I typically train DL models in Google Colab using their GPUs. I can subsequently download the .pkl files and run the trained model locally

Initially I tried to use wheel links in the requirements.txt as noted in this post so I could deploy to Cloud Functions instead of Cloud Run, but this wasn't working

In order to deploy to Cloud Run, we need to download Pytorch in different ways depending on the OS chosen. Here is a breakdown of the requirements needed for each OS. Note you will need to change the requirements.txt file to accommodate these

For local Windows - No Docker image running:

• Install non-CUDA Pytorch for Windows from site

For local Linux no CUDA - Docker image:

• Install non-CUDA Pytorch for Windows from site as the first line in the RUN section of Dockerfile. Also ensure that reference to torch and torchvision are removed from requirements.txt file

If we deploy to a CUDA-enabled virtual machine, there is nothing to change (install Pytorch through normal pip). However this won't be possible until my local machine is supports CUDA.

Trying Alpine version of python for now. UPDATE: Don't use Alpine - good explanation here

To build image (in the root directory of the app):

• docker build -t IMAGE_NAME .

To run a new container

• docker run --name CONTAINER_NAME IMAGE_NAME

To list images:

• docker images

To list containers:

• docker container ls

To stop and remove:

• docker stop CONTAINER_NAME
• docker rm CONTAINER_NAME

It's as easy as:

1. Build

• In local root directory, run: gcloud builds submit --tag gcr.io/PROJECT_NAME/fastai-api

2. Deploy to Cloud Run

• In local root directory, run: gcloud run deploy --image=gcr.io/PROJECT_NAME/fastai-api --platform managed
• Optional - include memory size (size is unit followed by G/M/K): gcloud run deploy --image gcr.io/PROJECT_NAME/fastai-api --memory 512M

Updating Cloud Run

• gcloud run services update fastai-api --memory 512M