Hey everyone,
We're a small team of six working on a startup project focused on computer vision and algorithms in our free time. Right now, we're using Roboflow for labeling and training models, but at $60 a month, it's not sustainable for us.
We're on the lookout for a more cost-effective solution that doesn't require too much setup time. I've heard about AWS SageMaker, but we're inexperienced with it and unsure if it offers what we need without breaking the bank.
Here's where we stand:
- We've got a small grant of 500 euros to spend, and we're open to using our own money if necessary. We're preparing for a demo soon to gauge interest before deciding on further investment.
- Currently, we store our datasets (about 40GB and growing) in an S3 bucket and use a VPS to host CVAT for labeling, costing us around 4-7 euros a month.
- We're using some basic repositories for data operations and training workflows centered on YOLO, RF-DETR, object detection, segmentation models, and some time series forecasting.
- We're considering renting VMs for training but want a user-friendly way to compare runs, possibly using MLFlow. However, setting up the entire pipeline has proven time-consuming.
What would you all suggest? Is AWS SageMaker a viable option for us, or should we just rent a VM on another platform? What frameworks do you recommend for our workflow, considering our budget limitations of under 500 euros for the next six months?
Thanks for any advice you can provide!
1 Answer
Honestly, I think if $60 a month is too steep for you, AWS SageMaker might not be the right choice. I once racked up a $500 monthly bill when I forgot to delete some resources. You could look into AWS's startup programs for credits, which might help you out, although I believe they require incorporation and incubator involvement, which can be tricky to secure.
And those POC credits are nice but the help usually lasts 2-3 months. After that, you'll likely need more funding again.
That sounds like a lot, especially without steady funding. Is it true that the startup program has those strict requirements? Seems like a hassle to jump through those hoops.