I have estimated the energy consumption needed to train and predict my two final submissions.\nComputer: i9-7900x (10 cores) + 2 x RTX2080 Ti\nHours: 100 h\nAverage power during training (train stage-2) + prediction (test stage-2): 720 W\nTotal energy consumption: 72 kWh
\n\nAny body else can estimate the energy consumption?
", "messages": [ { "id": 672772, "postDate": "2019-11-14T06:24:40.113Z", "content": "I have estimated the energy consumption needed to train and predict my two final submissions.\nComputer: i9-7900x (10 cores) + 2 x RTX2080 Ti\nHours: 100 h\nAverage power during training (train stage-2) + prediction (test stage-2): 720 W\nTotal energy consumption: 72 kWh
\n\nAny body else can estimate the energy consumption?
", "rawMarkdown": "I have estimated the energy consumption needed to train and predict my two final submissions.\nComputer: i9-7900x (10 cores) + 2 x RTX2080 Ti\nHours: **100 h**\nAverage power during training (train stage-2) + prediction (test stage-2): 720 W\nTotal energy consumption: **72 kWh**\n\nAny body else can estimate the energy consumption?", "votes": 14 }, { "id": 672825, "postDate": "2019-11-14T07:52:03.470Z", "content": "I have exact an number for the whole competition. It is 264.552 kWh. I use separate smart socket plug to measure the consumption of my home ML experiments. 😊
", "rawMarkdown": "I have exact an number for the whole competition. It is **264.552 kWh**. I use separate smart socket plug to measure the consumption of my home ML experiments. 😊 ", "votes": 4 }, { "id": 672780, "postDate": "2019-11-14T06:31:43.237Z", "content": "I have the same specs as you. Training + inference 6 models x 5Folds takes upto 160h :D.
\n\nAddition information: Training a single AI model can emit as much carbon as five cars in their lifetimes
", "rawMarkdown": "I have the same specs as you. Training + inference 6 models x 5Folds takes upto **160h** :D. \n\nAddition information: [*Training a single AI model can emit as much carbon as five cars in their lifetimes*](https://www.technologyreview.com/s/613630/training-a-single-ai-model-can-emit-as-much-carbon-as-five-cars-in-their-lifetimes/?fbclid=IwAR16Fu50k33bSf7rZWlFqxZ75wyqdhLRY17rf4PP8BqnnDUUQaw_3iqu0y0)", "votes": 4, "replies": [ { "id": 672796, "postDate": "2019-11-14T06:57:19.717Z", "content": "Thanks @backaggle, I think it's a efficient solution for 5th place solution: 115 kWh
", "rawMarkdown": "Thanks @backaggle, I think it's a efficient solution for 5th place solution: **115 kWh**", "votes": 2 }, { "id": 673375, "postDate": "2019-11-14T22:45:20.117Z", "content": "Isn't Google Cloud carbon-neutral though? If you're training using GPUs on Kaggle, your emissions are zeroed out because they're hosted by Google.
", "rawMarkdown": "Isn't Google Cloud carbon-neutral though? If you're training using GPUs on Kaggle, your emissions are zeroed out because they're hosted by Google." }, { "id": 673962, "postDate": "2019-11-15T18:01:07.947Z", "content": "@robroseknows different things \"emissions\" and \"energy consumption\". If Google generate their energy from renewable sources, they don't emit CO2, but they have energy consumption.
", "rawMarkdown": "@robroseknows different things \"emissions\" and \"energy consumption\". If Google generate their energy from renewable sources, they don't emit CO2, but they have energy consumption." }, { "id": 674069, "postDate": "2019-11-15T21:31:59.950Z", "content": "@amezet True. I was referring more to the paper Bac provided though.
", "rawMarkdown": "@amezet True. I was referring more to the paper Bac provided though." } ] }, { "id": 673012, "postDate": "2019-11-14T11:55:16.027Z", "content": "The solutions being posted by the winning teams are fascinating; I have a lot to learn from them and am very grateful for all the sharing. Congratulations to them all!
\n\nHowever, the amount of compute used and training time is frightening. As a comparison, a single model trained on reset50 for just 4 hrs 30 min on a preemptible Google Cloud VM with 256*256 JPEG images (as in Jeremy Howard's notebook ) achieves 0.65 on the private LB. With the benefit of hindsight, I would love to know how low resource consumption could go to achieve 0.50 on the private LB.
\n\nI'm out of Google Cloud credits 😂 .
", "rawMarkdown": "The solutions being posted by the winning teams are fascinating; I have a lot to learn from them and am very grateful for all the sharing. Congratulations to them all! \n\nHowever, the amount of compute used and training time is frightening. As a comparison, a single model trained on reset50 for just 4 hrs 30 min on a preemptible Google Cloud VM with 256*256 JPEG images (as in Jeremy Howard's [notebook](https://www.kaggle.com/jhoward/cleaning-the-data-for-rapid-prototyping-fastai) ) achieves 0.65 on the private LB. With the benefit of hindsight, I would love to know how low resource consumption could go to achieve 0.50 on the private LB.\n\nI'm out of Google Cloud credits 😂 ." }, { "id": 673934, "postDate": "2019-11-15T17:09:29.197Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 672825, "author_name": "Dmitry A. Grechka", "author_url": "", "post_date": "2019-11-14T07:52:03.470000", "content": "I have exact an number for the whole competition. It is 264.552 kWh. I use separate smart socket plug to measure the consumption of my home ML experiments. 😊
", "votes": 4, "replies": [] }, { "id": 672780, "author_name": "cab", "author_url": "", "post_date": "2019-11-14T06:31:43.237000", "content": "I have the same specs as you. Training + inference 6 models x 5Folds takes upto 160h :D.
\n\nAddition information: Training a single AI model can emit as much carbon as five cars in their lifetimes
", "votes": 4, "replies": [ { "id": 672796, "author_name": "Antonio Marin", "author_url": "", "post_date": "2019-11-14T06:57:19.717000", "content": "Thanks @backaggle, I think it's a efficient solution for 5th place solution: 115 kWh
", "votes": 2, "replies": [] }, { "id": 673375, "author_name": "Rob Rose", "author_url": "", "post_date": "2019-11-14T22:45:20.117000", "content": "Isn't Google Cloud carbon-neutral though? If you're training using GPUs on Kaggle, your emissions are zeroed out because they're hosted by Google.
", "votes": 0, "replies": [] }, { "id": 673962, "author_name": "Antonio Marin", "author_url": "", "post_date": "2019-11-15T18:01:07.947000", "content": "@robroseknows different things \"emissions\" and \"energy consumption\". If Google generate their energy from renewable sources, they don't emit CO2, but they have energy consumption.
", "votes": 0, "replies": [] }, { "id": 674069, "author_name": "Rob Rose", "author_url": "", "post_date": "2019-11-15T21:31:59.950000", "content": "@amezet True. I was referring more to the paper Bac provided though.
", "votes": 0, "replies": [] } ] }, { "id": 673012, "author_name": "Alison Davey", "author_url": "", "post_date": "2019-11-14T11:55:16.027000", "content": "The solutions being posted by the winning teams are fascinating; I have a lot to learn from them and am very grateful for all the sharing. Congratulations to them all!
\n\nHowever, the amount of compute used and training time is frightening. As a comparison, a single model trained on reset50 for just 4 hrs 30 min on a preemptible Google Cloud VM with 256*256 JPEG images (as in Jeremy Howard's notebook ) achieves 0.65 on the private LB. With the benefit of hindsight, I would love to know how low resource consumption could go to achieve 0.50 on the private LB.
\n\nI'm out of Google Cloud credits 😂 .
", "votes": 0, "replies": [] }, { "id": 673934, "author_name": "", "author_url": "", "post_date": "2019-11-15T17:09:29.197000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "672772": "I have estimated the energy consumption needed to train and predict my two final submissions.\nComputer: i9-7900x (10 cores) + 2 x RTX2080 Ti\nHours: **100 h**\nAverage power during training (train stage-2) + prediction (test stage-2): 720 W\nTotal energy consumption: **72 kWh**\n\nAny body else can estimate the energy consumption?", "672825": "I have exact an number for the whole competition. It is **264.552 kWh**. I use separate smart socket plug to measure the consumption of my home ML experiments. 😊 ", "672780": "I have the same specs as you. Training + inference 6 models x 5Folds takes upto **160h** :D. \n\nAddition information: [*Training a single AI model can emit as much carbon as five cars in their lifetimes*](https://www.technologyreview.com/s/613630/training-a-single-ai-model-can-emit-as-much-carbon-as-five-cars-in-their-lifetimes/?fbclid=IwAR16Fu50k33bSf7rZWlFqxZ75wyqdhLRY17rf4PP8BqnnDUUQaw_3iqu0y0)", "673012": "The solutions being posted by the winning teams are fascinating; I have a lot to learn from them and am very grateful for all the sharing. Congratulations to them all! \n\nHowever, the amount of compute used and training time is frightening. As a comparison, a single model trained on reset50 for just 4 hrs 30 min on a preemptible Google Cloud VM with 256*256 JPEG images (as in Jeremy Howard's [notebook](https://www.kaggle.com/jhoward/cleaning-the-data-for-rapid-prototyping-fastai) ) achieves 0.65 on the private LB. With the benefit of hindsight, I would love to know how low resource consumption could go to achieve 0.50 on the private LB.\n\nI'm out of Google Cloud credits 😂 .", "673934": "" } }