use full attention at the center of the unet

README.md

@@ -2,7 +2,9 @@
 
 ## 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. This implementation was transcribed from the official Tensorflow version <a href="https://github.com/hojonathanho/diffusion">here</a>. 
+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> 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>
 
@@ -97,3 +99,14 @@ Samples and model checkpoints will be logged to `./results` periodically
     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

@@ -22,15 +22,6 @@ try:
 except:
     APEX_AVAILABLE = False
 
-# constants
-
-SAVE_AND_SAMPLE_EVERY = 1000
-UPDATE_EMA_EVERY = 10
-EXTS = ['jpg', 'jpeg', 'png']
-
-RESULTS_FOLDER = Path('./results')
-RESULTS_FOLDER.mkdir(exist_ok = True)
-
 # helpers functions
 
 def exists(x):
@@ -100,69 +91,72 @@ class SinusoidalPosEmb(nn.Module):
         emb = torch.cat((emb.sin(), emb.cos()), dim=-1)
         return emb
 
-class Mish(nn.Module):
-    def forward(self, x):
-        return x * torch.tanh(F.softplus(x))
+def Upsample(dim):
+    return nn.ConvTranspose2d(dim, dim, 4, 2, 1)
 
-class Upsample(nn.Module):
-    def __init__(self, dim):
+def Downsample(dim):
+    return nn.Conv2d(dim, dim, 4, 2, 1)
+
+class LayerNorm(nn.Module):
+    def __init__(self, dim, eps = 1e-5):
         super().__init__()
-        self.conv = nn.ConvTranspose2d(dim, dim, 4, 2, 1)
+        self.eps = eps
+        self.g = nn.Parameter(torch.ones(1, dim, 1, 1))
+        self.b = nn.Parameter(torch.zeros(1, dim, 1, 1))
 
     def forward(self, x):
-        return self.conv(x)
+        var = torch.var(x, dim = 1, unbiased = False, keepdim = True)
+        mean = torch.mean(x, dim = 1, keepdim = True)
+        return (x - mean) / (var + self.eps).sqrt() * self.g + self.b
 
-class Downsample(nn.Module):
-    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):
+class PreNorm(nn.Module):
+    def __init__(self, dim, fn):
         super().__init__()
         self.fn = fn
-        self.g = nn.Parameter(torch.zeros(1))
+        self.norm = LayerNorm(dim)
 
     def forward(self, x):
-        return self.fn(x) * self.g
+        x = self.norm(x)
+        return self.fn(x)
 
 # building block modules
 
-class Block(nn.Module):
-    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)
+class ConvNextBlock(nn.Module):
+    """ https://arxiv.org/abs/2201.03545 """
 
-class ResnetBlock(nn.Module):
-    def __init__(self, dim, dim_out, *, time_emb_dim, groups = 8):
+    def __init__(self, dim, dim_out, *, time_emb_dim = None, mult = 2, norm = True):
         super().__init__()
         self.mlp = nn.Sequential(
-            Mish(),
-            nn.Linear(time_emb_dim, dim_out)
+            nn.GELU(),
+            nn.Linear(time_emb_dim, dim)
+        ) 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):
-        h = self.block1(x)
-        h += self.mlp(time_emb)[:, :, None, None]
-        h = self.block2(h)
+    def forward(self, x, time_emb = None):
+        h = self.ds_conv(x)
+
+        if exists(self.mlp):
+            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)
@@ -170,12 +164,38 @@ class LinearAttention(nn.Module):
 
     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, qkv=3)
-        k = k.softmax(dim=-1)
-        context = torch.einsum('bhdn,bhen->bhde', k, v)
-        out = torch.einsum('bhde,bhdn->bhen', context, q)
-        out = rearrange(out, 'b heads c (h w) -> b (heads c) h w', heads=self.heads, h=h, w=w)
+        qkv = self.to_qkv(x).chunk(3, dim = 1)
+        q, k, v = map(lambda t: rearrange(t, 'b (h c) x y -> b h c (x y)', h = self.heads), qkv)
+        q = q * self.scale
+
+        k = k.softmax(dim = -1)
+        context = torch.einsum('b h d n, b h e n -> b h d e', k, v)
+
+        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)
+
+class Attention(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)
+        self.to_out = nn.Conv2d(hidden_dim, dim, 1)
+
+    def forward(self, x):
+        b, c, h, w = x.shape
+        qkv = self.to_qkv(x).chunk(3, dim = 1)
+        q, k, v = map(lambda t: rearrange(t, 'b (h c) x y -> b h c (x y)', h = self.heads), qkv)
+        q = q * self.scale
+
+        sim = einsum('b h d i, b h d j -> b h i j', q, k)
+        sim = sim - sim.amax(dim = -1, keepdim = True).detach()
+        attn = sim.softmax(dim = -1)
+
+        out = einsum('b h i j, b h d j -> b h i d', attn, v)
+        out = rearrange(out, 'b h (x y) d -> b (h d) x y', x = h, y = w)
         return self.to_out(out)
 
 # model
@@ -186,8 +206,8 @@ class Unet(nn.Module):
         dim,
         out_dim = None,
         dim_mults=(1, 2, 4, 8),
-        groups = 8,
-        channels = 3
+        channels = 3,
+        with_time_emb = True
     ):
         super().__init__()
         self.channels = channels
@@ -195,12 +215,17 @@ class Unet(nn.Module):
         dims = [channels, *map(lambda m: dim * m, dim_mults)]
         in_out = list(zip(dims[:-1], dims[1:]))
 
-        self.time_pos_emb = SinusoidalPosEmb(dim)
-        self.mlp = nn.Sequential(
-            nn.Linear(dim, dim * 4),
-            Mish(),
-            nn.Linear(dim * 4, dim)
-        )
+        if with_time_emb:
+            time_dim = dim
+            self.time_mlp = nn.Sequential(
+                SinusoidalPosEmb(dim),
+                nn.Linear(dim, dim * 4),
+                nn.GELU(),
+                nn.Linear(dim * 4, dim)
+            )
+        else:
+            time_dim = None
+            self.time_mlp = None
 
         self.downs = nn.ModuleList([])
         self.ups = nn.ModuleList([])
@@ -210,42 +235,41 @@ class Unet(nn.Module):
             is_last = ind >= (num_resolutions - 1)
 
             self.downs.append(nn.ModuleList([
-                ResnetBlock(dim_in, dim_out, time_emb_dim = dim),
-                ResnetBlock(dim_out, dim_out, time_emb_dim = dim),
-                Residual(Rezero(LinearAttention(dim_out))),
+                ConvNextBlock(dim_in, dim_out, time_emb_dim = time_dim, norm = ind != 0),
+                ConvNextBlock(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()
             ]))
 
         mid_dim = dims[-1]
-        self.mid_block1 = ResnetBlock(mid_dim, mid_dim, time_emb_dim = dim)
-        self.mid_attn = Residual(Rezero(LinearAttention(mid_dim)))
-        self.mid_block2 = ResnetBlock(mid_dim, mid_dim, time_emb_dim = dim)
+        self.mid_block1 = ConvNextBlock(mid_dim, mid_dim, time_emb_dim = time_dim)
+        self.mid_attn = Residual(PreNorm(mid_dim, Attention(mid_dim)))
+        self.mid_block2 = ConvNextBlock(mid_dim, mid_dim, time_emb_dim = time_dim)
 
         for ind, (dim_in, dim_out) in enumerate(reversed(in_out[1:])):
             is_last = ind >= (num_resolutions - 1)
 
             self.ups.append(nn.ModuleList([
-                ResnetBlock(dim_out * 2, dim_in, time_emb_dim = dim),
-                ResnetBlock(dim_in, dim_in, time_emb_dim = dim),
-                Residual(Rezero(LinearAttention(dim_in))),
+                ConvNextBlock(dim_out * 2, dim_in, time_emb_dim = time_dim),
+                ConvNextBlock(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()
             ]))
 
         out_dim = default(out_dim, channels)
         self.final_conv = nn.Sequential(
-            Block(dim, dim),
+            ConvNextBlock(dim, dim),
             nn.Conv2d(dim, out_dim, 1)
         )
 
     def forward(self, x, time):
-        t = self.time_pos_emb(time)
-        t = self.mlp(t)
+        t = self.time_mlp(time) if exists(self.time_mlp) else None
 
         h = []
 
-        for resnet, resnet2, attn, downsample in self.downs:
-            x = resnet(x, t)
-            x = resnet2(x, t)
+        for convnext, convnext2, attn, downsample in self.downs:
+            x = convnext(x, t)
+            x = convnext2(x, t)
             x = attn(x)
             h.append(x)
             x = downsample(x)
@@ -254,10 +278,10 @@ class Unet(nn.Module):
         x = self.mid_attn(x)
         x = self.mid_block2(x, t)
 
-        for resnet, resnet2, attn, upsample in self.ups:
+        for convnext, convnext2, attn, upsample in self.ups:
             x = torch.cat((x, h.pop()), dim=1)
-            x = resnet(x, t)
-            x = resnet2(x, t)
+            x = convnext(x, t)
+            x = convnext2(x, t)
             x = attn(x)
             x = upsample(x)
 
@@ -365,7 +389,7 @@ 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_(0., 1.)
+            x_recon.clamp_(-1., 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
@@ -445,17 +469,18 @@ class GaussianDiffusion(nn.Module):
 # dataset classes
 
 class Dataset(data.Dataset):
-    def __init__(self, folder, image_size):
+    def __init__(self, folder, image_size, exts = ['jpg', 'jpeg', 'png']):
         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):
@@ -481,13 +506,19 @@ 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
@@ -506,6 +537,9 @@ 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):
@@ -523,10 +557,10 @@ class Trainer(object):
             'model': self.model.state_dict(),
             'ema': self.ema_model.state_dict()
         }
-        torch.save(data, str(RESULTS_FOLDER / f'model-{milestone}.pt'))
+        torch.save(data, str(self.results_folder / f'model-{milestone}.pt'))
 
     def load(self, milestone):
-        data = torch.load(str(RESULTS_FOLDER / f'model-{milestone}.pt'))
+        data = torch.load(str(self.results_folder / f'model-{milestone}.pt'))
 
         self.step = data['step']
         self.model.load_state_dict(data['model'])
@@ -545,15 +579,16 @@ class Trainer(object):
             self.opt.step()
             self.opt.zero_grad()
 
-            if self.step % UPDATE_EMA_EVERY == 0:
+            if self.step % self.update_ema_every == 0:
                 self.step_ema()
 
-            if self.step != 0 and self.step % SAVE_AND_SAMPLE_EVERY == 0:
-                milestone = self.step // SAVE_AND_SAMPLE_EVERY
+            if self.step != 0 and self.step % self.save_and_sample_every == 0:
+                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(RESULTS_FOLDER / f'sample-{milestone}.png'), nrow=6)
+                all_images = (all_images + 1) * 0.5
+                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 = '0.6.2',
+  version = '0.8.0',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',