Chapter 9 - model.to(device)

[jeffsnell]

Thank you for this class - I am learning a lot!!

I am running all your code on an Intel/Mac Pro (mps GPU) and everything runs well and nearly as fast as Colab in your video. I replace your device agnostic code block with the following:

# Setup device-agnostic code 
import torch
if torch.cuda.is_available():
    device = "cuda" # NVIDIA GPU
elif torch.backends.mps.is_available():
    device = "mps" # Apple GPU
else:
    device = "cpu" # Defaults to CPU if NVIDIA GPU/Apple GPU aren't available

# device = "cpu"
print(f"Using device: {device}")

I find that each time before calling the going_modular - engine.train(...) function - I need to add one line of code to move the model.to(device) or I get an error

I tried adding the model.to(device) code into the engine.py file - just before the model.train() call. But that does not seem to make a difference.

What am I missing here??

[pritesh2000]

model.to(device) used for sending model to target device.
When you are using GPU/ CPU of any type sending model to target device means model will run on GPU/CPU whichever you chose and based on that your model will take time computing results.

ex. GPU will take less time for computing the same results than CPU

Here you might get error, because your training data and model are on different device (GPU/CPU) so you have put them on same device to get computation done.

[jeffsnell]

I understand the purpose of model.to(device) - what I do not understand why when when I put that statement inside the engine.train(...) function it does not seem to work while when II put that statement before the engine.train(...) it does ??

Inside the engine.train(...) function I put model.to(device) just ahead of the model.train() call. But it does not work there...

[pritesh2000]

Can you share your code ? (error)

If you modified train function then share that too