add label

calibration.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f0eac4b0be8eb96b2a6fc7596727954ebe600951b3c4719b456948200d75f58e
+size 1999934

config.json

@@ -19,10 +19,12 @@
     512
   ],
   "id2label": {
-    "0": "LABEL_0"
+    "0": "Watermarked",
+    "1": "No watermark"
   },
   "label2id": {
-    "LABEL_0": 0
+    "Watermarked": 0,
+    "No watermark": 1
   },
   "layer_type": "basic",
   "model_type": "resnet",

export_detector.py

@@ -0,0 +1,63 @@
+import numpy as np
+import torch
+import sys
+from safetensors.torch import save_file
+from PIL import Image
+
+# usage:
+# export_detector.py logits.txt bzhdet/
+
+# read logits file
+data=(np.asarray([float(x) for x in open(sys.argv[1]).readlines()]))
+
+# negate for consistency with "1" = "watermarked"
+data = -data
+
+# sort and convert to safetensors format
+data = np.sort(data)
+data_min = data.min()
+data_max = data.max()
+data_len = data.shape[0]
+Ts = []
+for i in range(1,7):
+    t = 10**-i
+    T = data[int(t * data_len)]
+    print("t for p=%e: %f : %e" % (t, T, (data < T).mean()))
+    Ts.append(T)
+zero = np.where(np.diff(np.sign(data)))[0] + 1
+t = zero / data_len
+T = data[int(t * data_len)]
+print("t for p=%e: %f : %e" % (t, T, (data < T).mean()))
+
+data = torch.tensor(data, dtype=torch.float16)
+save_file({
+    "logits": data
+}, "calibration.safetensors")
+
+# read the detector
+from transformers import AutoModelForImageClassification, BlipImageProcessor
+
+detector_path = sys.argv[2]
+
+image_processor = BlipImageProcessor(do_resize=True,
+                                     size={ "width": 512, "height": 512 },
+                                     do_normalize=True,
+                                     image_mean = [ 0.5, 0.5, 0.5 ],
+                                     image_std = [ 0.5, 0.5, 0.5 ])
+
+detector = AutoModelForImageClassification.from_pretrained(detector_path)
+
+# make it output "1" for "watermarked"
+detector.eval()
+with torch.no_grad():
+    detector.classifier[1].weight.copy_(-detector.classifier[1].weight)
+
+#detector.push_to_hub("imatag/stable-signature-bzh-detector-resnet18")
+#image_processor.push_to_hub("imatag/stable-signature-bzh-detector-resnet18")
+examples = ['examples/not_watermarked.png', 'examples/watermarked.png']
+
+imgs = [ Image.open(example).convert("RGB") for example in examples ]
+inputs = image_processor(imgs, return_tensors="pt")
+with torch.no_grad():
+    p = torch.sigmoid(detector(**inputs).logits)
+print(p)