more skip connections, as in guided diffusion

denoising_diffusion_pytorch/continuous_time_gaussian_diffusion.py

@@ -1,3 +1,4 @@
+import math
 import torch
 from torch import sqrt
 from torch import nn, einsum
@@ -66,7 +67,7 @@ def beta_linear_log_snr(t):
     return -log(expm1(1e-4 + 10 * (t ** 2)))
 
 def alpha_cosine_log_snr(t, s = 0.008):
-    return -log((torch.cos((t + s) / (1 + s) * torch.pi * 0.5) ** -2) - 1, eps = 1e-5)
+    return -log((torch.cos((t + s) / (1 + s) * math.pi * 0.5) ** -2) - 1, eps = 1e-5)
 
 class learned_noise_schedule(nn.Module):
     """ described in section H and then I.2 of the supplementary material for variational ddpm paper """

denoising_diffusion_pytorch/denoising_diffusion_pytorch.py

@@ -19,6 +19,8 @@ from tqdm import tqdm
 from einops import rearrange, reduce
 from einops.layers.torch import Rearrange
 
+from ema_pytorch import EMA
+
 # helpers functions
 
 def exists(x):
@@ -50,21 +52,6 @@ def unnormalize_to_zero_to_one(t):
 
 # small helper modules
 
-class EMA():
-    def __init__(self, beta):
-        super().__init__()
-        self.beta = beta
-
-    def update_model_average(self, ma_model, current_model):
-        for current_params, ma_params in zip(current_model.parameters(), ma_model.parameters()):
-            old_weight, up_weight = ma_params.data, current_params.data
-            ma_params.data = self.update_average(old_weight, up_weight)
-
-    def update_average(self, old, new):
-        if old is None:
-            return new
-        return old * self.beta + (1 - self.beta) * new
-
 class Residual(nn.Module):
     def __init__(self, fn):
         super().__init__()
@@ -73,11 +60,14 @@ class Residual(nn.Module):
     def forward(self, x, *args, **kwargs):
         return self.fn(x, *args, **kwargs) + x
 
-def Upsample(dim):
-    return nn.ConvTranspose2d(dim, dim, 4, 2, 1)
+def Upsample(dim, dim_out = None):
+    return nn.Sequential(
+        nn.Upsample(scale_factor = 2, mode = 'nearest'),
+        nn.Conv2d(dim, default(dim_out, dim), 3, padding = 1)
+    )
 
-def Downsample(dim):
-    return nn.Conv2d(dim, dim, 4, 2, 1)
+def Downsample(dim, dim_out = None):
+    return nn.Conv2d(dim, default(dim_out, dim), 4, 2, 1)
 
 class LayerNorm(nn.Module):
     def __init__(self, dim, eps = 1e-5):
@@ -290,10 +280,10 @@ class Unet(nn.Module):
             is_last = ind >= (num_resolutions - 1)
 
             self.downs.append(nn.ModuleList([
-                block_klass(dim_in, dim_out, time_emb_dim = time_dim),
-                block_klass(dim_out, dim_out, time_emb_dim = time_dim),
-                Residual(PreNorm(dim_out, LinearAttention(dim_out))),
-                Downsample(dim_out) if not is_last else nn.Identity()
+                block_klass(dim_in, dim_in, time_emb_dim = time_dim),
+                block_klass(dim_in, dim_in, time_emb_dim = time_dim),
+                Residual(PreNorm(dim_in, LinearAttention(dim_in))),
+                Downsample(dim_in, dim_out) if not is_last else nn.Conv2d(dim_in, dim_out, 3, padding = 1)
             ]))
 
         mid_dim = dims[-1]
@@ -305,10 +295,10 @@ class Unet(nn.Module):
             is_last = ind == (len(in_out) - 1)
 
             self.ups.append(nn.ModuleList([
-                block_klass(dim_out * 2, dim_in, time_emb_dim = time_dim),
-                block_klass(dim_in, dim_in, time_emb_dim = time_dim),
-                Residual(PreNorm(dim_in, LinearAttention(dim_in))),
-                Upsample(dim_in) if not is_last else nn.Identity()
+                block_klass(dim_out + dim_in, dim_out, time_emb_dim = time_dim),
+                block_klass(dim_out + dim_in, dim_out, time_emb_dim = time_dim),
+                Residual(PreNorm(dim_out, LinearAttention(dim_out))),
+                Upsample(dim_out, dim_in) if not is_last else  nn.Conv2d(dim_out, dim_in, 3, padding = 1)
             ]))
 
         default_out_dim = channels * (1 if not learned_variance else 2)
@@ -327,9 +317,12 @@ class Unet(nn.Module):
 
         for block1, block2, attn, downsample in self.downs:
             x = block1(x, t)
+            h.append(x)
+
             x = block2(x, t)
             x = attn(x)
             h.append(x)
+
             x = downsample(x)
 
         x = self.mid_block1(x, t)
@@ -339,8 +332,11 @@ class Unet(nn.Module):
         for block1, block2, attn, upsample in self.ups:
             x = torch.cat((x, h.pop()), dim = 1)
             x = block1(x, t)
+
+            x = torch.cat((x, h.pop()), dim = 1)
             x = block2(x, t)
             x = attn(x)
+
             x = upsample(x)
 
         x = torch.cat((x, r), dim = 1)
@@ -368,7 +364,7 @@ def cosine_beta_schedule(timesteps, s = 0.008):
     """
     steps = timesteps + 1
     x = torch.linspace(0, timesteps, steps, dtype = torch.float64)
-    alphas_cumprod = torch.cos(((x / timesteps) + s) / (1 + s) * torch.pi * 0.5) ** 2
+    alphas_cumprod = torch.cos(((x / timesteps) + s) / (1 + s) * math.pi * 0.5) ** 2
     alphas_cumprod = alphas_cumprod / alphas_cumprod[0]
     betas = 1 - (alphas_cumprod[1:] / alphas_cumprod[:-1])
     return torch.clip(betas, 0, 0.999)
@@ -597,23 +593,22 @@ class Trainer(object):
         folder,
         *,
         ema_decay = 0.995,
-        image_size = 128,
         train_batch_size = 32,
         train_lr = 1e-4,
         train_num_steps = 100000,
         gradient_accumulate_every = 2,
         amp = False,
         step_start_ema = 2000,
-        update_ema_every = 10,
+        ema_update_every = 10,
         save_and_sample_every = 1000,
         results_folder = './results',
         augment_horizontal_flip = True
     ):
         super().__init__()
+        self.image_size = diffusion_model.image_size
+
         self.model = diffusion_model
-        self.ema = EMA(ema_decay)
-        self.ema_model = copy.deepcopy(self.model)
-        self.update_ema_every = update_ema_every
+        self.ema = EMA(diffusion_model, beta = ema_decay, update_every = ema_update_every)
 
         self.step_start_ema = step_start_ema
         self.save_and_sample_every = save_and_sample_every
@@ -623,9 +618,9 @@ class Trainer(object):
         self.gradient_accumulate_every = gradient_accumulate_every
         self.train_num_steps = train_num_steps
 
-        self.ds = Dataset(folder, image_size, augment_horizontal_flip = augment_horizontal_flip)
+        self.ds = Dataset(folder, self.image_size, augment_horizontal_flip = augment_horizontal_flip)
         self.dl = cycle(data.DataLoader(self.ds, batch_size = train_batch_size, shuffle = True, pin_memory = True, num_workers = cpu_count()))
-        self.opt = Adam(diffusion_model.parameters(), lr=train_lr)
+        self.opt = Adam(diffusion_model.parameters(), lr = train_lr)
 
         self.step = 0
 
@@ -635,22 +630,11 @@ class Trainer(object):
         self.results_folder = Path(results_folder)
         self.results_folder.mkdir(exist_ok = True)
 
-        self.reset_parameters()
-
-    def reset_parameters(self):
-        self.ema_model.load_state_dict(self.model.state_dict())
-
-    def step_ema(self):
-        if self.step < self.step_start_ema:
-            self.reset_parameters()
-            return
-        self.ema.update_model_average(self.ema_model, self.model)
-
     def save(self, milestone):
         data = {
             'step': self.step,
             'model': self.model.state_dict(),
-            'ema': self.ema_model.state_dict(),
+            'ema': self.ema.state_dict(),
             'scaler': self.scaler.state_dict()
         }
         torch.save(data, str(self.results_folder / f'model-{milestone}.pt'))
@@ -660,7 +644,7 @@ class Trainer(object):
 
         self.step = data['step']
         self.model.load_state_dict(data['model'])
-        self.ema_model.load_state_dict(data['ema'])
+        self.ema.load_state_dict(data['ema'])
         self.scaler.load_state_dict(data['scaler'])
 
     def train(self):
@@ -680,15 +664,15 @@ class Trainer(object):
                 self.scaler.update()
                 self.opt.zero_grad()
 
-                if self.step % self.update_ema_every == 0:
-                    self.step_ema()
+                self.ema.update()
 
                 if self.step != 0 and self.step % self.save_and_sample_every == 0:
-                    self.ema_model.eval()
+                    self.ema.ema_model.eval()
+                    with torch.no_grad():
+                        milestone = self.step // self.save_and_sample_every
+                        batches = num_to_groups(36, self.batch_size)
+                        all_images_list = list(map(lambda n: self.ema.ema_model.sample(batch_size=n), batches))
 
-                    milestone = self.step // self.save_and_sample_every
-                    batches = num_to_groups(36, self.batch_size)
-                    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, str(self.results_folder / f'sample-{milestone}.png'), nrow = 6)
                     self.save(milestone)

setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'denoising-diffusion-pytorch',
   packages = find_packages(),
-  version = '0.20.1',
+  version = '0.22.0',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',
@@ -16,6 +16,7 @@ setup(
   ],
   install_requires=[
     'einops',
+    'ema-pytorch',
     'pillow',
     'torch',
     'torchvision',