Has anyone successfully deployed the SD Inpainting model from model garden, or the associated Google tutorial notebook?
Click Open Notebook. The notebook states that for "Colab only" there is an extra notebook code block to run that does several "pip3 install". It also states that the section should be skipped if using Workbench.
1) Trying the Colab version:
I set up the default runtime config as so
Machine type: e2-standard-4
GPU type: —
Disk type: 100 GB Standard Disk (pd-standard)
I then run the "Colab only" block, which does various installations and then restarts the kernel.
I follow all of the remaining steps to get the notebook configured, including making sure that there is a google cloud project, that billing, the Vertex AI API and Compute Engine APIs are enabled (they are), creating a cloud storage bucket and service account. Then set the experiments environment variables accordingly.
I then run the remaining blocks and get errors saying that torch was not found when I execute the "Image-inpainting" section. Doing a !pip3 install torch gets me torch, but then when re-running it complains that it's the wrong version, and can't resolve other dependencies. Then the kernel crashes with "unknown error". Clearly there are either missing dependencies or mismatches between dependency versions between the notebook code and the environment.
2) Trying the Vertex Workbench version
For the Vertex option, the only environment clue the notebook gives is to use a "Python 3 GPU notebook with preinstalled Huggingface transformer libraries".
However, when picking an environment to launch the notebook there are no other clues as to which ones have those libraries, or even what a "GPU notebook" is, since most of them allow you to attach a GPU instance. There are quite a few available, and the documentation claims that they should all have the libraries needed for GPU and data science. This seems not to be the case. I tried both the tensorflow and pytorch versions, with attached Nvidia GPU.
As instructed, I skipped the "Colab only" code block on Vertex workbench, but I just got lots of errors that torch, diffusers, transformers and other libraries were not available. Doing pip installs to try to add them helped get further along, but just unearthed more errors. Similar to the Colab experiment, I think there are unresolved dependencies between the notebook and the environment, and probably misalignment of versions for different packages.
Has anyone successfully deployed this tutorial or this model? It seems to me that it should "just work", especially as it is a Google published notebook.
LinkedIn
Twitter
