![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2721247%2F768384d8cbf24a0e6c94586fe95053d5%2Fgpu.png?generation=1575660356465611&alt=media\")
\nWow 25GB of ram!\nAnd:\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2721247%2Fa3747c732299d68d13cd428b38d944d6%2Fgpu2.png?generation=1575660413437267&alt=media\")
\n12 hours of P100 access, which is longer than former kaggle (sometimes will still K80 but just need 1-2 times reset instead of 20+)!After oneday crash of my colab script, I found 1 tips:\n1. download my colab_script.ipynb and open it as text file\n2. the firstline will be: \n\n{\n \"nbformat\": 4,\n \"nbformat_minor\": 0,\n \"metadata\": {\n \"colab\": {\n \"name\": \"7 reload and train bert.ipynb\",\n \"provenance\": [],\n \"collapsed_sections\": [],\n \"toc_visible\": true,\n \"machine_shape\": \"hm\"\n },\n \"language_info\":\n...\n\ni did experiment and found only 1 difference: \"machine_shape\": \"hm\" is absent in normal script.\nSO, I add this key and value to see if those scripts become high-performance ones. and see:
For GPU:\n\nWow 25GB of ram!\nAnd:\n\n12 hours of P100 access, which is longer than former kaggle (sometimes will still K80 but just need 1-2 times reset instead of 20+)!
For TPU:\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2721247%2Fdcca580c2e17304cf0fa1bbada5f09a0%2Ftpu.png?generation=1575660508828088&alt=media\")
\n35GB ram 😍
Last words: Colab Premium, please take my money just as GabeN did!
", "messages": [ { "id": "689328", "postDate": "12/06/2019 19:30:10", "content": "After oneday crash of my colab script, I found 1 tips:\n1. download my colab_script.ipynb and open it as text file\n2. the firstline will be: \n\n{\n \"nbformat\": 4,\n \"nbformat_minor\": 0,\n \"metadata\": {\n \"colab\": {\n \"name\": \"7 reload and train bert.ipynb\",\n \"provenance\": [],\n \"collapsed_sections\": [],\n \"toc_visible\": true,\n \"machine_shape\": \"hm\"\n },\n \"language_info\":\n...\n\ni did experiment and found only 1 difference: \"machine_shape\": \"hm\" is absent in normal script.\nSO, I add this key and value to see if those scripts become high-performance ones. and see:
For GPU:\n\nWow 25GB of ram!\nAnd:\n\n12 hours of P100 access, which is longer than former kaggle (sometimes will still K80 but just need 1-2 times reset instead of 20+)!
For TPU:\n\n35GB ram 😍
Last words: Colab Premium, please take my money just as GabeN did!
", "rawMarkdown": "After oneday crash of my colab script, I found 1 tips:\n1. download my colab_script.ipynb and open it as text file\n2. the firstline will be: \n```\n{\n \"nbformat\": 4,\n \"nbformat_minor\": 0,\n \"metadata\": {\n \"colab\": {\n \"name\": \"7 reload and train bert.ipynb\",\n \"provenance\": [],\n \"collapsed_sections\": [],\n \"toc_visible\": true,\n \"machine_shape\": \"hm\"\n },\n \"language_info\":\n...\n```\ni did experiment and found only 1 difference: \"machine_shape\": \"hm\" is absent in normal script.\nSO, I add this key and value to see if those scripts become high-performance ones. and see:\n\nFor GPU:\n\nWow 25GB of ram!\nAnd:\n\n12 hours of P100 access, which is longer than former kaggle (sometimes will still K80 but **just need 1-2 times reset instead of 20+**)!\n\nFor TPU:\n\n35GB ram 😍 \n\nLast words: Colab Premium, please take my money just as GabeN did!", "votes": null }, { "id": "689341", "postDate": "12/06/2019 19:48:18", "content": "@httpwwwfszyc i will try your tip. Thanks :)
\n\njust need 1-2 reset instead of 20+ -- whats this?
", "rawMarkdown": "httpwwwfszyc i will try your tip. Thanks :)\n\n just need 1-2 reset instead of 20+ -- whats this?", "votes": null }, { "id": "689423", "postDate": "12/06/2019 23:10:59", "content": "when you start the script, GPU will be P100 in most of the time
", "rawMarkdown": "when you start the script, GPU will be P100 in most of the time", "votes": null }, { "id": "697959", "postDate": "12/18/2019 15:32:28", "content": "This is super information. This needs to be shared with more folks.\n+1 upvote.
", "rawMarkdown": "This is super information. This needs to be shared with more folks.\n+1 upvote.", "votes": null } ], "comments": [ { "id": 689341, "author_name": "seshurajup", "author_url": "", "post_date": "12/06/2019 19:48:18", "content": "@httpwwwfszyc i will try your tip. Thanks :)
\n\njust need 1-2 reset instead of 20+ -- whats this?
", "votes": null, "replies": [ { "id": 689423, "author_name": "httpwwwfszyc", "author_url": "", "post_date": "12/06/2019 23:10:59", "content": "when you start the script, GPU will be P100 in most of the time
", "votes": null, "replies": [] } ] }, { "id": 697959, "author_name": "rohitagarwal", "author_url": "", "post_date": "12/18/2019 15:32:28", "content": "This is super information. This needs to be shared with more folks.\n+1 upvote.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "689328": "After oneday crash of my colab script, I found 1 tips:\n1. download my colab_script.ipynb and open it as text file\n2. the firstline will be: \n```\n{\n \"nbformat\": 4,\n \"nbformat_minor\": 0,\n \"metadata\": {\n \"colab\": {\n \"name\": \"7 reload and train bert.ipynb\",\n \"provenance\": [],\n \"collapsed_sections\": [],\n \"toc_visible\": true,\n \"machine_shape\": \"hm\"\n },\n \"language_info\":\n...\n```\ni did experiment and found only 1 difference: \"machine_shape\": \"hm\" is absent in normal script.\nSO, I add this key and value to see if those scripts become high-performance ones. and see:\n\nFor GPU:\n\nWow 25GB of ram!\nAnd:\n\n12 hours of P100 access, which is longer than former kaggle (sometimes will still K80 but **just need 1-2 times reset instead of 20+**)!\n\nFor TPU:\n\n35GB ram 😍 \n\nLast words: Colab Premium, please take my money just as GabeN did!", "689341": "httpwwwfszyc i will try your tip. Thanks :)\n\n just need 1-2 reset instead of 20+ -- whats this?", "689423": "when you start the script, GPU will be P100 in most of the time", "697959": "This is super information. This needs to be shared with more folks.\n+1 upvote." }, "source": "meta" }