coco
46 lines
1.2 KiB
Python
46 lines
1.2 KiB
Python
import onnxruntime as ort
|
|
import torch
|
|
from PIL import Image
|
|
|
|
from onnxruntime.transformers import optimizer
|
|
from torchvision.transforms import Compose, Resize, CenterCrop, ToTensor, Normalize, InterpolationMode
|
|
|
|
model = "clip-image-encoder.onnx"
|
|
|
|
|
|
def test_original():
|
|
ort_session = ort.InferenceSession(model)
|
|
|
|
# 获取模型的输入名称
|
|
input_name = ort_session.get_inputs()[0].name
|
|
|
|
i = Image.open("../../image.jpg")
|
|
|
|
def _convert_image_to_rgb(image: Image):
|
|
return image.convert("RGB")
|
|
|
|
def _transform(n_px):
|
|
return Compose([
|
|
Resize(n_px, interpolation=InterpolationMode.NEAREST),
|
|
CenterCrop(n_px),
|
|
_convert_image_to_rgb,
|
|
ToTensor(),
|
|
Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)),
|
|
])
|
|
|
|
preprocess = _transform(224)
|
|
|
|
image_orig = preprocess(i).unsqueeze(0).to("cpu")
|
|
|
|
def to_numpy(tensor):
|
|
return tensor.detach().cpu().numpy() if tensor.requires_grad else tensor.cpu().numpy()
|
|
|
|
outputs = ort_session.run(None, {input_name: to_numpy(image_orig)})
|
|
|
|
print(outputs)
|
|
return outputs
|
|
|
|
|
|
if __name__ == '__main__':
|
|
test_original()
|