{ "id": 120658, "title": "substantial code competition", "url": "/competitions/tensorflow2-question-answering/discussion/120658", "author_name": "", "post_date": "2019-12-07T18:34:44.616672100Z", "votes": 10, "comment_count": 4, "views": 0, "content": "

I trained my current model about a month ago. After that, I experienced life like a ‘’blind‘’ man on and off.
\n\"\"\nRead the rules, check the code, optimize the algorithm, refactor the code and go back to reading the rules and start again. That's what I'll do later. So I summarize some key points may help:\n- Private dataset has about 3000 samples, so repeated redundant variables may cause memory shortage, Sometimes Notebook Exceeded Allowed Compute, sometimes 0.00 Lb score.\n- If you use tensorflow and don't use allow_growth, it's easy to run out of gpu memory.\n- When using jupyter, sometimes a submission will be generated even if there is a bug in the running. But the generated submission is sample_submission, which will result in 0.00lb. \n- To avoid this potentially misleading result, it might be better to use script.

\n\n

Finally, good luck guys

", "messages": [ { "id": "689961", "postDate": "12/07/2019 18:34:44", "content": "

I trained my current model about a month ago. After that, I experienced life like a ‘’blind‘’ man on and off.
\n\"\"\nRead the rules, check the code, optimize the algorithm, refactor the code and go back to reading the rules and start again. That's what I'll do later. So I summarize some key points may help:\n- Private dataset has about 3000 samples, so repeated redundant variables may cause memory shortage, Sometimes Notebook Exceeded Allowed Compute, sometimes 0.00 Lb score.\n- If you use tensorflow and don't use allow_growth, it's easy to run out of gpu memory.\n- When using jupyter, sometimes a submission will be generated even if there is a bug in the running. But the generated submission is sample_submission, which will result in 0.00lb. \n- To avoid this potentially misleading result, it might be better to use script.

\n\n

Finally, good luck guys

", "rawMarkdown": "I trained my current model about a month ago. After that, I experienced life like a ‘’blind‘’ man on and off. \n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1109229%2Fc75ab330c65dd6c0351fd0189c56b6cb%2FFireShot%20Capture%20001%20-%20TensorFlow%202.0%20Question%20Answering%20-%20Kaggle%20-%20www.kaggle.com.png?generation=1575742174278479&alt=media)\nRead the rules, check the code, optimize the algorithm, refactor the code and go back to reading the rules and start again. That's what I'll do later. So I summarize some key points may help:\n- Private dataset has about 3000 samples, so repeated redundant variables may cause memory shortage, Sometimes `Notebook Exceeded Allowed Compute`, sometimes `0.00 Lb score`.\n- If you use tensorflow and don't use `allow_growth `, it's easy to run out of gpu memory.\n- When using jupyter, sometimes a submission will be generated even if there is a bug in the running. But the generated submission is sample_submission, which will result in 0.00lb. \n- To avoid this potentially misleading result, it might be better to use script.\n\nFinally, good luck guys", "votes": null }, { "id": "690222", "postDate": "12/08/2019 06:41:18", "content": "

@wochidadonggua did you train your model from sratch? or finetuning?

", "rawMarkdown": "wochidadonggua did you train your model from sratch? or finetuning?", "votes": null }, { "id": "690790", "postDate": "12/09/2019 06:36:42", "content": "

finetuning

", "rawMarkdown": "finetuning", "votes": null }, { "id": "697744", "postDate": "12/18/2019 10:51:35", "content": "

@wochidadonggua Are you still facing this issue?

", "rawMarkdown": "wochidadonggua Are you still facing this issue?", "votes": null }, { "id": "697963", "postDate": "12/18/2019 15:47:56", "content": "

No

", "rawMarkdown": "No", "votes": null } ], "comments": [ { "id": 690222, "author_name": "axel81", "author_url": "", "post_date": "12/08/2019 06:41:18", "content": "

@wochidadonggua did you train your model from sratch? or finetuning?

", "votes": null, "replies": [ { "id": 690790, "author_name": "wochidadonggua", "author_url": "", "post_date": "12/09/2019 06:36:42", "content": "

finetuning

", "votes": null, "replies": [] } ] }, { "id": 697744, "author_name": "rohitagarwal", "author_url": "", "post_date": "12/18/2019 10:51:35", "content": "

@wochidadonggua Are you still facing this issue?

", "votes": null, "replies": [ { "id": 697963, "author_name": "wochidadonggua", "author_url": "", "post_date": "12/18/2019 15:47:56", "content": "

No

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "689961": "I trained my current model about a month ago. After that, I experienced life like a ‘’blind‘’ man on and off. \n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1109229%2Fc75ab330c65dd6c0351fd0189c56b6cb%2FFireShot%20Capture%20001%20-%20TensorFlow%202.0%20Question%20Answering%20-%20Kaggle%20-%20www.kaggle.com.png?generation=1575742174278479&alt=media)\nRead the rules, check the code, optimize the algorithm, refactor the code and go back to reading the rules and start again. That's what I'll do later. So I summarize some key points may help:\n- Private dataset has about 3000 samples, so repeated redundant variables may cause memory shortage, Sometimes `Notebook Exceeded Allowed Compute`, sometimes `0.00 Lb score`.\n- If you use tensorflow and don't use `allow_growth `, it's easy to run out of gpu memory.\n- When using jupyter, sometimes a submission will be generated even if there is a bug in the running. But the generated submission is sample_submission, which will result in 0.00lb. \n- To avoid this potentially misleading result, it might be better to use script.\n\nFinally, good luck guys", "690222": "wochidadonggua did you train your model from sratch? or finetuning?", "690790": "finetuning", "697744": "wochidadonggua Are you still facing this issue?", "697963": "No" }, "source": "meta" }