save samples and models to ./results path

README.md

@@ -6,6 +6,8 @@ Implementation of <a href="https://arxiv.org/abs/2006.11239">Denoising Diffusion
 
 <img src="./sample.png" width="500px"><img>
 
+[![PyPI version](https://badge.fury.io/py/denoising-diffusion-pytorch.svg)](https://badge.fury.io/py/denoising-diffusion-pytorch)
+
 ## Install
 
 ```bash
@@ -25,10 +27,8 @@ model = Unet(
 
 diffusion = GaussianDiffusion(
     model,
-    beta_start = 0.0001,
+    timesteps = 1000,   # number of steps
-    beta_end = 0.02,
+    loss_type = 'l1'    # L1 or L2
-    num_diffusion_timesteps = 1000,   # number of steps
-    loss_type = 'l1'                  # L1 or L2 (wavegrad paper claims l1 is better?)
 )
 
 training_images = torch.randn(8, 3, 128, 128)
@@ -52,10 +52,8 @@ model = Unet(
 
 diffusion = GaussianDiffusion(
     model,
-    beta_start = 0.0001,
+    timesteps = 1000,   # number of steps
-    beta_end = 0.02,
+    loss_type = 'l1'    # L1 or L2
-    num_diffusion_timesteps = 1000,   # number of steps
-    loss_type = 'l1'                  # L1 or L2
 ).cuda()
 
 trainer = Trainer(
@@ -73,8 +71,6 @@ trainer = Trainer(
 trainer.train()
 ```
 
-Todo: Command line tool for one-line training
-
 ## Citations
 
 ```bibtex
@@ -87,3 +83,15 @@ Todo: Command line tool for one-line training
     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
 
@@ -43,6 +46,14 @@ def cycle(dl):
         for data in dl:
             yield data
 
+def num_to_groups(num, divisor):
+    groups = num // divisor
+    remainder = num % divisor
+    arr = [divisor] * groups
+    if remainder > 0:
+        arr.append(remainder)
+    return arr
+
 def loss_backwards(fp16, loss, optimizer, **kwargs):
     if fp16:
         with amp.scale_loss(loss, optimizer) as scaled_loss:
@@ -110,12 +121,13 @@ class Downsample(nn.Module):
         return self.conv(x)
 
 class Rezero(nn.Module):
-    def __init__(self, dim):
+    def __init__(self, fn):
         super().__init__()
+        self.fn = fn
         self.g = nn.Parameter(torch.zeros(1))
 
     def forward(self, x):
-        return x * self.g
+        return self.fn(x) * self.g
 
 # building block modules
 
@@ -149,18 +161,17 @@ class ResnetBlock(nn.Module):
         return h + self.res_conv(x)
 
 class LinearAttention(nn.Module):
-    def __init__(self, dim, heads = 8, dim_head = 32):
+    def __init__(self, dim, heads = 4, dim_head = 32):
         super().__init__()
         self.heads = heads
         hidden_dim = dim_head * heads
-        self.to_qkv = nn.Conv2d(dim, hidden_dim, 1, bias = False)
+        self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias = False)
         self.to_out = nn.Conv2d(hidden_dim, dim, 1)
 
     def forward(self, x):
         b, c, h, w = x.shape
         qkv = self.to_qkv(x)
-        q, k, v = rearrange(qkv, 'b (qkv heads c) h w -> qkv b heads c (h w)', heads = self.heads)
+        q, k, v = rearrange(qkv, 'b (qkv heads c) h w -> qkv b heads c (h w)', heads = self.heads, qkv=3)
-        q = q.softmax(dim=-2)
         k = k.softmax(dim=-1)
         context = torch.einsum('bhdn,bhen->bhde', k, v)
         out = torch.einsum('bhde,bhdn->bhen', context, q)
@@ -255,24 +266,36 @@ 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, timesteps=1000, loss_type='l1', betas = None):
         super().__init__()
         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)
+            betas = cosine_beta_schedule(timesteps)
-
-        timesteps, = betas.shape
-        self.num_timesteps = int(timesteps)
-        self.loss_type = loss_type
 
         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))
@@ -349,7 +372,7 @@ class GaussianDiffusion(nn.Module):
 
     @torch.no_grad()
     def sample(self, image_size, batch_size = 16):
-        return self.p_sample_loop((16, 3, image_size, image_size))
+        return self.p_sample_loop((batch_size, 3, image_size, image_size))
 
     @torch.no_grad()
     def interpolate(self, x1, x2, t = None, lam = 0.5):
@@ -434,13 +457,16 @@ class Trainer(object):
         train_lr = 2e-5,
         train_num_steps = 100000,
         gradient_accumulate_every = 2,
-        fp16 = False
+        fp16 = False,
+        step_start_ema = 2000
     ):
         super().__init__()
         self.model = diffusion_model
         self.ema = EMA(ema_decay)
         self.ema_model = copy.deepcopy(self.model)
+        self.step_start_ema = step_start_ema
 
+        self.batch_size = train_batch_size
         self.image_size = image_size
         self.gradient_accumulate_every = gradient_accumulate_every
         self.train_num_steps = train_num_steps
@@ -463,7 +489,7 @@ class Trainer(object):
         self.ema_model.load_state_dict(self.model.state_dict())
 
     def step_ema(self):
-        if self.step < 2000:
+        if self.step < self.step_start_ema:
             self.reset_parameters()
             return
         self.ema.update_model_average(self.ema_model, self.model)
@@ -474,10 +500,10 @@ 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 / 'model-{milestone}.pt'))
 
         self.step = data['step']
         self.model.load_state_dict(data['model'])
@@ -499,10 +525,12 @@ class Trainer(object):
             if self.step % UPDATE_EMA_EVERY == 0:
                 self.step_ema()
 
-            if self.step % SAVE_AND_SAMPLE_EVERY == 0:
+            if self.step != 0 and self.step % SAVE_AND_SAMPLE_EVERY == 0:
                 milestone = self.step // SAVE_AND_SAMPLE_EVERY
-                all_images = self.ema_model.p_sample_loop((64, 3, self.image_size, self.image_size))
+                batches = num_to_groups(36, self.batch_size)
-                utils.save_image(all_images, f'./sample-{milestone}.png', nrow=8)
+                all_images_list = list(map(lambda n: self.ema_model.sample(self.image_size, batch_size=n), batches))
+                all_images = torch.cat(all_images_list, dim=0)
+                utils.save_image(all_images, str(RESULTS_FOLDER / '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 = '0.2.4',
+  version = '0.5.1',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',