How to make custom code in python utilize GPU while using Pytorch tensors and matrice functions

Question

I've created a CNN from scratch only using Pytorch tensors and matrix operation functions in the hope of utilizing GPU. To my surprise, the GPU stays 0% utilized and my training doesn't seem to be faster than running on my cpu.

Before Training:

While Training:

I've double checked whether CUDA is available and have installed it already.

Graphics card: Nvidia GEFORCE 2070 SUPER

Processor: Intel i5 10400

Coding Environment: Jupyter Notebook

Cuda & Cudnn Version: 11.0

Pytorch version: 1.6.0

Answer 1

You have to move your model and data to GPU using

model.cuda()
# and 
x = x.cuda()
y = y.cuda()

You seem to be doing this with-in the calls of forward and backwards. To make sure the model is going on to GPU, monitor the GPU usage continually using shell command

watch -n 5 nvidia-smi

Answer 2

I've created a CNN from scratch only using Pytorch tensors and matrix operation functions in the hope of utilizing GPU. To my surprise, the GPU stays 0% utilized and my training doesn't seem to be faster than running on my cpu.

Before Training:

While Training:

I've double checked whether CUDA is available and have installed it already.

Graphics card: Nvidia GEFORCE 2070 SUPER

Processor: Intel i5 10400

Coding Environment: Jupyter Notebook

Cuda & Cudnn Version: 11.0

Pytorch version: 1.6.0