SpaCy-based custom prediction on Google AI Platform

1
votes

I'm trying to run a custom prediction routine on Google's AI Platform, but always get an error when I include spaCy as a required package in my setup.py:

gcloud beta ai-platform versions create v1 --model MODEL_NAME --python-version=3.7 --runtime-version=1.15 --package-uris=gs://PATH_TO_PACKAGE --machine-type=mls1-c4-m2 --origin=gs://PATH_TO_MODEL --prediction-class=basic_predictor.BasicPredictor
Using endpoint [https://ml.googleapis.com/]
Creating version (this might take a few minutes)......failed.
ERROR: (gcloud.beta.ai-platform.versions.create) Create Version failed. Bad model detected with error:  "There was a problem processing the user code: basic_predictor.BasicPredictor cannot be found. Please make sure (1) prediction_class is the fully qualified function name, and (2) it uses the correct package name as provided by the package_uris: ['gs://PATH_TO_PACKAGE'] (Error code: 4)"

As soon as I remove spaCy as a dependency, the AI Platform is able to create the version, so it looks like incorrect function names or package names cannot be the problem. Obviously, my model relies on spaCy, so leaving it out is not an option.

Does anyone know how to fix this?

1
google-cloud-platformspacygoogle-ai-platform

1 Answers

0
votes

This seems to be an issue on how the dependencies are being installed on AI Platform prediction nodes. I replicated the issue and got the same error, I also tried to package the library as a tar.gz file but it failed in the same way.

I went ahead and reported this issue in GCP IssueTracker so the AI Platform team can investigate it, you can subscribe to it, to receive notifications whenever there's an update.