++

geometry/firsttest/entropy.py

@@ -0,0 +1,68 @@
+import numpy
+import PIL
+from PIL import Image
+import torch
+import torchvision
+
+
+class FeatureExtractor(torch.nn.Module):
+    def __init__(self):
+        super(FeatureExtractor, self).__init__()
+        tmp = torchvision.models.efficientnet_v2_s(weights="DEFAULT").features
+        del tmp[7]
+        del tmp[6]
+        del tmp[5]
+        self.features = tmp.cuda().half()
+
+    def forward(self, x):
+        x = x.cuda().half() / 255.0
+        x = (x - 0.5) / 0.25
+        return self.features(x)
+
+
+with torch.no_grad():
+    net = FeatureExtractor()
+
+    image = PIL.Image.open("/scratchf/DFC2015/BE_ORTHO_27032011_315130_56865.tif")
+    image = numpy.asarray(image.convert("RGB").copy())
+    image = torch.Tensor(image)[0:9920, 0:9920, :].clone()
+    image = torch.stack([image[:, :, 0], image[:, :, 1], image[:, :, 2]], dim=0)
+    image = image.unsqueeze(0)
+
+    print("extract data")
+    allfeatures = torch.zeros(128, 620, 620)
+    for r in range(10):
+        for c in range(10):
+            x = image[:, :, 992 * r : 992 * (r + 1), 992 * c : 992 * (c + 1)]
+            z = net(x)[0].cpu()
+            allfeatures[:, 62 * r : 62 * (r + 1), 62 * c : 62 * (c + 1)] = z
+
+    allfeatures = torch.nn.functional.avg_pool2d(
+        allfeatures.unsqueeze(0), kernel_size=2, stride=2
+    )
+    allfeatures = allfeatures[0]
+
+    print("extract stats")
+    allfeatures = allfeatures.cuda().half()
+
+    GRAM = torch.nn.functional.softmax(torch.matmul(allfeatures.transpose(0, 1), allfeatures),1)
+
+    del allfeatures
+    torch.diagonal(GRAM).fill_(-10)
+    assert GRAM.shape == (310 * 310, 310 * 310)
+
+    maxGRAM, _ = GRAM.max(1)
+    assert GRAM.shape[0] == 310 * 310
+    del GRAM
+
+    seuil = sorted(list(maxGRAM.cpu().numpy()))
+    seuil = float(seuil[100])
+    maxGRAM = maxGRAM.view(310, 310).cpu()
+    print((maxGRAM < seuil).float().sum())
+
+    image620 = torch.nn.functional.interpolate(image, size=310, mode="bilinear")
+    image620 = image620[0] / 255
+    torchvision.utils.save_image(image620, "build/image.png")
+
+    image620 *= (maxGRAM <= seuil).float().unsqueeze(0)
+    torchvision.utils.save_image(image620, "build/amer.png")