CLIPImageClassifier

CLIPImageClassifier wraps clip image model from transformers.

CLIPImageClassifier is initialized with the argument classes, these are the texts that we want to classify an image to one of them
The executor receives Documents with uri attribute. Each Document’s uri represent the path to an image. The executor will read the image
and classify it to one of the classes.

The result will be saved inside a new tag called class within the original document. The class tag is a dictionary that contains two things:

  • label: the chosen class from classes.
  • score: the confidence score in the chosen class given by the model.

Usage

Use the prebuilt images from Jina Hub in your Python code, add it to your Flow and classify your images according to chosen classes:

from jina import Flow
classes = ['this is a cat','this is a dog','this is a person']
f = Flow().add(
    uses='jinahub+docker://CLIPImageClassifier',
    uses_with={'classes':classes}
    )
docs = DocumentArray()
doc = Document(uri='/your/image/path')
docs.append(doc)

with f:
    f.post(on='/classify', inputs=docs, on_done=lambda resp: print(resp.docs[0].tags['class']['label']))

Returns

Document with class tag. This class tag which is a dict.It contains label which is an str and a float confidence score for the image.

GPU Usage

This executor also offers a GPU version. To use it, make sure to pass 'device'='cuda', as the initialization parameter, and gpus='all' when adding the containerized Executor to the Flow. See the Executor on GPU section of Jina documentation for more details.

Here’s how you would modify the example above to use a GPU:

from jina import Flow

classes = ['this is a cat','this is a dog','this is a person']	
f = Flow().add(
    uses='jinahub+docker://CLIPImageClassifier',
    uses_with={
    'classes':classes,
    'device':'cuda',
    'gpus':'all'
    }
    )
docs = DocumentArray()
doc = Document(uri='/your/image/path')
docs.append(doc)

with f:
    f.post(on='/classify', inputs=docs, on_done=lambda resp: print(resp.docs[0].tags['class']['label']))

Reference

CLIP Image model

GitHub

View Github