metadata
license: mit
Model convert from https://github.com/KichangKim/DeepDanbooru
Usage:
import cv2
import numpy as np
import onnxruntime as rt
from huggingface_hub import hf_hub_download
tagger_model_path = hf_hub_download(repo_id="skytnt/deepdanbooru_onnx", filename="deepdanbooru.onnx")
tagger_model = rt.InferenceSession(tagger_model_path, providers=['CUDAExecutionProvider', 'CPUExecutionProvider'])
tagger_model_meta = tagger_model.get_modelmeta().custom_metadata_map
tagger_tags = eval(tagger_model_meta['tags'])
def tagger_predict(image, score_threshold):
h, w = image.shape[:2]
r = min(512 / w, 512 / h)
h, w = int(h * r), int(w * r)
image = cv2.resize(image, (w, h))
pdx = 512 - w
pdy = 512 - h
img_new = np.full([512, 512, 3], 1, dtype=np.float32)
img_new[pdy // 2:pdy // 2 + h, pdx // 2:pdx // 2 + w] = image
image = img_new[np.newaxis, :]
probs = tagger_model.run(None, {"input_1": image})[0][0]
probs = probs.astype(np.float32)
res = []
for prob, label in zip(probs.tolist(), tagger_tags):
if prob < score_threshold:
continue
res.append(label)
return res
img = cv2.imread("test.jpg")
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img = img.astype(np.float32) / 255
tags = tagger_predict(img, 0.5)
print(tags)