Nice video. Although I think people would appreciate more realistic examples, with tips on optimizing using multiple GPUs, showing that adding more GPUs can save money (reduced training & inference time vs. money spent for multiple GPUs). I think that will justify your GPU options and will attract more people to use lambda cloud. MNIST training is such a poor example for 2xA6000 with 48G VRAM since it doesn't even require any GPU to train MNIST.
Though I agree with the point of example being naive, but it's so easy to get all this up and running, we can experiment on our own. Not everything is their job to demonstrate.
Hi. I'm trying to train my U-Net model to do deconvolution after observing before and after deconvolution of 3D microscopic images in .ims format. their resolutions are (32, 2048, 2048). The A100 of colab pro+ isn't enough to run the model for a batch size 16. not only that, i want to enlarge the dataset by introducing augmentations. So can the gpu instances of Lambda handle this task? please let me know asap. im on a deadline for a college project, and at my wit's end
For anyone experiencing issues, you can get around this by updating the remote version of jupyter notebook and ssh tunneling. Here's a tutorial. ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-ZhVdA2jSCuA.html
you showed using separate devices for separate jobs, but how do you use all the devices for the same job (training by distributing the batches across the devices, and taking the gradients from each of them together after each backpropagation)?
The easiest way is to use our new GPU cloud storage feature. You can then either SCP the data up or use sshfs to set up a bridge between your local drive and the server.