up

models/vision/ddpm/example.py

 #!/usr/bin/env python3
 import tempfile
+import sys
 
 from diffusers import GaussianDDPMScheduler, UNetModel
 from modeling_ddpm import DDPM
 
+model_id = sys.argv[1]
+folder = sys.argv[2]
+save = bool(int(sys.argv[3]))
 
-unet = UNetModel.from_pretrained("fusing/ddpm_dummy")
-sampler = GaussianDDPMScheduler.from_config("fusing/ddpm_dummy")
+unet = UNetModel.from_pretrained(model_id)
+sampler = GaussianDDPMScheduler.from_config(model_id)
 
 # compose Diffusion Pipeline
-ddpm = DDPM(unet, sampler)
-# generate / sample
+if save:
+    ddpm = DDPM(unet, sampler)
+    ddpm.save_pretrained(folder)
+
 image = ddpm()
-print(image)
 
+import PIL.Image
+import numpy as np
+image_processed = image.cpu().permute(0, 2, 3, 1)
+image_processed = (image_processed + 1.0) * 127.5
+image_processed = image_processed.numpy().astype(np.uint8)
+image_pil = PIL.Image.fromarray(image_processed[0])
+image_pil.save("test.png")
 
-# save and load with 0 extra code (handled by general `DiffusionPipeline` class)
-with tempfile.TemporaryDirectory() as tmpdirname:
-    ddpm.save_pretrained(tmpdirname)
-    print("Model saved")
-    ddpm_new = DDPM.from_pretrained(tmpdirname)
-    print("Model loaded")
-    print(ddpm_new)
+import ipdb; ipdb.set_trace()

models/vision/ddpm/modeling_ddpm.py

@@ -15,12 +15,43 @@
 
 
 from diffusers import DiffusionPipeline
+import tqdm
+import torch
 
 
 class DDPM(DiffusionPipeline):
-    def __init__(self, unet, gaussian_sampler):
-        super().__init__(unet=unet, gaussian_sampler=gaussian_sampler)
+    def __init__(self, unet, noise_scheduler):
+        super().__init__()
+        self.register_modules(unet=unet, noise_scheduler=noise_scheduler)
+
+    def __call__(self, generator=None, torch_device=None):
+        torch_device = "cuda" if torch.cuda.is_available() else "cpu"
+
+        self.unet.to(torch_device)
+        # 1. Sample gaussian noise
+        image = self.noise_scheduler.sample_noise((1, self.unet.in_channels, self.unet.resolution, self.unet.resolution), device=torch_device, generator=generator)
+        for t in tqdm.tqdm(reversed(range(len(self.noise_scheduler))), total=len(self.noise_scheduler)):
+            # i) define coefficients for time step t
+            clip_image_coeff = 1 / torch.sqrt(self.noise_scheduler.get_alpha_prod(t))
+            clip_noise_coeff = torch.sqrt(1 / self.noise_scheduler.get_alpha_prod(t) - 1)
+            image_coeff = (1 - self.noise_scheduler.get_alpha_prod(t - 1)) * torch.sqrt(self.noise_scheduler.get_alpha(t)) / (1 - self.noise_scheduler.get_alpha_prod(t))
+            clip_coeff = torch.sqrt(self.noise_scheduler.get_alpha_prod(t - 1)) * self.noise_scheduler.get_beta(t) / (1 - self.noise_scheduler.get_alpha_prod(t))
+
+            # ii) predict noise residual
+            with torch.no_grad():
+                noise_residual = self.unet(image, t)
+
+            # iii) compute predicted image from residual
+            # See 2nd formula at https://github.com/hojonathanho/diffusion/issues/5#issue-896554416 for comparison
+            pred_mean = clip_image_coeff * image - clip_noise_coeff * noise_residual
+            pred_mean = torch.clamp(pred_mean, -1, 1)
+            prev_image = clip_coeff * pred_mean + image_coeff * image
+
+            # iv) sample variance
+            prev_variance = self.noise_scheduler.sample_variance(t, prev_image.shape, device=torch_device, generator=generator)
+
+            # v) sample  x_{t-1} ~ N(prev_image, prev_variance)
+            sampled_prev_image = prev_image + prev_variance
+            image = sampled_prev_image
 
-    def __call__(self, batch_size=1):
-        image = self.gaussian_sampler.sample(self.unet, batch_size=batch_size)
         return image

src/diffusers/pipeline_utils.py

@@ -45,7 +45,7 @@ class DiffusionPipeline(Config):
 
     config_name = "model_index.json"
 
-    def __init__(self, **kwargs):
+    def register_modules(self, **kwargs):
         for name, module in kwargs.items():
             # retrive library
             library = module.__module__.split(".")[0]