Old changes i commit

super_resolution/models.py

@@ -200,8 +200,10 @@ def __init__(self, ref_image, scale_factor, p=20, s=1, n_proj=64, num_steps=500,
     def loss(self, image):
         from torchvision.transforms import Resize
         self.criteria.init()
-        sr_filters = self.criteria.rand.to(image.device)
-        lr_filters = Resize(self.p//self.scale_factor, antialias=True)(sr_filters.clone())
+        # sr_filters = self.criteria.rand.to(image.device)
+        # lr_filters = Resize(self.p//self.scale_factor, antialias=True)(sr_filters.clone())
+        lr_filters = self.criteria.rand.to(image.device)
+        sr_filters = Resize(self.p*self.scale_factor, antialias=True)(lr_filters.clone())
         projy = F.conv2d(image, sr_filters).transpose(1,0).reshape(self.n_proj, -1)
         projx = F.conv2d(self.ref_image, lr_filters).transpose(1,0).reshape(self.n_proj, -1)
 
@@ -213,6 +215,12 @@ def loss(self, image):
         loss = torch.abs(projx - projy).mean()
         return loss
 
+    def run(self, init_image):
+        img, losses = match_patch_distributions(init_image, self.loss, self.num_steps, self.lr)
+        if self.gradient_projector is not None:
+            img = self.gradient_projector(img)
+        img = torch.clip(img.detach(), -1, 1)
+        return img, losses
 
 class GD_gradient_projector:
     def __init__(self, corrupt_image, operator, n_steps=100, lr=0.0001, reg_weight=0):

super_resolution/scripts/test_projections_histogram.py

@@ -1,8 +1,10 @@
+import os
+import sys
 import torch
 from matplotlib import pyplot as plt
-
-from super_resolution.debug_utils import plot_hists, plot_img
-from super_resolution.predefined_filters import get_random_filters, get_gabor_filters, appply_filter, resize_filters
+sys.path.append(os.path.join(os.path.dirname(__file__), ".."))
+from sr_utils.debug_utils import plot_hists, plot_img
+from sr_utils.predefined_filters import get_random_filters, get_gabor_filters, appply_filter, resize_filters, normalize_filters
 
 
 def split_list_randomly(values):
@@ -13,22 +15,21 @@ def split_list_randomly(values):
     return l1,l2
 
 
-def compare_resized_histograms(p, reverse=True):
-    factor = 4
+def compare_resized_histograms(factor, p, reverse=True):
     low_res = Resize(int(im_size // factor), antialias=True)(hig_res)
 
     if reverse:
         p = factor * p
         factor = 1/factor
 
-    rand_projs = get_random_filters(p, n=50)
+    rand_projs = normalize_filters(get_random_filters(p, n=50))
     # naive_projs = get_naive_kernels(p)
-    gabor_projs = get_gabor_filters(p)
+    gabor_projs = normalize_filters(get_gabor_filters(p))
 
 
-    resized_rand_projs = resize_filters(rand_projs, p, factor)
+    resized_rand_projs = resize_filters(rand_projs, p, factor, normalize=True)
     # resized_naive_projs = resize_filters(naive_projs, p, factor, False)
-    resized_gabor_projs = resize_filters(gabor_projs, p, factor)
+    resized_gabor_projs = resize_filters(gabor_projs, p, factor, normalize=True)
 
     filters = [
                 (rand_projs[0], resized_rand_projs[0], f"Random"),
@@ -60,9 +61,7 @@ def compare_resized_histograms(p, reverse=True):
     plt.show()
 
 
-def hist_sanity():
-    p=5
-    factor=4
+def hist_sanity(factor, p):
     low_res = Resize(int(im_size // factor), antialias=True)(hig_res)
     low_res_big = Resize(im_size, antialias=True)(low_res)
 
@@ -108,9 +107,11 @@ def hist_sanity():
     device = torch.device("cpu")
     im_size = 1024
     nbins = 100
-    hig_res = load_image('../data/images/SR/fox2.jpg').to(device)
+    hig_res = load_image('data/images/SR/fox2.jpg').to(device)
     hig_res = Resize(im_size, antialias=True)(hig_res)
 
-    hist_sanity()
-    compare_resized_histograms(p=5, reverse=True)
-    compare_resized_histograms(p=5, reverse=False)
+    factor = 2
+    p=2
+    hist_sanity(factor, p)
+    compare_resized_histograms(factor, p, reverse=True)
+    compare_resized_histograms(factor, p, reverse=False)

super_resolution/sr_utils/predefined_filters.py

@@ -70,7 +70,7 @@ def get_random_filters(p, n=1, c=3):
     rand = torch.randn(n, c * p ** 2)  # (slice_size**2*ch)
     rand = rand / torch.norm(rand, dim=1, keepdim=True)  # noramlize to unit directions
     rand = rand.reshape(n, c, p, p)
-    rand /= torch.sum(rand, dim=(1,2,3), keepdim=True)
+    # rand /= torch.sum(rand, dim=(1,2,3), keepdim=True)
     return rand
 
 

super_resolution/test.py

@@ -74,8 +74,12 @@ def dump_image(img, path, normalize=True):
 
     # Define models
     models = [
+        # Fixe vs resample:
+        # DirectSWD(refernce_image, p=args.p, s=args.s, mode="Resample", num_steps=args.num_steps, n_proj=args.n_proj),
+        # DirectSWD(refernce_image, p=args.p, s=args.s, mode="Fixed", num_steps=args.num_steps, n_proj=args.n_proj),
+
         # # Patch sizes
-        DirectSWD(refernce_image, p=8, s=1, mode="Resample", num_steps=args.num_steps, n_proj=args.n_proj, lr=100),
+        # DirectSWD(refernce_image, p=8, s=1, mode="Resample", num_steps=args.num_steps, n_proj=args.n_proj, lr=100),
         # DirectSWD(refernce_image, p=16, s=2, mode="Resample", num_steps=args.num_steps, n_proj=args.n_proj),
         # DirectSWD(refernce_image, p=32, s=1, mode="Resample", num_steps=args.num_steps, n_proj=args.n_proj),
         # MSSWD(refernce_image, ps=(3, 5, 7, 9, 11), s=1, mode="Resample", num_steps=args.num_steps, n_proj=args.n_proj, lr=50),
@@ -101,8 +105,8 @@ def dump_image(img, path, normalize=True):
         #           gradient_projector=back_projector(corrupt_image, operator, n_steps=100), name="BackProject"),
 
         # # Self
-        # TwoScalesSWD(corrupt_image, scale_factor=4, p=args.p, s=args.s, mode="Resample", num_steps=args.num_steps, n_proj=args.n_proj, name="self-rescaled"),
-        # DirectSWD(corrupt_image, p=args.p, s=args.s, mode="Resample", num_steps=args.num_steps, n_proj=args.n_proj, name="self"),
+        TwoScalesSWD(corrupt_image, scale_factor=4, p=args.p, s=args.s, mode="Resample", num_steps=args.num_steps, n_proj=args.n_proj, name="self-rescaled"),
+        DirectSWD(corrupt_image, p=args.p, s=args.s, mode="Resample", num_steps=args.num_steps, n_proj=args.n_proj, name="self"),
         # DirectSWD(corrupt_image, p=args.p, s=args.s, mode="Resample", num_steps=args.num_steps, n_proj=args.n_proj, name="self-PGD",
         #           gradient_projector=gradient_projector),
         # DirectSWD(corrupt_image, p=args.p, s=args.s, mode="Resample", num_steps=args.num_steps, n_proj=args.n_proj, name="self-PGDreg",