initial commit

backend/misc/image_resize.py

@@ -0,0 +1,113 @@
+import torch
+import numpy as np
+
+from PIL import Image
+
+
+def bislerp(samples, width, height):
+    def slerp(b1, b2, r):
+        '''slerps batches b1, b2 according to ratio r, batches should be flat e.g. NxC'''
+
+        c = b1.shape[-1]
+
+        # norms
+        b1_norms = torch.norm(b1, dim=-1, keepdim=True)
+        b2_norms = torch.norm(b2, dim=-1, keepdim=True)
+
+        # normalize
+        b1_normalized = b1 / b1_norms
+        b2_normalized = b2 / b2_norms
+
+        # zero when norms are zero
+        b1_normalized[b1_norms.expand(-1, c) == 0.0] = 0.0
+        b2_normalized[b2_norms.expand(-1, c) == 0.0] = 0.0
+
+        # slerp
+        dot = (b1_normalized * b2_normalized).sum(1)
+        omega = torch.acos(dot)
+        so = torch.sin(omega)
+
+        # technically not mathematically correct, but more pleasing?
+        res = (torch.sin((1.0 - r.squeeze(1)) * omega) / so).unsqueeze(1) * b1_normalized + (torch.sin(r.squeeze(1) * omega) / so).unsqueeze(1) * b2_normalized
+        res *= (b1_norms * (1.0 - r) + b2_norms * r).expand(-1, c)
+
+        # edge cases for same or polar opposites
+        res[dot > 1 - 1e-5] = b1[dot > 1 - 1e-5]
+        res[dot < 1e-5 - 1] = (b1 * (1.0 - r) + b2 * r)[dot < 1e-5 - 1]
+        return res
+
+    def generate_bilinear_data(length_old, length_new, device):
+        coords_1 = torch.arange(length_old, dtype=torch.float32, device=device).reshape((1, 1, 1, -1))
+        coords_1 = torch.nn.functional.interpolate(coords_1, size=(1, length_new), mode="bilinear")
+        ratios = coords_1 - coords_1.floor()
+        coords_1 = coords_1.to(torch.int64)
+
+        coords_2 = torch.arange(length_old, dtype=torch.float32, device=device).reshape((1, 1, 1, -1)) + 1
+        coords_2[:, :, :, -1] -= 1
+        coords_2 = torch.nn.functional.interpolate(coords_2, size=(1, length_new), mode="bilinear")
+        coords_2 = coords_2.to(torch.int64)
+        return ratios, coords_1, coords_2
+
+    orig_dtype = samples.dtype
+    samples = samples.float()
+    n, c, h, w = samples.shape
+    h_new, w_new = (height, width)
+
+    # linear w
+    ratios, coords_1, coords_2 = generate_bilinear_data(w, w_new, samples.device)
+    coords_1 = coords_1.expand((n, c, h, -1))
+    coords_2 = coords_2.expand((n, c, h, -1))
+    ratios = ratios.expand((n, 1, h, -1))
+
+    pass_1 = samples.gather(-1, coords_1).movedim(1, -1).reshape((-1, c))
+    pass_2 = samples.gather(-1, coords_2).movedim(1, -1).reshape((-1, c))
+    ratios = ratios.movedim(1, -1).reshape((-1, 1))
+
+    result = slerp(pass_1, pass_2, ratios)
+    result = result.reshape(n, h, w_new, c).movedim(-1, 1)
+
+    # linear h
+    ratios, coords_1, coords_2 = generate_bilinear_data(h, h_new, samples.device)
+    coords_1 = coords_1.reshape((1, 1, -1, 1)).expand((n, c, -1, w_new))
+    coords_2 = coords_2.reshape((1, 1, -1, 1)).expand((n, c, -1, w_new))
+    ratios = ratios.reshape((1, 1, -1, 1)).expand((n, 1, -1, w_new))
+
+    pass_1 = result.gather(-2, coords_1).movedim(1, -1).reshape((-1, c))
+    pass_2 = result.gather(-2, coords_2).movedim(1, -1).reshape((-1, c))
+    ratios = ratios.movedim(1, -1).reshape((-1, 1))
+
+    result = slerp(pass_1, pass_2, ratios)
+    result = result.reshape(n, h_new, w_new, c).movedim(-1, 1)
+    return result.to(orig_dtype)
+
+
+def lanczos(samples, width, height):
+    images = [Image.fromarray(np.clip(255. * image.movedim(0, -1).cpu().numpy(), 0, 255).astype(np.uint8)) for image in samples]
+    images = [image.resize((width, height), resample=Image.Resampling.LANCZOS) for image in images]
+    images = [torch.from_numpy(np.array(image).astype(np.float32) / 255.0).movedim(-1, 0) for image in images]
+    result = torch.stack(images)
+    return result.to(samples.device, samples.dtype)
+
+
+def adaptive_resize(samples, width, height, upscale_method, crop):
+    if crop == "center":
+        old_width = samples.shape[3]
+        old_height = samples.shape[2]
+        old_aspect = old_width / old_height
+        new_aspect = width / height
+        x = 0
+        y = 0
+        if old_aspect > new_aspect:
+            x = round((old_width - old_width * (new_aspect / old_aspect)) / 2)
+        elif old_aspect < new_aspect:
+            y = round((old_height - old_height * (old_aspect / new_aspect)) / 2)
+        s = samples[:, :, y:old_height - y, x:old_width - x]
+    else:
+        s = samples
+
+    if upscale_method == "bislerp":
+        return bislerp(s, width, height)
+    elif upscale_method == "lanczos":
+        return lanczos(s, width, height)
+    else:
+        return torch.nn.functional.interpolate(s, size=(height, width), mode=upscale_method)