First time when I had only couple weights of my models I used Iafoss inference kernel https://www.kaggle.com/iafoss/grapheme-fast-ai-starter-inference with my model - this is the only one thing that was replaced in his notebook.\nNext I started to write my own inference kernel to perform ensemble according to the Peter's inference kernel with ensembling https://www.kaggle.com/pestipeti/fast-ensemble-5-folds-20-minutes and the notebook mentioned above. Yeah I know about these tips what it's obviously better to steal someone's OK-working kernel. But there was one problem that forced me to write my own inference kernel notebook: for some (unknown) reason Peter's inference kernel with my model and weights gave me submission file with all zeros and Iafoss kernel itself work only with one weight at the time and hence just can't produce an ensemble.\nHere we come, I replaced nearly 10-12 lines of code in Iafoss inference kernel with some stolen loop prediction accumulations inside of the dict from Peter's notebook and 3 (maybe even 4) days keep watching these mostly 0.5422, 0.0614, Submission CSV Not Found results. Want to notice that 0.5422 is a broken score too because it appears despite of how many weights I use, i.e. whether I want to perform an ensemble or not.\nSomebody help me out, thanks in advance.
", "messages": [ { "id": "726512", "postDate": "01/23/2020 03:27:26", "content": "Keep getting broken score: 0.5422, 0.0614\n\nFirst time when I had only couple weights of my models I used Iafoss inference kernel https://www.kaggle.com/iafoss/grapheme-fast-ai-starter-inference with my model - this is the only one thing that was replaced in his notebook.\nNext I started to write my own inference kernel to perform ensemble according to the Peter's inference kernel with ensembling https://www.kaggle.com/pestipeti/fast-ensemble-5-folds-20-minutes and the notebook mentioned above. Yeah I know about these tips what it's obviously better to steal someone's OK-working kernel. But there was one problem that forced me to write my own inference kernel notebook: for some (unknown) reason Peter's inference kernel with my model and weights gave me submission file with all zeros and Iafoss kernel itself work only with one weight at the time and hence just can't produce an ensemble.\nHere we come, I replaced nearly 10-12 lines of code in Iafoss inference kernel with some stolen loop prediction accumulations inside of the dict from Peter's notebook and 3 (maybe even 4) days keep watching these mostly 0.5422, 0.0614, Submission CSV Not Found results. Want to notice that 0.5422 is a broken score too because it appears despite of how many weights I use, i.e. whether I want to perform an ensemble or not.\nSomebody help me out, thanks in advance.
", "rawMarkdown": "#### Keep getting broken score: 0.5422, 0.0614\n\nFirst time when I had only couple weights of my models I used Iafoss inference kernel *https://www.kaggle.com/iafoss/grapheme-fast-ai-starter-inference* with my model - this is the only one thing that was replaced in his notebook.\nNext I started to write my own inference kernel to perform ensemble according to the Peter's inference kernel with ensembling *https://www.kaggle.com/pestipeti/fast-ensemble-5-folds-20-minutes* and the notebook mentioned above. Yeah I know about these tips what it's obviously better to steal someone's OK-working kernel. But there was one problem that forced me to write my own inference kernel notebook: **for some (unknown) reason Peter's inference kernel with my model and weights gave me submission file with all zeros and Iafoss kernel itself work only with one weight at the time and hence just can't produce an ensemble**.\nHere we come, I replaced nearly 10-12 lines of code in Iafoss inference kernel with some stolen loop prediction accumulations inside of the dict from Peter's notebook and 3 (maybe even 4) days keep watching these **mostly 0.5422, 0.0614, Submission CSV Not Found** results. Want to notice that 0.5422 is a broken score too because it appears despite of how many weights I use, i.e. whether I want to perform an ensemble or not.\nSomebody help me out, thanks in advance.", "votes": null }, { "id": "726552", "postDate": "01/23/2020 04:34:48", "content": "\n\n\nfor some (unknown) reason Peter's inference kernel with my model and weights gave me submission file with all zeros.
\n
It happened to me too. Use the train parquet files (instead of the test) for debugging. Probably one of your cell will fail during commit.
", "rawMarkdown": "> for some (unknown) reason Peter's inference kernel with my model and weights gave me submission file with all zeros.\n\nIt happened to me too. Use the train parquet files (instead of the test) for debugging. Probably one of your cell will fail during commit.", "votes": null } ], "comments": [ { "id": 726552, "author_name": "pestipeti", "author_url": "", "post_date": "01/23/2020 04:34:48", "content": "\n\n\nfor some (unknown) reason Peter's inference kernel with my model and weights gave me submission file with all zeros.
\n
It happened to me too. Use the train parquet files (instead of the test) for debugging. Probably one of your cell will fail during commit.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "726512": "#### Keep getting broken score: 0.5422, 0.0614\n\nFirst time when I had only couple weights of my models I used Iafoss inference kernel *https://www.kaggle.com/iafoss/grapheme-fast-ai-starter-inference* with my model - this is the only one thing that was replaced in his notebook.\nNext I started to write my own inference kernel to perform ensemble according to the Peter's inference kernel with ensembling *https://www.kaggle.com/pestipeti/fast-ensemble-5-folds-20-minutes* and the notebook mentioned above. Yeah I know about these tips what it's obviously better to steal someone's OK-working kernel. But there was one problem that forced me to write my own inference kernel notebook: **for some (unknown) reason Peter's inference kernel with my model and weights gave me submission file with all zeros and Iafoss kernel itself work only with one weight at the time and hence just can't produce an ensemble**.\nHere we come, I replaced nearly 10-12 lines of code in Iafoss inference kernel with some stolen loop prediction accumulations inside of the dict from Peter's notebook and 3 (maybe even 4) days keep watching these **mostly 0.5422, 0.0614, Submission CSV Not Found** results. Want to notice that 0.5422 is a broken score too because it appears despite of how many weights I use, i.e. whether I want to perform an ensemble or not.\nSomebody help me out, thanks in advance.", "726552": "> for some (unknown) reason Peter's inference kernel with my model and weights gave me submission file with all zeros.\n\nIt happened to me too. Use the train parquet files (instead of the test) for debugging. Probably one of your cell will fail during commit." }, "source": "meta" }