Thanks Jeff for this video, After it I see that Colab Pro has its uses, specially if you want to experiment/learn. The good thing I see is that you can get GPU's there with lots of VRAM for a bargain.
Thanks Jeff, didn't know Colab Pro was available in the UK now too. They do say on the info for the service that they do throttle you if you overuse the system compared to other people, so its basically going to boost other users if you've been hitting the service hard.
@@HeatonResearch I suspect it depends on available resources, so maybe all the cryptokiddies going elsewhere gives us more GPU available. I'm hoping they do the A100 soon too :) going to get Colab Pro this month for sure. Want to try and setup a github to Colab workflow.
Thank you for making all these videos Jeff! I would never have been able to know the things I know and be working on the projects I am were it not for your videos
Precisely the video I was looking for , subscribed. I probably can go with non high priced GPU's like rtx series for learning ML and deep learning. I am getting a good deal on Thinkpad E series gen 3 Amd ryzen 7 5800u with 16 gigs ram. Looks like colab is a good Idea to work on.
If you're using the non-Pro version of colab and still want access to P100's or V100's,factory restart your runtime a couple of times till you get those GPU's. (5-6 times is what usually works for me)
@@NamBarn2nd I don’t think it’s working anymore , with the launch of colab pro + , it seems like everyone on the free tier is limited to a Tesla K80 , i factory reset my runtime about 10 times , I could only get a K80 . If you really need access to a P100 you could use a Kaggle notebook , i very reliably get P100’s there
Great video. Thanks! Can you give your opinion on using gcloud and jupyter notebooks vs colab pro? I think you have many more options, but I wonder if the costs are better running a VM in GCloud vs using colab for basic ML apps.
Thank you so much for the video, for someone who is trying to get into ML, but struggling with budget especially with GPU prices going up, I've been trying to wiggle myself into getting a M1 Ipad pro, and running colab to get around the libraries incompability issues. would you reccomend your students to use such setup for your university subjects or stick to the traditional nvidia boarded x64 laptops?
Hey Jeff, I was wondering if you could do a video on creating your own dataset? I've been trying to do a side project of creating synthetic data and although I'm learning a lot, I feel it's going very slowly and I'm doing a lot of banging my head on the wall. Thanks!
I probably should do an advanced CoLab usage video. But yes. And there are ways around it. GDrive DOES cache some, but not perfect. Generally what I do is ONLY use GDrive for stuff I write... checkpoints... pickle... logs etc. Then pull down all of the training to the colab instance local storage. Of course, colab local storage is wiped as soon as you disconnect, but it is an ideal place to put images for GAN training.
@@HeatonResearch actually copying the checkpoints is a great tip. Also I think I am going to use your JavaScript code the next time. Great tip. Thanks!
Great video Jeff! One thing I haven't seen is a break down of what level of privacy is available on Colab. With most paid compute services like AWS, GCC, Azure, you're using storage like S3, which is usually pretty secure. When you're using colab, your storage is in G-drive, which I think is less secure for things like PII. Do you know any good breakdowns of the privacy of colab?
Two months ago, I always scheduled Tesla V100. But now, a very few times, I can schedule Tesla V100, and most of my time, Colab Pro only gives me Tesla P100. Maybe now GPUs are shortage, Colab Pro does not have enough Tesla V100 for all users.
I have the impression that colab has become faster. I use GANs and with the tuk-tuk graphics card (P-100) it takes now 25% less time, ameliorating from 9,2s/it to 6,9 s/it.
I love Google colab. works better for me than Kaggle. Kaggle doesn't save your run but colab does. You can see previous outputs/results of last run on colab but not on Kaggle. My local machine remains the best for me. But, to be honest... It has suffered in my hands. I just pity it. Always running. Sadly, it's just 8gb VRAM. When I'm running anything heavy, I don't do anything else. Please, is there a way to dedicate my GPU to Machine learning activities and let my default Xorg nouveau drivers power the rest of my application on Ubuntu 20.04? I want to be doing other things while training on local GPU.
Great overview, thanks for doing this!
I really wonder why I didn't come across this channel before!!
Excellent as always Jeff!! 👍
Thanks!
@@HeatonResearch the map of India you have used is the wrong one. please correct top north portion.
Thanks Jeff for this video, After it I see that Colab Pro has its uses, specially if you want to experiment/learn. The good thing I see is that you can get GPU's there with lots of VRAM for a bargain.
Really great straightforward and helpful - thank you so much!
Thanks, glad it was helpful! Happy new year!
I was hoping for a straight yes or no lol, but it was a helpful video. Thank you
Yes, it is very much worth it. IMHO
@@HeatonResearch great! Yeah seems like a lot for only ten bucks a month
Thanks Jeff, didn't know Colab Pro was available in the UK now too. They do say on the info for the service that they do throttle you if you overuse the system compared to other people, so its basically going to boost other users if you've been hitting the service hard.
Yes I am sure there is throttling to some degree. Fortunately, I have not hit it yet.
@@HeatonResearch I suspect it depends on available resources, so maybe all the cryptokiddies going elsewhere gives us more GPU available. I'm hoping they do the A100 soon too :) going to get Colab Pro this month for sure. Want to try and setup a github to Colab workflow.
Thank you for making all these videos Jeff! I would never have been able to know the things I know and be working on the projects I am were it not for your videos
Glad to help! Thank you.
Precisely the video I was looking for , subscribed. I probably can go with non high priced GPU's like rtx series for learning ML and deep learning. I am getting a good deal on Thinkpad E series gen 3 Amd ryzen 7 5800u with 16 gigs ram. Looks like colab is a good Idea to work on.
If you're using the non-Pro version of colab and still want access to P100's or V100's,factory restart your runtime a couple of times till you get those GPU's. (5-6 times is what usually works for me)
is it work in 2021 ?
@@NamBarn2nd I don’t think it’s working anymore , with the launch of colab pro + , it seems like everyone on the free tier is limited to a Tesla K80 , i factory reset my runtime about 10 times , I could only get a K80 . If you really need access to a P100 you could use a Kaggle notebook , i very reliably get P100’s there
@@supreethrao4028 i already subscription colab and now i gof p1000 😁
@@NamBarn2nd that’s great ! Maybe because I used colab GPUs a lot in the past week I’m getting nothing but K80’s
it works. Thanks
Thanks for the insight. Please update your map for India @1:54
please do the spot instance video!!
I’d love to see that too, great idea!
Okay, its on the list!
Based on minute 8:10 , for researching do you recommend building our own local machine for StyleGAN ?
Awesome video. I was thinking about doing this but so far all my projects aren't big or long enough to justify the purchase.
Pls make more videos like this. Thanks.
You could use terminal. It’s on bottom left corner.
Great point, and huge advantage to colab pro.
Training GANs in colab is such a game changer...
Great video. Thanks!
Can you give your opinion on using gcloud and jupyter notebooks vs colab pro? I think you have many more options, but I wonder if the costs are better running a VM in GCloud vs using colab for basic ML apps.
Thank you so much for the video, for someone who is trying to get into ML, but struggling with budget especially with GPU prices going up, I've been trying to wiggle myself into getting a M1 Ipad pro, and running colab to get around the libraries incompability issues.
would you reccomend your students to use such setup for your university subjects or stick to the traditional nvidia boarded x64 laptops?
Hey Jeff, I was wondering if you could do a video on creating your own dataset? I've been trying to do a side project of creating synthetic data and although I'm learning a lot, I feel it's going very slowly and I'm doing a lot of banging my head on the wall. Thanks!
Great video Jeff. Does google colab pro also suffer from slow uploads/downloads to the notebook instance, when not using the Google Drive?
I probably should do an advanced CoLab usage video. But yes. And there are ways around it. GDrive DOES cache some, but not perfect. Generally what I do is ONLY use GDrive for stuff I write... checkpoints... pickle... logs etc. Then pull down all of the training to the colab instance local storage. Of course, colab local storage is wiped as soon as you disconnect, but it is an ideal place to put images for GAN training.
@@HeatonResearch actually copying the checkpoints is a great tip. Also I think I am going to use your JavaScript code the next time. Great tip. Thanks!
Great video Jeff!
One thing I haven't seen is a break down of what level of privacy is available on Colab. With most paid compute services like AWS, GCC, Azure, you're using storage like S3, which is usually pretty secure. When you're using colab, your storage is in G-drive, which I think is less secure for things like PII. Do you know any good breakdowns of the privacy of colab?
Could you provide some info about AWS , comparing with Google colab pro?
Thanks a lot for your help.
Two months ago, I always scheduled Tesla V100. But now, a very few times, I can schedule Tesla V100, and most of my time, Colab Pro only gives me Tesla P100. Maybe now GPUs are shortage, Colab Pro does not have enough Tesla V100 for all users.
Can I use Google colab for computer vision like open cv?
Yes, OpenCV is installed in CoLab by default.
awesome
Even as a Colab Pro subscriber, I've been able to run one session at a time. Is that correct?
Hopefully Pro comes to Finland also.
fingers crossed.
Which will more fast for deep learning RTX 3080 laptop GPU or COLAB PRO GPU?
Colab
Sir is there a MS in ML program at Washington University, i didn’t saw on official website
Please make a spot instance video :)
Hello jeff can you help me out in calculating memory size from parameters of model.
colab pro disk space ?
Does colab pro prevent opening gpu after long transactions?
No, they just downgrade you to a slower GPU for awhile.
I remember when a p100 was completely available for the free tier, then they just downgraded us to a T4.
I now have T4 with the pro version. I remember getting P100 with free version, then pro version, now you need tp pay 50$ wtf for the same GPU
Colab pro or colab same keeps on disconnecting whenever it wishes.
Now there is even Colab Pro+ available that costs 50 bucks per month and offers background execution
I upgraded Pro to Pro+ but still got Tesla P100.
@@yifeipei5484 oof I think I'm gonna unsubcribe. I have T4 with Pro version and it's very slow
@@Nesggy Yeah, I agree with you. The quality of Colab Pro declines.
free tier offers 107GB.so
colab pro disk space ?
🤔🤔🤔🤔?
Pretty much the same, 109G:
/content# df -BG
Filesystem 1G-blocks Used Available Use% Mounted on
overlay 148G 39G 109G 27% /
tmpfs 1G 0G 1G 0% /dev
tmpfs 7G 0G 7G 0% /sys/fs/cgroup
shm 6G 0G 6G 0% /dev/shm
tmpfs 7G 1G 7G 1% /var/colab
/dev/sda1 154G 41G 114G 27% /opt/bin
tmpfs 7G 0G 7G 0% /proc/acpi
tmpfs 7G 0G 7G 0% /proc/scsi
tmpfs 7G 0G 7G 0% /sys/firmware
/co
@@HeatonResearch thanks
i need runtime much for colab free whre is the script ?
I have the impression that colab has become faster. I use GANs and with the tuk-tuk graphics card (P-100) it takes now 25% less time, ameliorating from 9,2s/it to 6,9 s/it.
In Russia K80 😔
Hey there are brazillians watching too! Best regards!
I love Google colab. works better for me than Kaggle. Kaggle doesn't save your run but colab does. You can see previous outputs/results of last run on colab but not on Kaggle.
My local machine remains the best for me. But, to be honest... It has suffered in my hands. I just pity it. Always running.
Sadly, it's just 8gb VRAM. When I'm running anything heavy, I don't do anything else.
Please, is there a way to dedicate my GPU to Machine learning activities and let my default Xorg nouveau drivers power the rest of my application on Ubuntu 20.04?
I want to be doing other things while training on local GPU.
I agree, I like the colab interface much better than Kaggle.
Indians love you Jeff 🇮🇳🇮🇳
Much love to India, especially with this COVID crisis.
@@HeatonResearch thank you for your support sir hope this will end soon
i got colab pro and it still gives me 12gb ram !!!
Great info, thanks! Please don't choose to do evil things to the rest of us :)
No, never!
for student its a little expensive😬😬
I get a Tesla K80 for colab PRO, not worth it!
It's limited to about 5 countries. Which is a pretty big con for most people.