[CLIP] Captioning Pipeline

[dsikka]

Note: there are currently no models for captioning on sparsezoo.
We have an open issue with open_clip to track some of the issues with CoCa models that have been brought up.

Summary

• Implement CLIPCaptioning and CLIPDecoder pipelines. These pipelines allow us to produce captions given an image. This leverages the previous CLIPVisual and CLIPText Pipelines that were implemented for zeroshot, with some modifications to make them more generic
• The captioning pipeline adds a _generate function which is adapted from open_clip and applies BeamSearch to build the caption: https://github.com/neuralmagic/open_clip/blob/onnx-edit/src/open_clip/coca_model.py
• One caveat is that in open_clip's implementation, the input sequence length is dynamic. We're using padded sequences
• Also, the exported onnx models are all originally from open_clip

Testing

• Added tests to the original clip tests
• Also ran the following script to generate captions for various images:

from deepsparse import BasePipeline, Pipeline
from deepsparse.clip import CLIPCaptionInput, CLIPCaptionPipeline, CLIPVisualInput

root = "caption_models"
model_path_visual = f"{root}/clip_visual.onnx"
model_path_text = f"{root}/clip_text.onnx"
model_path_decoder = f"{root}/clip_text_decoder.onnx"

kwargs = {
    "visual_model_path": model_path_visual,
    "text_model_path": model_path_text,
    "decoder_model_path": model_path_decoder,
}
pipeline = BasePipeline.create(task="clip_caption", **kwargs)

pipeline_input = CLIPCaptionInput(image=CLIPVisualInput(images="mountain.jpg"))
output = pipeline(pipeline_input)

Examples of images and the generated caption:

Caption: a view of mountains in the background .

Caption: an adult elephant and a baby elephant .

Caption: a cup of coffee .

The base branch was changed.