Running Notebooks on Local Windows Machine w/ Multiple GPUs

Course Q&A TensorFlow: Advanced Techniques Specialization Custom and Distributed Training with TF
1

Hello -

In the walk-through of C2_W4_Lab_1_basic-mirrored-strategy.ipynb of the Custom and Distributed Training with TensorFlow course, the instructor is using some IDE & running on his local Windows machine with multiple GPUs to demonstrate the features of the mirrored strategy. As multiple GPUs are unavailable in Coursera or Colab, I’d like to run this lab on my local machine which has multiple NVIDIA GPUs.

It would be most appreciated if someone in the community could point me toward the IDE (?) the instructor is using & the underlying framework used to run the notebook locally (Anaconda?).

Thanks, Matthew

2

He is using Colab, unless you have another screenshot you can share so we can have a look at.

In Colab its possible to use multiple gpu’s but I think you have to purchase a subscription. In your machine I have no idea how to do it.

3

Hi - Appreciate the reply. The instructor specifically said he was running on a Windows machine with a GPU that had only 4 cores (rather than the usual 8 cores) so he had to include the following code to avoid erroring out:
# Note that it generally has a minimum of 8 cores, but if your GPU has
# less, you need to set this. In this case one of my GPUs has 4 cores
os.environ[“TF_MIN_GPU_MULTIPROCESSOR_COUNT”] = “4”
He also noted he was not running in Colab (or Coursera).
Here’s the non-Colab walk-through:

image
image1920×1080 56.7 KB

Many of his walkthroughs do use Colab as seen here:
image
image1920×1080 72.8 KB

Thanks, Matthew

4

I cant see it well here, pycharm maybe…IDLE. Dont use them myself.

5

Hi, As @gent.spah said, the local environment seems to be pycharm, I have just a limited experience with It, used just one time.

But you can have access to multiple GPU’s in Kaggle, since I know you have access to two T4 GPUs for free.

I have a little guide:
https://www.kaggle.com/code/peremartramanonellas/using-multiple-gpu-s-with-tensorflow-on-kaggle

Hope it helps.

2 Likes
6

Hi - Appreciate the insights on Kaggle. I’ll give your link a spin.

I was hoping to run locally on my NVIDIA GPUs so that I wouldn’t be metered. I had an interesting GAN model I wanted to train. On Google Colab, I was metered out & offered the chance to pay for Colab Pro. ;}

I’d like to run in an environment I can manage: local datasets, no metering on model training, etc. I also wanted to benchmark this environment against vanilla Colab & Kaggle.

Looks like I’ll have to continue my search. Conda seems promising but Docker may be the best route. We’ll see.

Thanks again, Matthew