image is in range of 0 to 1

.gitignore

@@ -1,3 +1,6 @@
+# Generation results
+results/
+
 # Byte-compiled / optimized / DLL files
 __pycache__/
 *.py[cod]

README.md

@@ -27,10 +27,9 @@ model = Unet(
 
 diffusion = GaussianDiffusion(
     model,
-    beta_start = 0.0001,
-    beta_end = 0.02,
-    num_diffusion_timesteps = 1000,   # number of steps
-    loss_type = 'l1'                  # L1 or L2 (wavegrad paper claims l1 is better?)
+    image_size = 128,
+    timesteps = 1000,   # number of steps
+    loss_type = 'l1'    # L1 or L2
 )
 
 training_images = torch.randn(8, 3, 128, 128)
@@ -38,7 +37,7 @@ loss = diffusion(training_images)
 loss.backward()
 # after a lot of training
 
-sampled_images = diffusion.sample(128, batch_size = 4)
+sampled_images = diffusion.sample(batch_size = 4)
 sampled_images.shape # (4, 3, 128, 128)
 ```
 
@@ -54,19 +53,17 @@ model = Unet(
 
 diffusion = GaussianDiffusion(
     model,
-    beta_start = 0.0001,
-    beta_end = 0.02,
-    num_diffusion_timesteps = 1000,   # number of steps
-    loss_type = 'l1'                  # L1 or L2
+    image_size = 128,
+    timesteps = 1000,   # number of steps
+    loss_type = 'l1'    # L1 or L2
 ).cuda()
 
 trainer = Trainer(
     diffusion,
     'path/to/your/images',
-    image_size = 128,
     train_batch_size = 32,
     train_lr = 2e-5,
-    train_num_steps = 100000,         # total training steps
+    train_num_steps = 700000,         # total training steps
     gradient_accumulate_every = 2,    # gradient accumulation steps
     ema_decay = 0.995,                # exponential moving average decay
     fp16 = True                       # turn on mixed precision training with apex
@@ -75,15 +72,28 @@ trainer = Trainer(
 trainer.train()
 ```
 
+Samples and model checkpoints will be logged to `./results` periodically
+
 ## Citations
 
 ```bibtex
 @misc{ho2020denoising,
-    title={Denoising Diffusion Probabilistic Models},
-    author={Jonathan Ho and Ajay Jain and Pieter Abbeel},
-    year={2020},
-    eprint={2006.11239},
-    archivePrefix={arXiv},
-    primaryClass={cs.LG}
+    title   = {Denoising Diffusion Probabilistic Models},
+    author  = {Jonathan Ho and Ajay Jain and Pieter Abbeel},
+    year    = {2020},
+    eprint  = {2006.11239},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.LG}
+}
+```
+
+```bibtex
+@inproceedings{anonymous2021improved,
+    title   = {Improved Denoising Diffusion Probabilistic Models},
+    author  = {Anonymous},
+    booktitle = {Submitted to International Conference on Learning Representations},
+    year    = {2021},
+    url     = {https://openreview.net/forum?id=-NEXDKk8gZ},
+    note    = {under review}
 }
 ```

denoising_diffusion_pytorch/denoising_diffusion_pytorch.py

@@ -26,7 +26,10 @@ except:
 
 SAVE_AND_SAMPLE_EVERY = 1000
 UPDATE_EMA_EVERY = 10
-EXTS = ['jpg', 'png']
+EXTS = ['jpg', 'jpeg', 'png']
+
+RESULTS_FOLDER = Path('./results')
+RESULTS_FOLDER.mkdir(exist_ok = True)
 
 # helpers functions
 
@@ -178,9 +181,18 @@ class LinearAttention(nn.Module):
 # model
 
 class Unet(nn.Module):
-    def __init__(self, dim, out_dim = None, dim_mults=(1, 2, 4, 8), groups = 8):
+    def __init__(
+        self,
+        dim,
+        out_dim = None,
+        dim_mults=(1, 2, 4, 8),
+        groups = 8,
+        channels = 3
+    ):
         super().__init__()
-        dims = [3, *map(lambda m: dim * m, dim_mults)]
+        self.channels = channels
+
+        dims = [channels, *map(lambda m: dim * m, dim_mults)]
         in_out = list(zip(dims[:-1], dims[1:]))
 
         self.time_pos_emb = SinusoidalPosEmb(dim)
@@ -219,7 +231,7 @@ class Unet(nn.Module):
                 Upsample(dim_in) if not is_last else nn.Identity()
             ]))
 
-        out_dim = default(out_dim, 3)
+        out_dim = default(out_dim, channels)
         self.final_conv = nn.Sequential(
             Block(dim, dim),
             nn.Conv2d(dim, out_dim, 1)
@@ -263,24 +275,47 @@ def noise_like(shape, device, repeat=False):
     noise = lambda: torch.randn(shape, device=device)
     return repeat_noise() if repeat else noise()
 
+def cosine_beta_schedule(timesteps, s = 0.008):
+    """
+    cosine schedule
+    as proposed in https://openreview.net/forum?id=-NEXDKk8gZ
+    """
+    steps = timesteps + 1
+    x = np.linspace(0, steps, steps)
+    alphas_cumprod = np.cos(((x / steps) + s) / (1 + s) * np.pi * 0.5) ** 2
+    alphas_cumprod = alphas_cumprod / alphas_cumprod[0]
+    betas = 1 - (alphas_cumprod[1:] / alphas_cumprod[:-1])
+    return np.clip(betas, a_min = 0, a_max = 0.999)
+
 class GaussianDiffusion(nn.Module):
-    def __init__(self, denoise_fn, beta_start=0.0001, beta_end=0.02, num_diffusion_timesteps=1000, loss_type='l1', betas = None):
+    def __init__(
+        self,
+        denoise_fn,
+        *,
+        image_size,
+        channels = 3,
+        timesteps = 1000,
+        loss_type = 'l1',
+        betas = None
+    ):
         super().__init__()
+        self.channels = channels
+        self.image_size = image_size
         self.denoise_fn = denoise_fn
 
         if exists(betas):
-            self.np_betas = betas.detach().cpu().numpy() if isinstance(betas, torch.Tensor) else betas
+            betas = betas.detach().cpu().numpy() if isinstance(betas, torch.Tensor) else betas
         else:
-            self.np_betas = betas = np.linspace(beta_start, beta_end, num_diffusion_timesteps).astype(np.float64)
-
-        timesteps, = betas.shape
-        self.num_timesteps = int(timesteps)
-        self.loss_type = loss_type
+            betas = cosine_beta_schedule(timesteps)
 
         alphas = 1. - betas
         alphas_cumprod = np.cumprod(alphas, axis=0)
         alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
 
+        timesteps, = betas.shape
+        self.num_timesteps = int(timesteps)
+        self.loss_type = loss_type
+
         to_torch = partial(torch.tensor, dtype=torch.float32)
 
         self.register_buffer('betas', to_torch(betas))
@@ -330,7 +365,7 @@ class GaussianDiffusion(nn.Module):
         x_recon = self.predict_start_from_noise(x, t=t, noise=self.denoise_fn(x, t))
 
         if clip_denoised:
-            x_recon.clamp_(-1., 1.)
+            x_recon.clamp_(0., 1.)
 
         model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
         return model_mean, posterior_variance, posterior_log_variance
@@ -356,8 +391,10 @@ class GaussianDiffusion(nn.Module):
         return img
 
     @torch.no_grad()
-    def sample(self, image_size, batch_size = 16):
-        return self.p_sample_loop((batch_size, 3, image_size, image_size))
+    def sample(self, batch_size = 16):
+        image_size = self.image_size
+        channels = self.channels
+        return self.p_sample_loop((batch_size, channels, image_size, image_size))
 
     @torch.no_grad()
     def interpolate(self, x1, x2, t = None, lam = 0.5):
@@ -400,7 +437,8 @@ class GaussianDiffusion(nn.Module):
         return loss
 
     def forward(self, x, *args, **kwargs):
-        b, *_, device = *x.shape, x.device
+        b, c, h, w, device, img_size, = *x.shape, x.device, self.image_size
+        assert h == img_size and w == img_size, f'height and width of image must be {img_size}'
         t = torch.randint(0, self.num_timesteps, (b,), device=device).long()
         return self.p_losses(x, t, *args, **kwargs)
 
@@ -452,7 +490,7 @@ class Trainer(object):
         self.step_start_ema = step_start_ema
 
         self.batch_size = train_batch_size
-        self.image_size = image_size
+        self.image_size = diffusion_model.image_size
         self.gradient_accumulate_every = gradient_accumulate_every
         self.train_num_steps = train_num_steps
 
@@ -485,10 +523,10 @@ class Trainer(object):
             'model': self.model.state_dict(),
             'ema': self.ema_model.state_dict()
         }
-        torch.save(data, f'./model-{milestone}.pt')
+        torch.save(data, str(RESULTS_FOLDER / f'model-{milestone}.pt'))
 
     def load(self, milestone):
-        data = torch.load(f'./model-{milestone}.pt')
+        data = torch.load(str(RESULTS_FOLDER / f'model-{milestone}.pt'))
 
         self.step = data['step']
         self.model.load_state_dict(data['model'])
@@ -513,9 +551,9 @@ class Trainer(object):
             if self.step != 0 and self.step % SAVE_AND_SAMPLE_EVERY == 0:
                 milestone = self.step // SAVE_AND_SAMPLE_EVERY
                 batches = num_to_groups(36, self.batch_size)
-                all_images_list = list(map(lambda n: self.ema_model.sample(self.image_size, batch_size=n), batches))
+                all_images_list = list(map(lambda n: self.ema_model.sample(batch_size=n), batches))
                 all_images = torch.cat(all_images_list, dim=0)
-                utils.save_image(all_images, f'./sample-{milestone}.png', nrow=6)
+                utils.save_image(all_images, str(RESULTS_FOLDER / f'sample-{milestone}.png'), nrow=6)
                 self.save(milestone)
 
             self.step += 1

setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'denoising-diffusion-pytorch',
   packages = find_packages(),
-  version = '0.4.0',
+  version = '0.6.2',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',