14 minutes to run with 5 EfficientNetB4 ensemble).", "votes": null }, { "id": "1254227", "postDate": "03/27/2021 12:53:28", "content": "Hey, maybe I didn't read the rules carefully enough or something. But when I submitted a TPU Inference kernel, it said:
\nYour Notebook cannot use TPU in this competition.\nYour Notebook's runtime of 203 minutes exceeds this competition's TPU max of 120 minutes.\nBut, then when I submitted a much \"lighter\" Inference Kernel (by lighter I mean a kernel that took less time), It allowed me to submit (a TPU Kernel).
\nAlso, I'm using a SEResnet101 Backbone
Do you use both my train and inference notebook as a template or just a train notebook and do inference in yours?
", "rawMarkdown": "Do you use both my train and inference notebook as a template or just a train notebook and do inference in yours?", "votes": null }, { "id": "1254286", "postDate": "03/27/2021 13:46:30", "content": "I created a separate Train Notebook using your kernel as a template and then forked your Inference Kernel again to create an Inference Notebook.
", "rawMarkdown": "I created a separate Train Notebook using your kernel as a template and then forked your Inference Kernel again to create an Inference Notebook.", "votes": null }, { "id": "1254329", "postDate": "03/27/2021 14:39:07", "content": "That's strange.
\nIf you don't mind sharing, maybe I can find what's the problem is by looking through your notebook.
", "rawMarkdown": "That's strange.\n\nIf you don't mind sharing, maybe I can find what's the problem is by looking through your notebook.", "votes": null } ], "comments": [ { "id": 1254200, "author_name": "nickuzmenkov", "author_url": "", "post_date": "03/27/2021 12:29:45", "content": "Hello!
\nDo you mean GPU (TPUs are not allowed for inference)? What backbone you're using? (e.g. the last version of my inference notebook takes 14 minutes to run with 5 EfficientNetB4 ensemble).
", "votes": null, "replies": [ { "id": 1254227, "author_name": "sauravmaheshkar", "author_url": "", "post_date": "03/27/2021 12:53:28", "content": "Hey, maybe I didn't read the rules carefully enough or something. But when I submitted a TPU Inference kernel, it said:
\nYour Notebook cannot use TPU in this competition.\nYour Notebook's runtime of 203 minutes exceeds this competition's TPU max of 120 minutes.\nBut, then when I submitted a much \"lighter\" Inference Kernel (by lighter I mean a kernel that took less time), It allowed me to submit (a TPU Kernel).
\nAlso, I'm using a SEResnet101 Backbone
Do you use both my train and inference notebook as a template or just a train notebook and do inference in yours?
", "votes": null, "replies": [] }, { "id": 1254286, "author_name": "sauravmaheshkar", "author_url": "", "post_date": "03/27/2021 13:46:30", "content": "I created a separate Train Notebook using your kernel as a template and then forked your Inference Kernel again to create an Inference Notebook.
", "votes": null, "replies": [] }, { "id": 1254329, "author_name": "nickuzmenkov", "author_url": "", "post_date": "03/27/2021 14:39:07", "content": "That's strange.
\nIf you don't mind sharing, maybe I can find what's the problem is by looking through your notebook.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1254073": "I've been @nickuzmenkov's [kernel](https://www.kaggle.com/nickuzmenkov/pp2021-tpu-tf-inference/) as a Template, with a heavier model architecture (Not made public yet). And have been facing some errors with the TPU time limit. \n\n\nThe time limit is 120 minutes for a TPU kernel, and my Inference kernel takes 203 minutes (almost double). Any ideas from the community on how I can speed this up ?", "1254200": "Hello! \n\nDo you mean GPU (TPUs are not allowed for inference)? What backbone you're using? (e.g. the last version of my inference notebook takes 14 minutes to run with 5 EfficientNetB4 ensemble).", "1254227": "Hey, maybe I didn't read the rules carefully enough or something. But when I submitted a TPU Inference kernel, it said:\n\n```\nYour Notebook cannot use TPU in this competition.\nYour Notebook's runtime of 203 minutes exceeds this competition's TPU max of 120 minutes.\n```\n\nBut, then when I submitted a much \"lighter\" Inference Kernel (by lighter I mean a kernel that took less time), It allowed me to submit (**a TPU Kernel**).\n\nAlso, I'm using a `SEResnet101` Backbone", "1254263": "Do you use both my train and inference notebook as a template or just a train notebook and do inference in yours?", "1254286": "I created a separate Train Notebook using your kernel as a template and then forked your Inference Kernel again to create an Inference Notebook.", "1254329": "That's strange.\n\nIf you don't mind sharing, maybe I can find what's the problem is by looking through your notebook." }, "source": "meta" }