waifu2x: Add a option to use only bicubic in train

tuxzz · Feb 8, 2023 · 9a4d6e2 · 9a4d6e2
1 parent f8618d8
commit 9a4d6e2
Show file tree

Hide file tree

Showing 2 changed files with 12 additions and 1 deletion.
diff --git a/waifu2x/training/dataset.py b/waifu2x/training/dataset.py
@@ -29,6 +29,7 @@
 #    "vision.bicubic_no_antialias",
 )
 INTERPOLATION_NEAREST = "box"
+INTERPOLATION_BICUBIC = "catrom"
 #INTERPOLATION_MODE_WEIGHTS = (1/3, 1/3, 1/6, 1/16, 1/3, 1/12)  # noqa: E226
 INTERPOLATION_MODE_WEIGHTS = (1/3, 1/3, 1/6, 1/16, 1/3)  # noqa: E226
 
@@ -72,6 +73,7 @@ def __call__(self, x, y):
             interpolation = random.choices(INTERPOLATION_MODES, weights=INTERPOLATION_MODE_WEIGHTS, k=1)[0]
         else:
             interpolation = self.interpolation
+
         if self.scale_factor == 2:
             if not self.training:
                 blur = 1 + self.blur_shift / 4
@@ -156,6 +158,7 @@ def __init__(self, input_dir,
                  scale_factor,
                  tile_size, num_samples=None,
                  da_jpeg_p=0, da_scale_p=0, da_chshuf_p=0, da_unsharpmask_p=0, da_grayscale_p=0,
+                 bicubic_only=False,
                  deblur=0, resize_blur_p=0.1,
                  noise_level=-1, style=None,
                  training=True):
@@ -173,7 +176,12 @@ def __init__(self, input_dir,
             else:
                 jpeg_transform = TP.Identity()
             if scale_factor > 1:
+                if bicubic_only:
+                    interpolation = INTERPOLATION_BICUBIC
+                else:
+                    interpolation = None  # random
                 random_downscale_x = RandomDownscaleX(scale_factor=scale_factor,
+                                                      interpolation=interpolation,
                                                       blur_shift=deblur, resize_blur_p=resize_blur_p)
                 random_downscale_x_nearest = RandomDownscaleX(scale_factor=scale_factor,
                                                               interpolation=INTERPOLATION_NEAREST)
@@ -208,7 +216,7 @@ def __init__(self, input_dir,
             ])
         else:
             self.gt_transforms = TS.Identity()
-            interpolation = "catrom"
+            interpolation = INTERPOLATION_BICUBIC
             if scale_factor > 1:
                 downscale_x = RandomDownscaleX(scale_factor=scale_factor,
                                                blur_shift=deblur,

diff --git a/waifu2x/training/trainer.py b/waifu2x/training/trainer.py
@@ -58,6 +58,7 @@ def create_dataloader(self, type):
                 input_dir=path.join(self.args.data_dir, "train"),
                 model_offset=model_offset,
                 scale_factor=scale_factor,
+                bicubic_only=self.args.b4b,
                 style=self.args.style,
                 noise_level=self.args.noise_level,
                 tile_size=self.args.size,
@@ -240,6 +241,8 @@ def register(subparsers, default_parser):
     parser.add_argument("--hard-example", type=str, default="linear",
                         choices=["none", "linear", "top10", "top20"],
                         help="hard example mining for training data sampleing")
+    parser.add_argument("--b4b", action="store_true",
+                        help="use only bicubic downsampling for bicubic downsampling restoration")
 
     parser.set_defaults(
         batch_size=16,