higher epsilon for fp16 in layernorm

README.md

@@ -1,4 +1,4 @@
-<img src="./denoising-diffusion.png" width="500px"></img>
+<img src="./images/denoising-diffusion.png" width="500px"></img>
 
 ## Denoising Diffusion Probabilistic Model, in Pytorch
 
@@ -10,7 +10,7 @@ Youtube AI Educators - <a href="https://www.youtube.com/watch?v=W-O7AZNzbzQ">Yan
 
 <a href="https://huggingface.co/blog/annotated-diffusion">Annotated code</a> by Research Scientists / Engineers from <a href="https://huggingface.co/">🤗 Huggingface</a>
 
-<img src="./sample.png" width="500px"><img>
+<img src="./images/sample.png" width="500px"><img>
 
 [![PyPI version](https://badge.fury.io/py/denoising-diffusion-pytorch.svg)](https://badge.fury.io/py/denoising-diffusion-pytorch)
 
@@ -69,7 +69,7 @@ trainer = Trainer(
     diffusion,
     'path/to/your/images',
     train_batch_size = 32,
-    train_lr = 1e-4,
+    train_lr = 8e-5,
     train_num_steps = 700000,         # total training steps
     gradient_accumulate_every = 2,    # gradient accumulation steps
     ema_decay = 0.995,                # exponential moving average decay

denoising_diffusion_pytorch/denoising_diffusion_pytorch.py

@@ -58,6 +58,9 @@ def convert_image_to(img_type, image):
         return image.convert(img_type)
     return image
 
+def l2norm(t):
+    return F.normalize(t, dim = -1)
+
 # normalization functions
 
 def normalize_to_neg_one_to_one(img):
@@ -86,16 +89,15 @@ 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):
+    def __init__(self, dim):
         super().__init__()
-        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):
+        eps = 1e-5 if x.dtype == torch.float32 else 1e-3
         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
+        return (x - mean) * (var + eps).rsqrt() * self.g
 
 class PreNorm(nn.Module):
     def __init__(self, dim, fn):
@@ -208,6 +210,8 @@ class LinearAttention(nn.Module):
         k = k.softmax(dim = -1)
 
         q = q * self.scale
+        v = v / (h * w)
+
         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)
@@ -215,9 +219,9 @@ class LinearAttention(nn.Module):
         return self.to_out(out)
 
 class Attention(nn.Module):
-    def __init__(self, dim, heads = 4, dim_head = 32):
+    def __init__(self, dim, heads = 4, dim_head = 32, scale = 16):
         super().__init__()
-        self.scale = dim_head ** -0.5
+        self.scale = scale
         self.heads = heads
         hidden_dim = dim_head * heads
         self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias = False)
@@ -227,10 +231,10 @@ class Attention(nn.Module):
         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()
+        q, k = map(l2norm, (q, k))
+
+        sim = einsum('b h d i, b h d j -> b h i j', q, k) * self.scale
         attn = sim.softmax(dim = -1)
 
         out = einsum('b h i j, b h d j -> b h i d', attn, v)
@@ -476,7 +480,7 @@ class GaussianDiffusion(nn.Module):
 
     def predict_noise_from_start(self, x_t, t, x0):
         return (
-            (x0 - extract(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t) / \
+            (extract(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - x0) / \
             extract(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape)
         )
 
@@ -533,11 +537,10 @@ class GaussianDiffusion(nn.Module):
         return img
 
     @torch.no_grad()
-    def ddim_sample(self, shape, clip_denoised = False):
+    def ddim_sample(self, shape, clip_denoised = True):
         batch, device, total_timesteps, sampling_timesteps, eta, objective = shape[0], self.betas.device, self.num_timesteps, self.sampling_timesteps, self.ddim_sampling_eta, self.objective
 
         times = torch.linspace(0., total_timesteps, steps = sampling_timesteps + 2)[:-1]
-
         times = list(reversed(times.int().tolist()))
         time_pairs = list(zip(times[:-1], times[1:]))
 
@@ -551,12 +554,17 @@ class GaussianDiffusion(nn.Module):
 
             pred_noise, x_start, *_ = self.model_predictions(img, time_cond)
 
-            c1 = eta * ((1 - alpha / alpha_next) * (1 - alpha_next) / (1 - alpha)).sqrt()
-            c2 = ((1 - alpha_next) - torch.square(c1)).sqrt()
+            if clip_denoised:
+                x_start.clamp_(-1., 1.)
+
+            sigma = eta * ((1 - alpha / alpha_next) * (1 - alpha_next) / (1 - alpha)).sqrt()
+            c = ((1 - alpha_next) - sigma ** 2).sqrt()
+
+            noise = torch.randn_like(img) if time_next > 0 else 0.
 
             img = x_start * alpha_next.sqrt() + \
-                  c1 * torch.randn_like(img) + \
-                  c2 * pred_noise
+                  c * pred_noise + \
+                  sigma * noise
 
         img = unnormalize_to_zero_to_one(img)
         return img
@@ -677,6 +685,7 @@ class Trainer(object):
         train_num_steps = 100000,
         ema_update_every = 10,
         ema_decay = 0.995,
+        adam_betas = (0.9, 0.99),
         save_and_sample_every = 1000,
         num_samples = 25,
         results_folder = './results',
@@ -711,11 +720,12 @@ class Trainer(object):
         self.ds = Dataset(folder, self.image_size, augment_horizontal_flip = augment_horizontal_flip, convert_image_to = convert_image_to)
         dl = DataLoader(self.ds, batch_size = train_batch_size, shuffle = True, pin_memory = True, num_workers = cpu_count())
 
+        dl = self.accelerator.prepare(dl)
         self.dl = cycle(dl)
 
         # optimizer
 
-        self.opt = Adam(diffusion_model.parameters(), lr = train_lr)
+        self.opt = Adam(diffusion_model.parameters(), lr = train_lr, betas = adam_betas)
 
         # for logging results in a folder periodically
 
@@ -731,7 +741,7 @@ class Trainer(object):
 
         # prepare model, dataloader, optimizer with accelerator
 
-        self.model, self.dl, self.opt = self.accelerator.prepare(self.model, self.dl, self.opt)
+        self.model, self.opt = self.accelerator.prepare(self.model, self.opt)
 
     def save(self, milestone):
         if not self.accelerator.is_local_main_process:
@@ -768,14 +778,19 @@ class Trainer(object):
 
             while self.step < self.train_num_steps:
 
+                total_loss = 0.
+
                 for _ in range(self.gradient_accumulate_every):
                     data = next(self.dl).to(device)
 
                     with self.accelerator.autocast():
                         loss = self.model(data)
-                        self.accelerator.backward(loss / self.gradient_accumulate_every)
+                        loss = loss / self.gradient_accumulate_every
+                        total_loss += loss.item()
 
-                pbar.set_description(f'loss: {loss.item():.4f}')
+                    self.accelerator.backward(loss)
+
+                pbar.set_description(f'loss: {total_loss:.4f}')
 
                 accelerator.wait_for_everyone()
 

setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'denoising-diffusion-pytorch',
   packages = find_packages(),
-  version = '0.25.0',
+  version = '0.26.5',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',