Image classification for the masses
Install via pip: pip install zeroshot
For GPU support, pip install zeroshot[torch]
N.B. In theory ONNX supports GPU, but the restrictions on CUDA version are iffy at best, and so for easiest results just use PyTorch. If you're brave, instead pip install onnxruntime-gpu.
First, go to app.usezeroshot.com and create a classifier. Check out the video on the landing page for an example.
Then, in Python (image should be an RGB numpy array with channels last):
import zeroshot
# Create the classifier and preprocessing function.
classifier = zeroshot.Classifier("model-uuid-goes-here")
preprocess_fn = zeroshot.create_preprocess_fn()
# Run the model!
prediction = classifier.predict(preprocess_fn(image))
print(f"The image is class {prediction}")You can also download the classifier and save it somewhere locally so you don't need to hit the server each time. Hit "download model" in the web-app and save the json file somewhere. You can then instead do:
classifier = zeroshot.Classifier("/home/user/path/to/model.json")- To use a GPU, install the torch backend with
pip install zeroshot[torch] - If you are hitting issues with torch trying to run on CPU, try disabling XFormers by setting XFORMERS_DISABLED=1 in your ENV varaibles.
See the docs folder for some details on how things work under the hood.
If you need help or just want to chat, join the Moonshine Labs Slack server and come hang out in the #zeroshot channel.