access Nvidia cloud GPU for FREE - 3 ways for Machine Learning in the cloud 💸

do you think you have to spend the big bucks to start your machine learning journey well i don't think so and let's have a look how you can train even large-scale machine learning models for free my name is jesper ramj and i have a phd in machine learning for geoscience and currently i'm a machine learning engineer in the space industry and yeah i love this stuff let's start with google google has colab and is it free yes it they're k80 gpus but you have to be aware it has a couple of limits you get one gpu or one tpu which is google's fancy machine learning hardware they're really cool i have a couple of notebooks i'll leave them in the comments so you can check them out and they're easy to train but yeah you get one so be aware of that there are limitations to maximum of 12 hours which is fine because you can get model checkpoints for whatever you're doing and save it out into your google drive you can hook that up to google app of course google is making it easy in their own environment to do it but you have to be aware after 12 hours everything gets shut down and you better have something saved otherwise the 12 hours rent for free if you're running close to your fantastic icml paper you might not want to take this route because those resources are allocated on a um as available basis so when a lot of people are in the google cloud which always happens when it's close to a big conference deadline you might not be able to train it on google collab because there are no gpus or tpus available and that's really important to consider if you are doing research how do you do the model persistence let's have a look first of all we need to connect google drive and then if we're training with cameras for example we have to implement model checkpoints and we have to save them to our google drive then next time we can actually start from that model checkpoint now listen to me real close it's important that you do not try to trick google google is known to ban people without recourse do you know the support number for google there isn't any if you get thrown out you are thrown out for good i have tried to scrape scholar on a work computer and that computer was forever not going to go to the scholar website there is no way for you to get back and that would be a shame because it's free it's a free tpu to play with if you're in the us there's a bonus you can get kolepro and there you get your gpu for 24 hours and you sometimes even get better gpus but yeah don't count on it just think that you're getting getting a k80 architecture maybe you you're lucky probably not the next one is cargo notebooks now you can take a kaggle notebook and they can be private your data can be private everything's behind locked gates so you don't have to share it with anyone and you get p100s at the moment and those are some really beefy gpus and you can also once again get tpus because cargo is owned by google there are a couple of tricks that make your life so much easier because you essentially have a contingent that you can use in a week and only turn on the gpu when you actually plan to do training you can do all the coding without gpu or tpu turned on and then you get a tesla p100 and you have 30 hours per week so you can train your model for 30 hours and get something nice out now the cool part about kaggle is also that they do have an api so you can submit data and notebooks from your computer and get everything set up that way and there was a time when cargo was super slow to use on the computer and the website was just super slow but the api was perfectly fine so yeah working on everything offline and then uploading it for the computation perfectly valid and completely free you do not have to be a grandmaster not even an expert you need an account and you get all of that especially when you're using cargo data sets and you're currently not partaking in a competition because that would go out of the budget that you have for calculation on kaggle this is a really good resource for you for free to use now if you want to scale up the entire thing to something more serious like aws the amazon cloud or gcp the google cloud that means you can use a couple of tricks that help you save a lot of money because on aws you get the t1 micro and this t1 micro is free but you can install the deep learning ami on it and that means you have all the deep learning architecture there you can code on there you can ssh into it and once you're ready you upgrade you change your instance that means you have to stop the instance change the instance type to a gpu instance and then you can start your training that means you're not paying for anything until you changed that instance to a for example a p3 instance and then you start your training on that saves you major money because all the development is not done on on the clock of amazon don't get me wrong you can't use them for training those free ones they're tiny they have one virtual cpu no gpu they have i think a gigabyte of ram and yeah don't even try it it's it's gonna cry you're you're gonna make that instance cry and you don't want to make that instant cry it's very time also another tip if you're watching this before a hackathon for example be aware that for some amazon instances you have to apply before so when i wanted to use the p2 instances yes that's how old i am i had to apply for them look at the documentation before and be certain that you don't have to like ask somewhere or submit something like they let everyone in but because they're more expensive um the gpu instances or specialized instances often are behind another wall so you don't accidentally rack up a lot of money here's one for researchers and students especially google has educational programs and i for example won a grant from them for using their tpu and gpu infrastructures there was in effect equivalent to a couple hundred thousand dollars to train models on their architecture and you get that for free you have to apply you have to tell them what your what you want to do but those educational programs from google can be extremely valuable and you should definitely see if you can use them for your masters for your phd or if you're uh early career researcher and don't really have the grand money to shell out the big box i personally like these three options for experimentation it's really easy to use and the notebook interface just makes it super convenient it's nice to try something out just have it run for an hour or two and see what's happening without really worrying about sending it to a queue or something like that so it can be really nice for for hackathons for your personal experimentation it's super neat that it's so easy to switch between cpu gpu and tpu like with one click and you get it switched to the instance that you need and scaling up is usually the next step so you can go into the cloud or if you want to buy some hardware i have a video on that i leave that in the card and you can see if you want to buy a beefier laptop or maybe even a desktop try it out register for google collab set up an instance train a small model and really see what's happening

Channel: Jesper Dramsch – Real-world Machine Learning

Views: 40,851

Keywords: ai, deep learning, machine learning, data science, python, free gpu, gpu gratis, gpu, cloud, tesla, nvidia tesla, google colab, kaggle notebooks, kaggle cloud, kaggle code, p100, k80, neural network training, colab, colab pro, free gpu training, ML for free, machine learning in the cloud, ML in the cloud, broke ML, broke machine learning

Id: Ld8vPbtyJWQ

Length: 8min 32sec (512 seconds)

Published: Thu Jul 29 2021