optimization of LN model and simple CNN

RGC_simulations/MEI.py

@@ -285,14 +285,14 @@ def deepdraw(net, base_img, octaves, random_crop=True, original_size=None,
             
             # stopping iterations if mean activation of net declines for generated image
             
-            mean_activation = net(src).mean()
-            if i>1 and mean_activation <= mean_activation_hist:
-                image = image_hist
-                break
-            else:
-                image_hist = image
-                mean_activation_hist = net(src).mean()
-                continue
+            #mean_activation = net(src).mean()
+            #if i>1 and mean_activation <= mean_activation_hist:
+            #    image = image_hist
+            #    break
+            #else:
+            #    image_hist = image
+            #    mean_activation_hist = net(src).mean()
+            #    continue
                 
     # returning the resulting image
     return unprocess(image, mu=bias, sigma=scale)
@@ -344,7 +344,7 @@ def contrast_tuning(model, img, bias, scale, min_contrast=0.01, n=1000, linear=T
     return cont, vals, lim_contrast
 
 
-def MEI_multi_seed(dataset_name, dat, dataloaders, models, n_seeds, MEIParameter, TargetUnit, bk_color, track=False):
+def MEI_multi_seed(dataset_name, dat, dataloaders, models, n_seeds, MEIParameter, TargetUnit, bk_color, init_gen_image, track=False):
     """
     dataset_name        : string    # string with dataset_name of dataset used for training
     dat                 : object    # FileTreeDataset object with data schema of data
@@ -399,7 +399,10 @@ def MEI_multi_seed(dataset_name, dat, dataloaders, models, n_seeds, MEIParameter
     # the background color of the initial image
     background_color = np.float32([bk_color] * channels)
     # generate initial random image
-    gen_image = np.random.normal(background_color, 8, (original_h, original_w, channels))
+    if init_gen_image is not None:
+        gen_image = init_gen_image
+    else:
+        gen_image = np.random.normal(background_color, 8, (original_h, original_w, channels))
     gen_image = np.clip(gen_image, 0, 255)
 
     # generate class visualization via octavewise gradient ascent