[CLIP] Validation Script

[dsikka]

Question - should this script include some sort of metric generation for zeroshot classification on some dataset? Right now, the script just takes in local samples

Summary

• Validation script to run the CLIP Zeroshot and CLIP Caption generation pipelines
• Both tasks are supported and the default task is Zeroshot classifiation
• The script adds an extra layer where the user can provide text files with paths to images and sample classes. Examples of text files are provided in this PR: src/deepsparse/clip/sample_classes.txt and src/deepsparse/clip/sample_images.txt. Both are used as defaults

Testing

• Tested locally using the following commands. The default sample_classes.txt was used and the sample_images.txt was updated to include the thailand.jpg (an image of elephants) from the yolact samples images.

Caption Generation:

python validation.py --task caption --decoder-model /home/dsikka/deepsparse/tests/deepsparse/pipelines/final_models/clip_text_decoder.onnx --text-model /home/dsikka/deepsparse/tests/deepsparse/pipelines/final_models/clip_text.onnx --visual-model /home/dsikka/deepsparse/tests/deepsparse/pipelines/final_models/clip_visual.onnx

Output:

Class prediction for /home/dsikka/deepsparse/src/deepsparse/yolact/sample_images/thailand.jpg, an adult elephant and a baby elephant 

Zeroshot Classification:

python validation.py --visual-model ~/deepsparse/tests/deepsparse/pipelines/zeroshot_research/visual/model.onnx --text-model ~/deepsparse/tests/deepsparse/pipelines/zeroshot_research/text/model.onnx

Output:

Class prediction for /home/dsikka/deepsparse/src/deepsparse/yolact/sample_images/thailand.jpg, an elephant