SAE: dssim kernel size now depends on resolution

2019-03-12 09:49:40 +04:00
parent fd3b9add2f
commit 46ff33bf89
3 changed files with 12 additions and 8 deletions
@@ -276,7 +276,7 @@ NLayerDiscriminator = nnlib.NLayerDiscriminator
            return func
        nnlib.style_loss = style_loss  
        
-        def dssim(k1=0.01, k2=0.03, max_value=1.0):
+        def dssim(kernel_size=11, k1=0.01, k2=0.03, max_value=1.0):
            # port of tf.image.ssim to pure keras in order to work on plaidML backend.

            def func(y_true, y_pred):
@@ -295,7 +295,7 @@ NLayerDiscriminator = nnlib.NLayerDiscriminator
                    g = K.tile (g, (1,1,ch,1))
                    return g
                              
-                kernel = _fspecial_gauss(11,1.5)
+                kernel = _fspecial_gauss(kernel_size,1.5)

                def reducer(x):
                    return K.depthwise_conv2d(x, kernel, strides=(1, 1), padding='valid')