CLIP-ViT-H-14-laion2B-s32B-b79K (ONNX)

ONNX export of laion/CLIP-ViT-H-14-laion2B-s32B-b79K, produced with optimum-onnx (task=zero-shot-image-classification). Weights are unmodified; only the graph/serialization format changed. See the base model for license (MIT) and details.

Usage

from transformers import CLIPProcessor
import onnxruntime as ort

processor = CLIPProcessor.from_pretrained("brishen/CLIP-ViT-H-14-laion2B-s32B-b79K-onnx")
sess = ort.InferenceSession("model.onnx")  # + model.onnx_data alongside it

inputs = processor(text=["a photo of a cat", "a photo of a dog"], images=image, return_tensors="np", padding=True)
logits_per_image = sess.run(["logits_per_image"], dict(inputs))[0]

Or via optimum.onnxruntime.ORTModelForZeroShotImageClassification.