Large scale controlnet #260
Conversation
PhilippeMoussalli
requested review from
RobbeSneyders,
NielsRogge and
ChristiaensBert
yes we still need to dissect the current pipeline.py |
| url: | ||
| description: The url of the backend clip retrieval service, defaults to the public service | ||
| type: str | ||
| default: https://knn.laion.ai/knn-service |
There was a problem hiding this comment.
It doesn't seem like the main.py script has a default
There was a problem hiding this comment.
The defaults defined here translate internally to defaults defined in the argparser since kfp always requires a given provided argument if specified and cannot be empty.
parser.add_argument("--url", default="https://knn.laion.ai/knn-service")
The values defined in the argument parser generally take precedence over the default values defined in the main.py file so adding them there can be a bit misleading (e.g. if the user attempts to change them, the default values won't be used).
PhilippeMoussalli
force-pushed
the
large-scale-controlnet
branch
from
19a9422 to
c81400c
Compare
PhilippeMoussalli
force-pushed
the
large-scale-controlnet
branch
from
c81400c to
858701e
Compare
| @@ -1,6 +1,6 @@ | |||
| name: Download images | |||
| description: Component that downloads images based on URLs | |||
| image: ghcr.io/ml6team/download_images:dev | |||
| image: ghcr.io/ml6team/download_images:latest | |||
There was a problem hiding this comment.
| image: ghcr.io/ml6team/download_images:latest | |
| image: ghcr.io/ml6team/download_images:dev |
The images on main should be fixed to dev, which corresponds to the latest main version. latest corresponds to the latest release.
There was a problem hiding this comment.
right forget to revert this back!
|
|
||
| return color_seg | ||
| crop_bytes = io.BytesIO() | ||
| image.save(crop_bytes, format="JPEG") |
There was a problem hiding this comment.
Does this actually save the image to disk?
There was a problem hiding this comment.
No, this just saves it to crop_bytes which is a BytesIO object (in-memory buffer to store the image in binary format)
PR for running the controlnet pipeline end-to-end on KFP. Some observations when doing the pipeline testing: - Tested with @ChristiaensBert VM and it runs really nice and much faster than the public clip service. - I could not test everything end to end locally since the GPU component are difficult to run locally -> switched to KFP to leverage the GPU VMs - I had to rebuild images using the build and tag images in the `scripts` folder. I think we still need to modify the script to enable only building specified components since it currently default to all components in the `components` directory which might take some time to build - The local runner does not seem to do the subset checking yet and we still need to expand the CLI to be able to use the kfp runner (currently not supported). Although the CLI is really nice overall :) - Pipeline runs fine and writes the dataset to the hub but fails at the end since it expects an output manifest. This can be resolved with this [ticket](#221). We should prioritize this. Notes: - Changed the segmentation to output a segmentation image instead of a segmentation array since that's the output expected for controlnet training Things to do: - Estimate how much the job would cost
PR for running the controlnet pipeline end-to-end on KFP.
Some observations when doing the pipeline testing:
scriptsfolder. I think we still need to modify the script to enable only building specified components since it currently default to all components in thecomponentsdirectory which might take some time to buildNotes:
Things to do: