Experimental convnext discriminator support

1 files changed, 35 insertions, 0 deletions

diff --git a/models/convnext/discriminator.py b/models/convnext/discriminator.py
new file mode 100644
index 0000000..7dbbe3a
--- /dev/null
+++ b/models/convnext/discriminator.py
@@ -0,0 +1,35 @@
+import torch
+from timm.models import ConvNeXt
+from timm.data.constants import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
+
+from torch.nn import functional as F
+
+
+class ConvNeXtDiscriminator():
+    def __init__(self, model: ConvNeXt, input_size: int) -> None:
+        self.net = model
+
+        self.input_size = input_size
+
+        self.img_mean = torch.tensor(IMAGENET_DEFAULT_MEAN).view(1, -1, 1, 1)
+        self.img_std = torch.tensor(IMAGENET_DEFAULT_STD).view(1, -1, 1, 1)
+
+    def get_score(self, img):
+        img_mean = self.img_mean.to(device=img.device, dtype=img.dtype)
+        img_std = self.img_std.to(device=img.device, dtype=img.dtype)
+
+        img = ((img+1.)/2.).sub(img_mean).div(img_std)
+
+        img = F.interpolate(img, size=(self.input_size, self.input_size), mode='bicubic', align_corners=True)
+        pred = self.net(img)
+        return torch.softmax(pred, dim=-1)[:, 1]
+
+    def get_all(self, img):
+        img_mean = self.img_mean.to(device=img.device, dtype=img.dtype)
+        img_std = self.img_std.to(device=img.device, dtype=img.dtype)
+
+        img = ((img + 1.) / 2.).sub(img_mean).div(img_std)
+
+        img = F.interpolate(img, size=(self.input_size, self.input_size), mode='bicubic', align_corners=True)
+        pred = self.net(img)
+        return pred