Skip to content

Latest commit

 

History

History
86 lines (51 loc) · 2.14 KB

ROCm_Instructions.md

File metadata and controls

86 lines (51 loc) · 2.14 KB
Raw

ROCm Instructions

I am forking this repo to give instructions on using DeOldify on AMD hardware via the ROCm platform.

Note: Getting DeOldify to work on AMD cards is difficult and you may need to play arround with the various requirements for it to work.

Instructions:

  1. Install ROCm

  2. Install Pytorch

  3. Install vision

  4. Install fast.ai

  • Clone fastai
  • git checkout fastai
  1. Test DeOldify
  • Clone DeOldify
  • Install prerequisites (tensorboardX and jupyter lab)
  • Start jupyter lab and monitor GPU usage

1. + 2. Install ROCm + Pytorch

I have created a script to install ROCm and Pytorch; however, since its completion, Pytorch released an official pip package you can use. I have not tried this, but it should work. Make sure to use the "ROCm" option in Pytorch.

Official release:

https://pytorch.org/get-started/locally/

My Script:

git clone https://github.com/computerguy2030/pytorch-rocm-amd.git
cd pytorch-rocm-amd
bash amd\_build\_script.sh

3. Install vision

git clone https://github.com/pytorch/vision.git
sudo python3 setup.py install

4. Install fast.ai

I have created a fork of the fast.ai repo and reverted changes to the required v1.0.51 for DeOldify: git clone https://github.com/computerguy2030/fastai1.git

Original:
Clone:

git clone https://github.com/fastai/fastai1.git

Checkout the correct version as advised by DeOldify:

git checkout c6bae03a697df565e9d30877069d41a6b813d98a

Install:

sudo python3 setup.py install

5. Start DeOldify Fun!!

Clone:

git clone https://github.com/jantic/DeOldify.git

Install prerequisites:

pip3 install jupyterlab tensorboardX

Start Jupyter Lab and that you are using GPU not CPU

  • Run rocm-smi to check power draw, temp and other information
  • Use radeontop to check GPU utlization

These tools should allow you to monitor your GPU to ensure the your code in the Jupyter Lab is executing on your AMD GPU.

6. Have fun!!!

jupyter lab

I have found much interesting B+W footage on archive.org and recommend it as a video and image source for your experiments.