Standarized scripts

GPDM.py

@@ -73,7 +73,7 @@ def __init__(self,
         self.decay_steps = decay_steps
 
         init_name = 'img' if os.path.exists(self.init) else self.init
-        self.name = f'R-{resize}_S-{pyr_factor}->{coarse_dim}_I-{init_name}+I(0,{noise_sigma})'
+        self.name = f'R-{resize}x{pyr_factor}->{coarse_dim}_I-{init_name}+I(0,{noise_sigma})'
 
     def _get_target_pyramid(self, target_img_path):
         """Reads an image and create a pyraimd out of it. Ordered in increasing image size"""

distribution_metrics/patch_swd.py

@@ -34,13 +34,14 @@ def extract_patches(x, patch_size, stride):
 
 
 class PatchSWDLoss(torch.nn.Module):
-    def __init__(self, patch_size=7, stride=1, num_proj=256, use_convs=True):
+    def __init__(self, patch_size=7, stride=1, num_proj=256, use_convs=True, mask_patches_factor=0):
         super(PatchSWDLoss, self).__init__()
         self.name = f"ConvSWDLoss(p-{patch_size}:{stride})"
         self.patch_size = patch_size
         self.stride = stride
         self.num_proj = num_proj
         self.use_convs = use_convs
+        self.mask_patches_factor = mask_patches_factor
 
     def forward(self, x, y, mask=None):
         b, c, h, w = x.shape
@@ -58,10 +59,10 @@ def forward(self, x, y, mask=None):
 
         if mask is not None:
             # duplicate patches that touches the mask by a factor
-            mask_patches_factor = 0
+
             mask_patches = extract_patches(mask, self.patch_size, self.stride)[0] # in [-1,1]
             mask_patches = torch.any(mask_patches > 0, dim=-1)
-            projy = torch.cat([projy[:, ~mask_patches]] + [projy[:, mask_patches]]*mask_patches_factor, dim=1)
+            projy = torch.cat([projy[:, ~mask_patches]] + [projy[:, mask_patches]]*self.mask_patches_factor, dim=1)
 
         projx, projy = duplicate_to_match_lengths(projx, projy)
 

scripts/duplicate_objects.py

@@ -4,28 +4,30 @@
 
 import distribution_metrics
 from GPDM import GPDM
-from utils import save_image
+from utils import save_image, get_file_name
 
 
-input_and_target_images = [
-    ('images/retargeting/teo.jpg', 'images/edit_inputs/teo_mask.png', 28),
-    ('images/edit_inputs/oranges.jpg', 'images/edit_inputs/oranges_mask.jpg', 28),
-]
-def main():
+def edit_images(input_and_target_images, out_dir):
     """
     Smaller patch size adds variablility but may ruin large objects
     """
-    criteria = distribution_metrics.PatchSWDLoss(patch_size=7, stride=1, num_proj=128)
-    model = GPDM(pyr_factor=0.85, coarse_dim=28, lr=0.05, num_steps=300, init='noise', noise_sigma=1.5, resize=0)
-    output_dir = f'outputs/remove_objects/{criteria.name}_{model.name}'
-    for (target_image_path, input_image_path, coarse_dim) in input_and_target_images:
-        fname, ext = os.path.splitext(os.path.basename(target_image_path))[:2]
 
-        debug_dir = f'{output_dir}/debug_images/{fname}'
+    criteria = distribution_metrics.PatchSWDLoss(patch_size=7, stride=1, num_proj=256, mask_patches_factor=3)
 
-        result = model.run(target_image_path, criteria, debug_dir, mask_img_path=input_image_path)
+    for (target_image_path, mask_image_path, coarse_dim) in input_and_target_images:
+        model = GPDM(coarse_dim=coarse_dim, pyr_factor=0.85, lr=0.03, num_steps=300, init='noise', noise_sigma=0, resize=256)
+
+        result = model.run(target_image_path, criteria, mask_img_path=mask_image_path, debug_dir=None)
+        output_path = os.path.join(out_dir, f'{get_file_name(target_image_path)}.png')
+        save_image(result, output_path)
 
-        save_image(result, f'{output_dir}/generated_images/{fname}{ext}')
 
 if __name__ == '__main__':
-    main()
+    input_and_target_images = [
+        ('images/retargeting/teo.jpg', 'images/edit_inputs/teo_mask.png', 28),
+        ('images/edit_inputs/oranges.jpg', 'images/edit_inputs/oranges_mask.jpg', 28),
+        ('images/edit_inputs/tree.png', 'images/edit_inputs/tree_mask.png', 28),
+        ('images/manipulations/soccer2_org.jpg', 'images/manipulations/soccer2_mask.jpg', 28),
+    ]
+    outputs_dir = f'outputs/duplicate_images'
+    edit_images(input_and_target_images, outputs_dir)

scripts/edit_images.py

@@ -6,31 +6,29 @@
 from GPDM import GPDM
 from utils import save_image, get_file_name
 
-input_and_target_images = [
-    ('images/edit_inputs/tree.png', 'images/edit_inputs/tree_edit.png', 128),
-    ('images/SIGD16/7.jpg', 'images/edit_inputs/balls_edit.jpg', 128),
-    ('images/SIGD16/4.jpg', 'images/edit_inputs/birds_dusk_edit.jpg', 128),
-    ('images/Places50/15.jpg', 'images/edit_inputs/stone_edit.png', 256),
-    ('images/edit_inputs/swiming.jpg', 'images/edit_inputs/swiming_edit.jpg', 128),
-    ('images/retargeting/teo.jpg', 'images/edit_inputs/teo_edit.jpg', 64),
-]
 
-def main():
+def edit_images(input_and_target_images, out_dir):
     """
     Smaller patch size adds variablility but may ruin large objects
     """
     criteria = distribution_metrics.PatchSWDLoss(patch_size=7, stride=1, num_proj=256)
-    # criteria = distribution_metrics.PatchCoherentLoss(patch_size=7, stride=3, mode='detached')
-    outputs_dir = 'outputs/image_editing'
-    for (target_image_path, input_image_path, coarse_dim) in input_and_target_images:
-            model = GPDM(coarse_dim=coarse_dim, pyr_factor=0.85, lr=0.03, num_steps=300, init=input_image_path, noise_sigma=0, resize=0)
-            fname, ext = os.path.splitext(os.path.basename(input_image_path))[:2]
 
-            debug_dir = f'{outputs_dir}/debug_images/{get_file_name(input_image_path)}-to-{get_file_name(target_image_path)}/{criteria.name}_{model.name}'
-            result = model.run(target_image_path, criteria, debug_dir)
+    for (target_image_path, input_image_path, coarse_dim) in input_and_target_images:
+        model = GPDM(coarse_dim=coarse_dim, pyr_factor=0.85, lr=0.03, num_steps=300, init=input_image_path, noise_sigma=0, resize=256)
 
-            save_image(result, f'{outputs_dir}/generated_images/{fname}${0}{ext}')
+        result = model.run(target_image_path, criteria, debug_dir=None)
+        output_path = os.path.join(out_dir, f'{get_file_name(input_image_path)}.png')
+        save_image(result, output_path)
 
 
 if __name__ == '__main__':
-    main()
+    input_and_target_images = [
+        ('images/edit_inputs/tree.png', 'images/edit_inputs/tree_edit.png', 128),
+        ('images/SIGD16/7.jpg', 'images/edit_inputs/balls_edit.jpg', 128),
+        ('images/SIGD16/4.jpg', 'images/edit_inputs/birds_dusk_edit.jpg', 128),
+        ('images/Places50/15.jpg', 'images/edit_inputs/stone_edit.png', 256),
+        ('images/edit_inputs/swiming.jpg', 'images/edit_inputs/swiming_edit.jpg', 32),
+        ('images/retargeting/teo.jpg', 'images/edit_inputs/teo_edit.jpg', 64),
+    ]
+    outputs_dir = f'outputs/edit_images'
+    edit_images(input_and_target_images, outputs_dir)

scripts/resample_periodic_textures.py

@@ -6,27 +6,31 @@
 from GPDM import GPDM
 from utils import save_image
 
-image_paths = [
+
+def resample_textures(image_paths, out_dir, n_reps=1):
+    """
+    Smaller patch size adds variablility but may ruin large objects
+    """
+    criteria = distribution_metrics.PatchSWDLoss(patch_size=7, stride=1, num_proj=256)
+    model = GPDM(pyr_factor=0.85, coarse_dim=35, lr=0.05, num_steps=300, init='noise', noise_sigma=1.5, resize=0)
+
+    for input_image_path in image_paths:
+        for i in range(n_reps):
+
+            result = model.run(input_image_path, criteria, debug_dir=None)
+
+            save_image(result, os.path.join(out_dir, model.name, str(i), os.path.basename(input_image_path)))
+
+if __name__ == '__main__':
+    image_paths = [
         'images/textures/olives.png',
         'images/textures/tomatos.png',
         'images/textures/green_waves.jpg',
         'images/textures/cobbles.jpeg',
         'images/style_transfer/style/brick.jpg',
         'images/style_transfer/style/mondrian.jpg',
         'images/style_transfer/style/rug.jpeg',
-
     ]
 
-def main():
-    criteria = distribution_metrics.PatchSWDLoss(patch_size=7, stride=1, num_proj=128)
-    for input_image_path in image_paths:
-        model = GPDM(pyr_factor=0.75, coarse_dim=32, resize=256, scale_factor=(2, 2), lr=0.03, num_steps=200, init='noise', noise_sigma=1.5, decay_steps=100)
-        output_dir = f'outputs/periodic_textures/{criteria.name}_{model.name}'
-        fname, ext = os.path.splitext(os.path.basename(input_image_path))[:2]
-        debug_dir = f'{output_dir}/debug_images/{fname}'
-        result = model.run(input_image_path, criteria, debug_dir)
-
-        save_image(result, f'{output_dir}/generated_images/{fname}{ext}')
-
-if __name__ == '__main__':
-    main()
+    out_dir = "outputs/resample_textures"
+    resample_textures(image_paths, out_dir)

scripts/reshuffle.py

@@ -6,30 +6,23 @@
 from GPDM import GPDM
 from utils import save_image
 
-image_paths = []
-dataset_dir = 'images/SIGD16'
-# dataset_dir = 'images/Places50'
-# dataset_dir = 'images/reshuffle'
-# dataset_dir = '/home/ariel/university/GPDM/GPDM/images/HQ_16'
-image_paths += [os.path.join(dataset_dir, x) for x in os.listdir(dataset_dir)]
-
-
-def main():
+def reshuffle(dataset_dir, out_dir, n_reps=1):
     """
     Smaller patch size adds variablility but may ruin large objects
     """
-    criteria = distribution_metrics.PatchSWDLoss(patch_size=7, stride=1, num_proj=512)
-    model = GPDM(pyr_factor=0.85, coarse_dim=28, lr=0.05, num_steps=150, init='noise', noise_sigma=1.5, resize=0)
-    output_dir = f'outputs/reshuffle/{os.path.basename(dataset_dir)}_{criteria.name}_{model.name}'
-    for input_image_path in image_paths:
-        for i in range(5):
-            fname, ext = os.path.splitext(os.path.basename(input_image_path))[:2]
+    criteria = distribution_metrics.PatchSWDLoss(patch_size=7, stride=1, num_proj=256)
+    model = GPDM(pyr_factor=0.85, coarse_dim=35, lr=0.05, num_steps=300, init='noise', noise_sigma=1.5, resize=0)
 
-            debug_dir = f'{output_dir}/debug_images/{fname}-{i}'
+    image_paths = [os.path.join(dataset_dir, x) for x in os.listdir(dataset_dir)]
+    for input_image_path in image_paths:
+        for i in range(n_reps):
 
-            result = model.run(input_image_path, criteria, debug_dir)
+            result = model.run(input_image_path, criteria, debug_dir=None)
 
-            save_image(result, f'{output_dir}/generated_images/{fname}${i}{ext}')
+            save_image(result, os.path.join(out_dir, f"{os.path.basename(dataset_dir)}_{model.name}", str(i), os.path.basename(input_image_path)))
 
 if __name__ == '__main__':
-    main()
+    out_dir = "outputs/reshuffle"
+    reshuffle('/home/ariel/university/GPDM/images/SIGD16', out_dir)
+    reshuffle('/home/ariel/university/GPDM/images/Places50', out_dir)
+    reshuffle('/home/ariel/university/GPDM/images/reshuffle', out_dir)

scripts/reshuffle_HD_image.py

@@ -6,27 +6,30 @@
 from GPDM import GPDM
 from utils import save_image
 
-image_paths = []
-dataset_dir = '/home/ariel/university/GPDM/images/HQ_16'
-image_paths += [os.path.join(dataset_dir, x) for x in os.listdir(dataset_dir)]
+criteria = distribution_metrics.PatchSWDLoss(patch_size=11, stride=1, num_proj=64)
+model = GPDM(pyr_factor=0.85, coarse_dim=44, lr=0.03, num_steps=500, init='noise', noise_sigma=0.15, resize=1024,
+             scale_factor=(1, 1))
 
-
-def main():
+def reshuffle_hq(image_paths, out_dir, n_reps=1):
     """
     Smaller patch size adds variablility but may ruin large objects
     """
-    criteria = distribution_metrics.PatchSWDLoss(patch_size=11, stride=1, num_proj=64)
-    model = GPDM(pyr_factor=0.85, coarse_dim=44, lr=0.03, num_steps=500, init='noise', noise_sigma=0.15, resize=1024, scale_factor=(1,1))
-    output_dir = f'outputs/reshuffle/{os.path.basename(dataset_dir)}_{criteria.name}_{model.name}'
-    for input_image_path in image_paths:
-        for i in range(1):
-            fname, ext = os.path.splitext(os.path.basename(input_image_path))[:2]
+    criteria = distribution_metrics.PatchSWDLoss(patch_size=7, stride=1, num_proj=256)
+    model = GPDM(pyr_factor=0.85, coarse_dim=35, lr=0.05, num_steps=300, init='noise', noise_sigma=1.5, resize=0)
 
-            debug_dir = f'{output_dir}/debug_images/{fname}-{i}'
+    for input_image_path in image_paths:
+        for i in range(n_reps):
 
-            result = model.run(input_image_path, criteria, debug_dir)
+            result = model.run(input_image_path, criteria, debug_dir=None)
 
-            save_image(result, f'{output_dir}/generated_images/{fname}${i}{ext}')
+            save_image(result, os.path.join(out_dir, model.name, str(i), os.path.basename(input_image_path)))
 
 if __name__ == '__main__':
-    main()
+
+    image_paths = []
+    dataset_dir = '/home/ariel/university/GPDM/images/HQ_16'
+    image_paths += [os.path.join(dataset_dir, x) for x in os.listdir(dataset_dir)]
+    image_paths = [x for x in image_paths if ('rio' in x or 'lambs' in x or 'safari' in x)]
+
+    out_dir = "outputs/reshuffle_HQ"
+    reshuffle_hq(image_paths, out_dir)

scripts/retarget.py

@@ -6,37 +6,35 @@
 from GPDM import GPDM
 from utils import save_image
 
-images = [
-    'images/retargeting/fish.png',
-    'images/retargeting/fruit.png',
-    'images/retargeting/corn.png',
-    'images/retargeting/pinguins.png',
-    'images/retargeting/SupremeCourt.jpeg',
-    'images/retargeting/mountains.jpg',
-    'images/retargeting/jerusalem.jpg',
-    'images/retargeting/teo.jpg',
-    'images/retargeting/house.jpg',
-    'images/Places50/50.jpg',
-
-]
-images += [os.path.join('images/SIGD16/', x) for x in os.listdir('images/SIGD16/')]
-
-def main():
+
+def retarget(image_paths, out_dir, n_reps=1):
     """
     Smaller patch size adds variablility but may ruin large objects
     """
-    outputs_dir = 'outputs/retarget/'
-    criteria = distribution_metrics.PatchSWDLoss(patch_size=7, stride=1, num_proj=64)
-    for input_image_path in images:
-        for scale_factor in [(2,0.5), (1,1.5), (1,2.5), (1,0.5)]:
-            for i in range(2):
-                model = GPDM(resize=256, coarse_dim=21, pyr_factor=0.85, scale_factor=scale_factor, lr=0.05, num_steps=300, decay_steps=150, init='target', noise_sigma=1.5)
+    criteria = distribution_metrics.PatchSWDLoss(patch_size=7, stride=1, num_proj=256)
+
+    for input_image_path in image_paths:
+        for scale_factor in [(1, 2), (1, 0.5), (2, 0.5)]:
+            for i in range(n_reps):
+                model = GPDM(resize=0, coarse_dim=28, pyr_factor=0.75, scale_factor=scale_factor, lr=0.05, num_steps=300, init='target', noise_sigma=1.5)
+
+                result = model.run(input_image_path, criteria, debug_dir=None)
 
-                fname, ext = os.path.splitext(os.path.basename(input_image_path))[:2]
-                debug_dir = f'{outputs_dir}/{criteria.name}_{model.name}/debug_images/{fname}-{scale_factor}'
+                save_image(result, os.path.join(out_dir, model.name, str(i), str(scale_factor) + os.path.basename(input_image_path)))
 
-                result = model.run(input_image_path, criteria, debug_dir)
 
-                save_image(result, f'{outputs_dir}/{criteria.name}_{model.name}/generated_images/{fname}_{scale_factor}${i}{ext}')
 if __name__ == '__main__':
-    main()
+    image_pats = [
+        'images/retargeting/fish.png',
+        'images/retargeting/fruit.png',
+        'images/retargeting/corn.png',
+        'images/retargeting/pinguins.png',
+        'images/retargeting/SupremeCourt.jpeg',
+        'images/retargeting/mountains.jpg',
+        'images/retargeting/jerusalem.jpg',
+        'images/retargeting/teo.jpg',
+        'images/retargeting/house.jpg',
+        'images/Places50/50.jpg',
+    ]
+    out_dir = "outputs/retarget"
+    retarget(image_pats, out_dir, 1)