Given the size of this dataset, working on Kaggle or a local machine (like my 16 GB laptop) probably won't be feasible. Recently, however, I signed up for Google Cloud and after registering for a paid account, I got $300 of compute credits. You can provision some very beefy machines there for not a lot of cost. Adding in GPUs is where the real costs seem to add up, but I was able to run some large GAN models (for a different project) over a week and only spent about half my credits. You can also transfer your notebooks directly to Google Cloud from Kaggle and they provide a template for data transfer as well.
\nLooking at machine configurations, something on the order of 96 CPUs and 386 GB of memory runs around $2.50 - $3.50 per hour. 100 hours of free compute may be available assuming you get the introductory offer credits.
\nQuick update - I'm finding that I'm bumping into quota limits on Google Cloud. I'm able to provision a machine with 24 CPUs, but not more. I have requested an increase in that limit - we'll see if it is approved. 🤞 Regardless, 24 CPUs is way more than I have available locally or on Kaggle. 👍
\nAnother update - getting quotas increased isn't too difficult as long as the request isn't extravagant. Moderate requests are approved through an automated system within a few minutes.
\nUpdate 1 May 2024
\nAnother possibility is Saturn Cloud. They have preconfigured instances with Python or R. I used them in the past and was really impressed with the interface and the tools. I'm not sure how much free compute they offer in the Hosted Free plan, but they give a certain number of free hours compute per month and then pay as you go after that.
Given the size of this dataset, working on Kaggle or a local machine (like my 16 GB laptop) probably won't be feasible. Recently, however, I signed up for Google Cloud and after registering for a paid account, I got $300 of compute credits. You can provision some very beefy machines there for not a lot of cost. Adding in GPUs is where the real costs seem to add up, but I was able to run some large GAN models (for a different project) over a week and only spent about half my credits. You can also transfer your notebooks directly to Google Cloud from Kaggle and they provide a template for data transfer as well.
\nLooking at machine configurations, something on the order of 96 CPUs and 386 GB of memory runs around $2.50 - $3.50 per hour. 100 hours of free compute may be available assuming you get the introductory offer credits.
\nQuick update - I'm finding that I'm bumping into quota limits on Google Cloud. I'm able to provision a machine with 24 CPUs, but not more. I have requested an increase in that limit - we'll see if it is approved. 🤞 Regardless, 24 CPUs is way more than I have available locally or on Kaggle. 👍
\nAnother update - getting quotas increased isn't too difficult as long as the request isn't extravagant. Moderate requests are approved through an automated system within a few minutes.
\nUpdate 1 May 2024
\nAnother possibility is Saturn Cloud. They have preconfigured instances with Python or R. I used them in the past and was really impressed with the interface and the tools. I'm not sure how much free compute they offer in the Hosted Free plan, but they give a certain number of free hours compute per month and then pay as you go after that.
the trick to cost saving is really:
\nyou don't just blindly start training. for every decision to use the cloud to train, you are pretty confidently that the training should yield an useful/successful experiment
\n(for every failed experiment, it is $$$ wasted. this is actually the same of any project management)
", "rawMarkdown": "the trick to cost saving is really:\n1. use cloud for training and not for development\n2. before training in the cloud, you should have already estimated the end results of training and are pretty confident of the results\n\nyou don't just blindly start training. for every decision to use the cloud to train, you are pretty confidently that the training should yield an useful/successful experiment\n\n(for every failed experiment, it is $$$ wasted. this is actually the same of any project management)", "votes": null }, { "id": "2781622", "postDate": "04/28/2024 21:31:54", "content": "@hengck23 can you elaborate point 1 please? I did not get you properly!
", "rawMarkdown": "hengck23 can you elaborate point 1 please? I did not get you properly!", "votes": null }, { "id": "2782952", "postDate": "04/29/2024 14:12:01", "content": "Correct me if I'm wrong, but doesn't Google still charge you for the used credits, after they expire? Like, if you use $50 from the initial $300 credits, you'd still get a bill of $50, after a few months.
", "rawMarkdown": "Correct me if I'm wrong, but doesn't Google still charge you for the used credits, after they expire? Like, if you use $50 from the initial $300 credits, you'd still get a bill of $50, after a few months.", "votes": null }, { "id": "2783238", "postDate": "04/29/2024 16:27:55", "content": "Not that I'm aware. I think the longer term bill may come from inadvertently leaving an instance running or leaving data stored, both of which will create surprise bills.
\nOf course, I could be wrong. I've not seen that happen (yet) and I'll double check the details there. Also if anyone else has experience with this, please let us know.
", "rawMarkdown": "Not that I'm aware. I think the longer term bill may come from inadvertently leaving an instance running or leaving data stored, both of which will create surprise bills.\n\nOf course, I could be wrong. I've not seen that happen (yet) and I'll double check the details there. Also if anyone else has experience with this, please let us know.", "votes": null }, { "id": "2785731", "postDate": "05/01/2024 02:27:18", "content": "I think the point is to make sure that we develop features, and the code can run properly in the local environment before running it on the cloud, which would cost money to debug and develop.
", "rawMarkdown": "I think the point is to make sure that we develop features, and the code can run properly in the local environment before running it on the cloud, which would cost money to debug and develop.", "votes": null }, { "id": "2785854", "postDate": "05/01/2024 04:11:49", "content": "Received my 300 dollar credit almost 4 years ago - spent 200 than and still have 100. The 100 has been there for 4 years - just be certain that you have deleted all storage, etc - check the billing for a couple of months. I have never been 'charged for the 200\" so pretty darn sure it's real value that goggle offered to try their cloud.
", "rawMarkdown": "Received my 300 dollar credit almost 4 years ago - spent 200 than and still have 100. The 100 has been there for 4 years - just be certain that you have deleted all storage, etc - check the billing for a couple of months. I have never been 'charged for the 200\" so pretty darn sure it's real value that goggle offered to try their cloud.", "votes": null }, { "id": "2790587", "postDate": "05/03/2024 08:55:43", "content": "thans bro I appreciate it
", "rawMarkdown": "thans bro I appreciate it", "votes": null } ], "comments": [ { "id": 2781044, "author_name": "hengck23", "author_url": "", "post_date": "04/28/2024 14:22:07", "content": "the trick to cost saving is really:
\nyou don't just blindly start training. for every decision to use the cloud to train, you are pretty confidently that the training should yield an useful/successful experiment
\n(for every failed experiment, it is $$$ wasted. this is actually the same of any project management)
", "votes": null, "replies": [ { "id": 2781622, "author_name": "ravi20076", "author_url": "", "post_date": "04/28/2024 21:31:54", "content": "@hengck23 can you elaborate point 1 please? I did not get you properly!
", "votes": null, "replies": [ { "id": 2785731, "author_name": "faithk7u", "author_url": "", "post_date": "05/01/2024 02:27:18", "content": "I think the point is to make sure that we develop features, and the code can run properly in the local environment before running it on the cloud, which would cost money to debug and develop.
", "votes": null, "replies": [] } ] } ] }, { "id": 2782952, "author_name": "tornikeonoprishvili", "author_url": "", "post_date": "04/29/2024 14:12:01", "content": "Correct me if I'm wrong, but doesn't Google still charge you for the used credits, after they expire? Like, if you use $50 from the initial $300 credits, you'd still get a bill of $50, after a few months.
", "votes": null, "replies": [ { "id": 2783238, "author_name": "kirkdco", "author_url": "", "post_date": "04/29/2024 16:27:55", "content": "Not that I'm aware. I think the longer term bill may come from inadvertently leaving an instance running or leaving data stored, both of which will create surprise bills.
\nOf course, I could be wrong. I've not seen that happen (yet) and I'll double check the details there. Also if anyone else has experience with this, please let us know.
", "votes": null, "replies": [] }, { "id": 2785854, "author_name": "pcjimmmy", "author_url": "", "post_date": "05/01/2024 04:11:49", "content": "Received my 300 dollar credit almost 4 years ago - spent 200 than and still have 100. The 100 has been there for 4 years - just be certain that you have deleted all storage, etc - check the billing for a couple of months. I have never been 'charged for the 200\" so pretty darn sure it's real value that goggle offered to try their cloud.
", "votes": null, "replies": [ { "id": 2790587, "author_name": "akifahiner", "author_url": "", "post_date": "05/03/2024 08:55:43", "content": "thans bro I appreciate it
", "votes": null, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2780927": "Given the size of this dataset, working on Kaggle or a local machine (like my 16 GB laptop) probably won't be feasible. Recently, however, I signed up for Google Cloud and after registering for a paid account, I got $300 of compute credits. You can provision some very beefy machines there for not a lot of cost. Adding in GPUs is where the real costs seem to add up, but I was able to run some large GAN models (for a different project) over a week and only spent about half my credits. You can also transfer your notebooks directly to Google Cloud from Kaggle and they provide a template for data transfer as well. \n\nLooking at machine configurations, something on the order of 96 CPUs and 386 GB of memory runs around $2.50 - $3.50 per hour. 100 hours of free compute may be available assuming you get the introductory offer credits.\n\nQuick update - I'm finding that I'm bumping into quota limits on Google Cloud. I'm able to provision a machine with 24 CPUs, but not more. I have requested an increase in that limit - we'll see if it is approved. 🤞 Regardless, 24 CPUs is way more than I have available locally or on Kaggle. 👍\n\nAnother update - getting quotas increased isn't too difficult as long as the request isn't extravagant. Moderate requests are approved through an automated system within a few minutes.\n\n**Update 1 May 2024**\nAnother possibility is [Saturn Cloud](https://saturncloud.io/plans/saturn_cloud_plans/). They have preconfigured instances with Python or R. I used them in the past and was really impressed with the interface and the tools. I'm not sure how much free compute they offer in the Hosted Free plan, but they give a certain number of free hours compute per month and then pay as you go after that.", "2781044": "the trick to cost saving is really:\n1. use cloud for training and not for development\n2. before training in the cloud, you should have already estimated the end results of training and are pretty confident of the results\n\nyou don't just blindly start training. for every decision to use the cloud to train, you are pretty confidently that the training should yield an useful/successful experiment\n\n(for every failed experiment, it is $$$ wasted. this is actually the same of any project management)", "2781622": "hengck23 can you elaborate point 1 please? I did not get you properly!", "2782952": "Correct me if I'm wrong, but doesn't Google still charge you for the used credits, after they expire? Like, if you use $50 from the initial $300 credits, you'd still get a bill of $50, after a few months.", "2783238": "Not that I'm aware. I think the longer term bill may come from inadvertently leaving an instance running or leaving data stored, both of which will create surprise bills.\n\nOf course, I could be wrong. I've not seen that happen (yet) and I'll double check the details there. Also if anyone else has experience with this, please let us know.", "2785731": "I think the point is to make sure that we develop features, and the code can run properly in the local environment before running it on the cloud, which would cost money to debug and develop.", "2785854": "Received my 300 dollar credit almost 4 years ago - spent 200 than and still have 100. The 100 has been there for 4 years - just be certain that you have deleted all storage, etc - check the billing for a couple of months. I have never been 'charged for the 200\" so pretty darn sure it's real value that goggle offered to try their cloud.", "2790587": "thans bro I appreciate it" }, "source": "meta" }