Creating and running a Docker image

This documentation is unstable (see Issue #3). We believe some of these issues exist in BentoML, so this documentation should be used for experimental purpose only.

From BentoML documentation To create a docker image and run it with a container:

# Find the local path of the latest version PyTorchModel saved bundle
saved_path=$(bentoml get PyTorchModel:latest --print-location --quiet)

echo $saved_path # your_path

# Build docker image using saved_path directory as the build context, replace the
# {username} below to your docker hub account name
docker build -t {username}/pytorch_model $saved_path

# Run a container with the docker image built and expose port 5000
docker run -p 5000:5000 {username}/pytorch_model

FAQs

1. I am getting the following error: No matching distribution found for torchvision==0.7.0+cpu

In your requirements.txt located in the saved_path mentioned above, check if the file contains torchvision==0.7.0+cpu. If it does, add the following line above it - --find-links https://download.pytorch.org/whl/torch_stable.html See Issue #3 for a detailed response.

2. I am getting the following error: No matching distribution found for torch==1.6.0+cpu

In your requirements.txt located in the saved_path mentioned above, check if the file contains torch==1.6.0+cpu. If it does, add the following line above it - --find-links https://download.pytorch.org/whl/torch_stable.html

3. I am getting this error when the container runs: ModuleNotFoundError: No module named 'transformers'

Check if your requirements.txt located in the saved_path mentioned above, contains transformers==3.2.0. If not, add it.

4. Error while deployment: FileNotFoundError: [Errno 2] No such file or directory: 'meta.bin'

Seems silly but, what worked for us was adding the meta.bin file to all 3 directories in the saved_path directory. See Issue #3 for a detailed response.

5. ModuleNotFoundError: No module named 'sklearn'

Adding scikit-learn==0.22 to the requirements.txt file in saved_path directory fixes the issue.

As it stands, our model consumes a lot of memory and cannot be deployed with Heroku's free version (see Issue 5).