![\"https://cdn.wccftech.com/wp-content/uploads/2018/03/NVIDIA-DGX-2-2.png\"](\"https://cdn.wccftech.com/wp-content/uploads/2018/03/NVIDIA-DGX-2-2.png\")
In this competition, we faced a lack of GPU, we used the only GCP available. And it was observed that more GPU could obviously help us a lot. For example, to train longer, use different parameters, image resolutions, and finally to get better results for sure (the tendency to that we faced also).
\n\nIn this competition, we faced a lack of GPU, we used the only GCP available. And it was observed that more GPU could obviously help us a lot. For example, to train longer, use different parameters, image resolutions, and finally to get better results for sure (the tendency to that we faced also).
\n\nWell it is crucial of course - I didn't have local GPU and used only the kernels - 30 hr/week - thats it. I could have performed far better if I had at least one GPU :)
", "rawMarkdown": "Well it is crucial of course - I didn't have local GPU and used only the kernels - 30 hr/week - thats it. I could have performed far better if I had at least one GPU :) ", "votes": 3 }, { "id": 676993, "postDate": "2019-11-19T17:33:08.640Z", "content": "I highly recommend vast.ai. Just tried it out in the last couple days. Didn't end up getting anything out of it but 4x2080ti for ~90cents/hour is way better than the other cloud providers.
", "rawMarkdown": "I highly recommend vast.ai. Just tried it out in the last couple days. Didn't end up getting anything out of it but 4x2080ti for ~90cents/hour is way better than the other cloud providers. ", "votes": 3, "replies": [ { "id": 676996, "postDate": "2019-11-19T17:35:22.443Z", "content": "~1$/h seems to be very good. \nThank you very much for such an advice.
", "rawMarkdown": "~1$/h seems to be very good. \nThank you very much for such an advice.", "votes": 1 } ] }, { "id": 676965, "postDate": "2019-11-19T16:55:45.613Z", "content": "> Where to get >1 GPUs for such competitions? Any advice?
\n\nIf you have some spare cash if think this would be a great choice.
\n\n
@carlolepelaars It's fantastic! \nUnfortunately, I don't have 200 000 $ for this system (((
", "rawMarkdown": "@carlolepelaars It's fantastic! \nUnfortunately, I don't have 200 000 $ for this system (((", "votes": 3 }, { "id": 677013, "postDate": "2019-11-19T17:55:44.180Z", "content": "Maybe if we do well in a competition like the Zillow Prize a few years back. 😊
\n\nhttps://www.kaggle.com/c/zillow-prize-1
", "rawMarkdown": "Maybe if we do well in a competition like the Zillow Prize a few years back. 😊 \n\nhttps://www.kaggle.com/c/zillow-prize-1", "votes": 1 }, { "id": 677017, "postDate": "2019-11-19T18:00:02.013Z", "rawMarkdown": "", "votes": 2, "isDeleted": true } ] }, { "id": 677072, "postDate": "2019-11-19T18:54:23.897Z", "content": "Well, you definitely need GPU for image competition. But for this competition, the computation resource is not crucial. My resnet34-FPN could get public 0.670 and private 0.665. The the whole pipeline can be trained on a 2080ti in ~7.5 hours (1 min/epoch x 30 epoch x 5 fold x 3 models = 450 mins).
", "rawMarkdown": "Well, you definitely need GPU for image competition. But for this competition, the computation resource is not crucial. My resnet34-FPN could get public 0.670 and private 0.665. The the whole pipeline can be trained on a 2080ti in ~7.5 hours (1 min/epoch x 30 epoch x 5 fold x 3 models = 450 mins). ", "votes": 2, "replies": [ { "id": 677105, "postDate": "2019-11-19T19:59:33.060Z", "rawMarkdown": "", "votes": 2, "isDeleted": true }, { "id": 677194, "postDate": "2019-11-19T22:59:36.837Z", "content": "But i bet it took a decent amount of experimentation to come up with that pipeline/optimize hyperparameters
", "rawMarkdown": "But i bet it took a decent amount of experimentation to come up with that pipeline/optimize hyperparameters" }, { "id": 677329, "postDate": "2019-11-20T03:48:52.140Z", "content": "Not really. I got that result 21 days ago, that was about 7 days after I joined the competition. I was using one 2080ti.
", "rawMarkdown": "Not really. I got that result 21 days ago, that was about 7 days after I joined the competition. I was using one 2080ti." } ] } ], "comments": [ { "id": 677158, "author_name": "Georgi Pamukov", "author_url": "", "post_date": "2019-11-19T21:33:40.303000", "content": "Well it is crucial of course - I didn't have local GPU and used only the kernels - 30 hr/week - thats it. I could have performed far better if I had at least one GPU :)
", "votes": 3, "replies": [] }, { "id": 676993, "author_name": "ryches", "author_url": "", "post_date": "2019-11-19T17:33:08.640000", "content": "I highly recommend vast.ai. Just tried it out in the last couple days. Didn't end up getting anything out of it but 4x2080ti for ~90cents/hour is way better than the other cloud providers.
", "votes": 3, "replies": [ { "id": 676996, "author_name": "Mukharbek Organokov", "author_url": "", "post_date": "2019-11-19T17:35:22.443000", "content": "~1$/h seems to be very good. \nThank you very much for such an advice.
", "votes": 1, "replies": [] } ] }, { "id": 676965, "author_name": "Carlo", "author_url": "", "post_date": "2019-11-19T16:55:45.613000", "content": "> Where to get >1 GPUs for such competitions? Any advice?
\n\nIf you have some spare cash if think this would be a great choice.
\n\n
@carlolepelaars It's fantastic! \nUnfortunately, I don't have 200 000 $ for this system (((
", "votes": 3, "replies": [] }, { "id": 677013, "author_name": "Carlo", "author_url": "", "post_date": "2019-11-19T17:55:44.180000", "content": "Maybe if we do well in a competition like the Zillow Prize a few years back. 😊
\n\nhttps://www.kaggle.com/c/zillow-prize-1
", "votes": 1, "replies": [] }, { "id": 677017, "author_name": "", "author_url": "", "post_date": "2019-11-19T18:00:02.013000", "content": "", "votes": 2, "replies": [] } ] }, { "id": 677072, "author_name": "Xuan Cao", "author_url": "", "post_date": "2019-11-19T18:54:23.897000", "content": "Well, you definitely need GPU for image competition. But for this competition, the computation resource is not crucial. My resnet34-FPN could get public 0.670 and private 0.665. The the whole pipeline can be trained on a 2080ti in ~7.5 hours (1 min/epoch x 30 epoch x 5 fold x 3 models = 450 mins).
", "votes": 2, "replies": [ { "id": 677105, "author_name": "", "author_url": "", "post_date": "2019-11-19T19:59:33.060000", "content": "", "votes": 2, "replies": [] }, { "id": 677194, "author_name": "sh", "author_url": "", "post_date": "2019-11-19T22:59:36.837000", "content": "But i bet it took a decent amount of experimentation to come up with that pipeline/optimize hyperparameters
", "votes": 0, "replies": [] }, { "id": 677329, "author_name": "Xuan Cao", "author_url": "", "post_date": "2019-11-20T03:48:52.140000", "content": "Not really. I got that result 21 days ago, that was about 7 days after I joined the competition. I was using one 2080ti.
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "676922": "In this competition, we faced a lack of GPU, we used the only GCP available. And it was observed that **more GPU could obviously help us a lot**. For example, to train longer, use different parameters, image resolutions, and finally **to get better results** for sure (the tendency to that we faced also). \n\n- *Please share your experience and thoughts.*\n- *Did someone use >1 GPUs in GCP, e.g.?*\n- *Where to get >1 GPUs for such competitions? Any advice?*", "677158": "Well it is crucial of course - I didn't have local GPU and used only the kernels - 30 hr/week - thats it. I could have performed far better if I had at least one GPU :) ", "676993": "I highly recommend vast.ai. Just tried it out in the last couple days. Didn't end up getting anything out of it but 4x2080ti for ~90cents/hour is way better than the other cloud providers. ", "676965": "> Where to get >1 GPUs for such competitions? Any advice?\n\nIf you have some spare cash if think this would be a great choice.\n\n\n\n**Datasheet:**\nhttps://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/dgx-2/dgx-2-print-datasheet-738070-nvidia-a4-web-uk.pdf", "677072": "Well, you definitely need GPU for image competition. But for this competition, the computation resource is not crucial. My resnet34-FPN could get public 0.670 and private 0.665. The the whole pipeline can be trained on a 2080ti in ~7.5 hours (1 min/epoch x 30 epoch x 5 fold x 3 models = 450 mins). " } }