++

geometry/firsttest/smaller.py

@@ -23,10 +23,9 @@ def forward(self, x):
 def computeDist(A):
     A_expanded = A.unsqueeze(1)
     A_diff = A_expanded - A_expanded.permute(0, 2, 1)
-    D = A_diff.pow(2)
+    D = A_diff.pow(2) 
     return D.mean(0)
 
-
 with torch.no_grad():
     net = FeatureExtractor()
 
@@ -51,26 +50,26 @@ def computeDist(A):
 
     print("extract stats")
     allfeatures.cuda().half()
-    # norm = torch.sqrt((allfeatures**2).sum(0).unsqueeze(0)).half()
-    # allfeatures = allfeatures / (norm + 0.0001)
+    norm = torch.sqrt((allfeatures**2).sum(0).unsqueeze(0)).half()
+    meannorm = norm.flatten().mean()
 
     allfeatures = allfeatures.flatten(1)
     assert allfeatures.shape == (128, 310 * 310)
 
-    # GRAM = torch.matmul(torch.transpose(allfeatures, 0, 1), allfeatures)
-    GRAM = computeDist(allfeatures)
+    #GRAM = torch.matmul(torch.transpose(allfeatures, 0, 1), allfeatures)
+    GRAM = computeDist( allfeatures)
     del allfeatures
     torch.diagonal(GRAM).fill_(-1)
     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 = seuil[-100]
     maxGRAM = maxGRAM.view(310, 310)
-    print((maxGRAM > seuil).float().sum())
+    print((maxGRAM>seuil).float().sum())
 
     image620 = torch.nn.functional.interpolate(image, size=310, mode="bilinear")
     image620 = image620[0] / 255