update with new and improved cosine noise scheduler

.gitignore

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

README.md

@@ -2,9 +2,7 @@
 
 ## Denoising Diffusion Probabilistic Model, in Pytorch
 
-Implementation of <a href="https://arxiv.org/abs/2006.11239">Denoising Diffusion Probabilistic Model</a> in Pytorch. It is a new approach to generative modeling that may <a href="https://ajolicoeur.wordpress.com/the-new-contender-to-gans-score-matching-with-langevin-sampling/">have the potential</a> to rival GANs. It uses denoising score matching to estimate the gradient of the data distribution, followed by Langevin sampling to sample from the true distribution.
+Implementation of <a href="https://arxiv.org/abs/2006.11239">Denoising Diffusion Probabilistic Model</a> in Pytorch. It is a new approach to generative modeling that may <a href="https://ajolicoeur.wordpress.com/the-new-contender-to-gans-score-matching-with-langevin-sampling/">have the potential</a> to rival GANs. It uses denoising score matching to estimate the gradient of the data distribution, followed by Langevin sampling to sample from the true distribution. This implementation was transcribed from the official Tensorflow version <a href="https://github.com/hojonathanho/diffusion">here</a>. 
-
-This implementation was transcribed from the official Tensorflow version <a href="https://github.com/hojonathanho/diffusion">here</a> and then modified to use <a href="https://arxiv.org/abs/2201.03545">ConvNext</a> blocks instead of Resnets.
 
 <img src="./sample.png" width="500px"><img>
 
@@ -29,9 +27,8 @@ model = Unet(
 
 diffusion = GaussianDiffusion(
     model,
-    image_size = 128,
     timesteps = 1000,   # number of steps
-    loss_type = 'l1'    # L1 or L2
+    loss_type = 'l1'                  # L1 or L2
 )
 
 training_images = torch.randn(8, 3, 128, 128)
@@ -39,7 +36,7 @@ loss = diffusion(training_images)
 loss.backward()
 # after a lot of training
 
-sampled_images = diffusion.sample(batch_size = 4)
+sampled_images = diffusion.sample(128, batch_size = 4)
 sampled_images.shape # (4, 3, 128, 128)
 ```
 
@@ -55,7 +52,6 @@ model = Unet(
 
 diffusion = GaussianDiffusion(
     model,
-    image_size = 128,
     timesteps = 1000,   # number of steps
     loss_type = 'l1'    # L1 or L2
 ).cuda()
@@ -63,9 +59,10 @@ diffusion = GaussianDiffusion(
 trainer = Trainer(
     diffusion,
     'path/to/your/images',
+    image_size = 128,
     train_batch_size = 32,
     train_lr = 2e-5,
-    train_num_steps = 700000,         # total training steps
+    train_num_steps = 100000,         # 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
@@ -74,39 +71,27 @@ trainer = Trainer(
 trainer.train()
 ```
 
-Samples and model checkpoints will be logged to `./results` periodically
-
 ## Citations
 
 ```bibtex
 @misc{ho2020denoising,
-    title   = {Denoising Diffusion Probabilistic Models},
+    title={Denoising Diffusion Probabilistic Models},
-    author  = {Jonathan Ho and Ajay Jain and Pieter Abbeel},
+    author={Jonathan Ho and Ajay Jain and Pieter Abbeel},
-    year    = {2020},
+    year={2020},
-    eprint  = {2006.11239},
+    eprint={2006.11239},
-    archivePrefix = {arXiv},
+    archivePrefix={arXiv},
-    primaryClass = {cs.LG}
+    primaryClass={cs.LG}
 }
 ```
 
 ```bibtex
-@inproceedings{anonymous2021improved,
+@inproceedings{
-    title   = {Improved Denoising Diffusion Probabilistic Models},
+    anonymous2021improved,
-    author  = {Anonymous},
+    title={Improved Denoising Diffusion Probabilistic Models},
-    booktitle = {Submitted to International Conference on Learning Representations},
+    author={Anonymous},
-    year    = {2021},
+    booktitle={Submitted to International Conference on Learning Representations},
-    url     = {https://openreview.net/forum?id=-NEXDKk8gZ},
+    year={2021},
-    note    = {under review}
+    url={https://openreview.net/forum?id=-NEXDKk8gZ},
-}
+    note={under review}
-```
-
-```bibtex
-@misc{liu2022convnet,
-    title   = {A ConvNet for the 2020s},
-    author  = {Zhuang Liu and Hanzi Mao and Chao-Yuan Wu and Christoph Feichtenhofer and Trevor Darrell and Saining Xie},
-    year    = {2022},
-    eprint  = {2201.03545},
-    archivePrefix = {arXiv},
-    primaryClass = {cs.CV}
 }
 ```

denoising_diffusion_pytorch/denoising_diffusion_pytorch.py

@@ -22,6 +22,12 @@ try:
 except:
     APEX_AVAILABLE = False
 
+# constants
+
+SAVE_AND_SAMPLE_EVERY = 1000
+UPDATE_EMA_EVERY = 10
+EXTS = ['jpg', 'jpeg', 'png']
+
 # helpers functions
 
 def exists(x):
@@ -91,72 +97,69 @@ class SinusoidalPosEmb(nn.Module):
         emb = torch.cat((emb.sin(), emb.cos()), dim=-1)
         return emb
 
-def Upsample(dim):
+class Mish(nn.Module):
-    return nn.ConvTranspose2d(dim, dim, 4, 2, 1)
+    def forward(self, x):
+        return x * torch.tanh(F.softplus(x))
 
-def Downsample(dim):
+class Upsample(nn.Module):
-    return nn.Conv2d(dim, dim, 4, 2, 1)
+    def __init__(self, dim):
-
-class LayerNorm(nn.Module):
-    def __init__(self, dim, eps = 1e-5):
         super().__init__()
-        self.eps = eps
+        self.conv = nn.ConvTranspose2d(dim, dim, 4, 2, 1)
-        self.g = nn.Parameter(torch.ones(1, dim, 1, 1))
-        self.b = nn.Parameter(torch.zeros(1, dim, 1, 1))
 
     def forward(self, x):
-        var = torch.var(x, dim = 1, unbiased = False, keepdim = True)
+        return self.conv(x)
-        mean = torch.mean(x, dim = 1, keepdim = True)
-        return (x - mean) / (var + self.eps).sqrt() * self.g + self.b
 
-class PreNorm(nn.Module):
+class Downsample(nn.Module):
-    def __init__(self, dim, fn):
+    def __init__(self, dim):
+        super().__init__()
+        self.conv = nn.Conv2d(dim, dim, 3, 2, 1)
+
+    def forward(self, x):
+        return self.conv(x)
+
+class Rezero(nn.Module):
+    def __init__(self, fn):
         super().__init__()
         self.fn = fn
-        self.norm = LayerNorm(dim)
+        self.g = nn.Parameter(torch.zeros(1))
 
     def forward(self, x):
-        x = self.norm(x)
+        return self.fn(x) * self.g
-        return self.fn(x)
 
 # building block modules
 
-class ConvNextBlock(nn.Module):
+class Block(nn.Module):
-    """ https://arxiv.org/abs/2201.03545 """
+    def __init__(self, dim, dim_out, groups = 8):
+        super().__init__()
+        self.block = nn.Sequential(
+            nn.Conv2d(dim, dim_out, 3, padding=1),
+            nn.GroupNorm(groups, dim_out),
+            Mish()
+        )
+    def forward(self, x):
+        return self.block(x)
 
-    def __init__(self, dim, dim_out, *, time_emb_dim = None, mult = 2, norm = True):
+class ResnetBlock(nn.Module):
+    def __init__(self, dim, dim_out, *, time_emb_dim, groups = 8):
         super().__init__()
         self.mlp = nn.Sequential(
-            nn.GELU(),
+            Mish(),
-            nn.Linear(time_emb_dim, dim)
+            nn.Linear(time_emb_dim, dim_out)
-        ) if exists(time_emb_dim) else None
-
-        self.ds_conv = nn.Conv2d(dim, dim, 7, padding = 3, groups = dim)
-
-        self.net = nn.Sequential(
-            LayerNorm(dim) if norm else nn.Identity(),
-            nn.Conv2d(dim, dim_out * mult, 3, padding = 1),
-            nn.GELU(),
-            nn.Conv2d(dim_out * mult, dim_out, 3, padding = 1)
         )
 
+        self.block1 = Block(dim, dim_out)
+        self.block2 = Block(dim_out, dim_out)
         self.res_conv = nn.Conv2d(dim, dim_out, 1) if dim != dim_out else nn.Identity()
 
-    def forward(self, x, time_emb = None):
+    def forward(self, x, time_emb):
-        h = self.ds_conv(x)
+        h = self.block1(x)
-
+        h += self.mlp(time_emb)[:, :, None, None]
-        if exists(self.mlp):
+        h = self.block2(h)
-            assert exists(time_emb), 'time emb must be passed in'
-            condition = self.mlp(time_emb)
-            h = h + rearrange(condition, 'b c -> b c 1 1')
-
-        h = self.net(h)
         return h + self.res_conv(x)
 
 class LinearAttention(nn.Module):
     def __init__(self, dim, heads = 4, dim_head = 32):
         super().__init__()
-        self.scale = dim_head ** -0.5
         self.heads = heads
         hidden_dim = dim_head * heads
         self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias = False)
@@ -164,45 +167,28 @@ class LinearAttention(nn.Module):
 
     def forward(self, x):
         b, c, h, w = x.shape
-        qkv = self.to_qkv(x).chunk(3, dim = 1)
+        qkv = self.to_qkv(x)
-        q, k, v = map(lambda t: rearrange(t, 'b (h c) x y -> b h c (x y)', h = self.heads), qkv)
+        q, k, v = rearrange(qkv, 'b (qkv heads c) h w -> qkv b heads c (h w)', heads = self.heads, qkv=3)
-        q = q * self.scale
+        k = k.softmax(dim=-1)
-
+        context = torch.einsum('bhdn,bhen->bhde', k, v)
-        k = k.softmax(dim = -1)
+        out = torch.einsum('bhde,bhdn->bhen', context, q)
-        context = torch.einsum('b h d n, b h e n -> b h d e', k, v)
+        out = rearrange(out, 'b heads c (h w) -> b (heads c) h w', heads=self.heads, h=h, w=w)
-
-        out = torch.einsum('b h d e, b h d n -> b h e n', context, q)
-        out = rearrange(out, 'b h c (x y) -> b (h c) x y', h = self.heads, x = h, y = w)
         return self.to_out(out)
 
 # model
 
 class Unet(nn.Module):
-    def __init__(
+    def __init__(self, dim, out_dim = None, dim_mults=(1, 2, 4, 8), groups = 8):
-        self,
-        dim,
-        out_dim = None,
-        dim_mults=(1, 2, 4, 8),
-        channels = 3,
-        with_time_emb = True
-    ):
         super().__init__()
-        self.channels = channels
+        dims = [3, *map(lambda m: dim * m, dim_mults)]
-
-        dims = [channels, *map(lambda m: dim * m, dim_mults)]
         in_out = list(zip(dims[:-1], dims[1:]))
 
-        if with_time_emb:
+        self.time_pos_emb = SinusoidalPosEmb(dim)
-            time_dim = dim
+        self.mlp = nn.Sequential(
-            self.time_mlp = nn.Sequential(
+            nn.Linear(dim, dim * 4),
-                SinusoidalPosEmb(dim),
+            Mish(),
-                nn.Linear(dim, dim * 4),
+            nn.Linear(dim * 4, dim)
-                nn.GELU(),
+        )
-                nn.Linear(dim * 4, dim)
-            )
-        else:
-            time_dim = None
-            self.time_mlp = None
 
         self.downs = nn.ModuleList([])
         self.ups = nn.ModuleList([])
@@ -212,41 +198,42 @@ class Unet(nn.Module):
             is_last = ind >= (num_resolutions - 1)
 
             self.downs.append(nn.ModuleList([
-                ConvNextBlock(dim_in, dim_out, time_emb_dim = time_dim, norm = ind != 0),
+                ResnetBlock(dim_in, dim_out, time_emb_dim = dim),
-                ConvNextBlock(dim_out, dim_out, time_emb_dim = time_dim),
+                ResnetBlock(dim_out, dim_out, time_emb_dim = dim),
-                Residual(PreNorm(dim_out, LinearAttention(dim_out))),
+                Residual(Rezero(LinearAttention(dim_out))),
                 Downsample(dim_out) if not is_last else nn.Identity()
             ]))
 
         mid_dim = dims[-1]
-        self.mid_block1 = ConvNextBlock(mid_dim, mid_dim, time_emb_dim = time_dim)
+        self.mid_block1 = ResnetBlock(mid_dim, mid_dim, time_emb_dim = dim)
-        self.mid_attn = Residual(PreNorm(mid_dim, LinearAttention(mid_dim)))
+        self.mid_attn = Residual(Rezero(LinearAttention(mid_dim)))
-        self.mid_block2 = ConvNextBlock(mid_dim, mid_dim, time_emb_dim = time_dim)
+        self.mid_block2 = ResnetBlock(mid_dim, mid_dim, time_emb_dim = dim)
 
         for ind, (dim_in, dim_out) in enumerate(reversed(in_out[1:])):
             is_last = ind >= (num_resolutions - 1)
 
             self.ups.append(nn.ModuleList([
-                ConvNextBlock(dim_out * 2, dim_in, time_emb_dim = time_dim),
+                ResnetBlock(dim_out * 2, dim_in, time_emb_dim = dim),
-                ConvNextBlock(dim_in, dim_in, time_emb_dim = time_dim),
+                ResnetBlock(dim_in, dim_in, time_emb_dim = dim),
-                Residual(PreNorm(dim_in, LinearAttention(dim_in))),
+                Residual(Rezero(LinearAttention(dim_in))),
                 Upsample(dim_in) if not is_last else nn.Identity()
             ]))
 
-        out_dim = default(out_dim, channels)
+        out_dim = default(out_dim, 3)
         self.final_conv = nn.Sequential(
-            ConvNextBlock(dim, dim),
+            Block(dim, dim),
             nn.Conv2d(dim, out_dim, 1)
         )
 
     def forward(self, x, time):
-        t = self.time_mlp(time) if exists(self.time_mlp) else None
+        t = self.time_pos_emb(time)
+        t = self.mlp(t)
 
         h = []
 
-        for convnext, convnext2, attn, downsample in self.downs:
+        for resnet, resnet2, attn, downsample in self.downs:
-            x = convnext(x, t)
+            x = resnet(x, t)
-            x = convnext2(x, t)
+            x = resnet2(x, t)
             x = attn(x)
             h.append(x)
             x = downsample(x)
@@ -255,10 +242,10 @@ class Unet(nn.Module):
         x = self.mid_attn(x)
         x = self.mid_block2(x, t)
 
-        for convnext, convnext2, attn, upsample in self.ups:
+        for resnet, resnet2, attn, upsample in self.ups:
             x = torch.cat((x, h.pop()), dim=1)
-            x = convnext(x, t)
+            x = resnet(x, t)
-            x = convnext2(x, t)
+            x = resnet2(x, t)
             x = attn(x)
             x = upsample(x)
 
@@ -289,19 +276,8 @@ def cosine_beta_schedule(timesteps, s = 0.008):
     return np.clip(betas, a_min = 0, a_max = 0.999)
 
 class GaussianDiffusion(nn.Module):
-    def __init__(
+    def __init__(self, denoise_fn, timesteps=1000, loss_type='l1', betas = None):
-        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):
@@ -392,10 +368,8 @@ class GaussianDiffusion(nn.Module):
         return img
 
     @torch.no_grad()
-    def sample(self, batch_size = 16):
+    def sample(self, image_size, batch_size = 16):
-        image_size = self.image_size
+        return self.p_sample_loop((batch_size, 3, image_size, 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):
@@ -438,26 +412,24 @@ class GaussianDiffusion(nn.Module):
         return loss
 
     def forward(self, x, *args, **kwargs):
-        b, c, h, w, device, img_size, = *x.shape, x.device, self.image_size
+        b, *_, device = *x.shape, x.device
-        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)
 
 # dataset classes
 
 class Dataset(data.Dataset):
-    def __init__(self, folder, image_size, exts = ['jpg', 'jpeg', 'png']):
+    def __init__(self, folder, image_size):
         super().__init__()
         self.folder = folder
         self.image_size = image_size
-        self.paths = [p for ext in exts for p in Path(f'{folder}').glob(f'**/*.{ext}')]
+        self.paths = [p for ext in EXTS for p in Path(f'{folder}').glob(f'**/*.{ext}')]
 
         self.transform = transforms.Compose([
             transforms.Resize(image_size),
             transforms.RandomHorizontalFlip(),
             transforms.CenterCrop(image_size),
-            transforms.ToTensor(),
+            transforms.ToTensor()
-            transforms.Lambda(lambda t: (t * 2) - 1)
         ])
 
     def __len__(self):
@@ -483,22 +455,16 @@ class Trainer(object):
         train_num_steps = 100000,
         gradient_accumulate_every = 2,
         fp16 = False,
-        step_start_ema = 2000,
+        step_start_ema = 2000
-        update_ema_every = 10,
-        save_and_sample_every = 1000,
-        results_folder = './results'
     ):
         super().__init__()
         self.model = diffusion_model
         self.ema = EMA(ema_decay)
         self.ema_model = copy.deepcopy(self.model)
-        self.update_ema_every = update_ema_every
-
         self.step_start_ema = step_start_ema
-        self.save_and_sample_every = save_and_sample_every
 
         self.batch_size = train_batch_size
-        self.image_size = diffusion_model.image_size
+        self.image_size = image_size
         self.gradient_accumulate_every = gradient_accumulate_every
         self.train_num_steps = train_num_steps
 
@@ -514,9 +480,6 @@ class Trainer(object):
         if fp16:
             (self.model, self.ema_model), self.opt = amp.initialize([self.model, self.ema_model], self.opt, opt_level='O1')
 
-        self.results_folder = Path(results_folder)
-        self.results_folder.mkdir(exist_ok = True)
-
         self.reset_parameters()
 
     def reset_parameters(self):
@@ -534,10 +497,10 @@ class Trainer(object):
             'model': self.model.state_dict(),
             'ema': self.ema_model.state_dict()
         }
-        torch.save(data, str(self.results_folder / f'model-{milestone}.pt'))
+        torch.save(data, f'./model-{milestone}.pt')
 
     def load(self, milestone):
-        data = torch.load(str(self.results_folder / f'model-{milestone}.pt'))
+        data = torch.load(f'./model-{milestone}.pt')
 
         self.step = data['step']
         self.model.load_state_dict(data['model'])
@@ -556,16 +519,15 @@ class Trainer(object):
             self.opt.step()
             self.opt.zero_grad()
 
-            if self.step % self.update_ema_every == 0:
+            if self.step % UPDATE_EMA_EVERY == 0:
                 self.step_ema()
 
-            if self.step != 0 and self.step % self.save_and_sample_every == 0:
+            if self.step != 0 and self.step % SAVE_AND_SAMPLE_EVERY == 0:
-                milestone = self.step // self.save_and_sample_every
+                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(batch_size=n), batches))
+                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)
-                all_images = (all_images + 1) * 0.5
+                utils.save_image(all_images, f'./sample-{milestone}.png', nrow=6)
-                utils.save_image(all_images, str(self.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 = '0.7.1',
+  version = '0.5.0',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',