switch back to regular attention, given @rromb results

Added gradient clipping.

README.md

@@ -8,8 +8,12 @@ This implementation was transcribed from the official Tensorflow version <a href
 
 Youtube AI Educators - <a href="https://www.youtube.com/watch?v=W-O7AZNzbzQ">Yannic Kilcher</a> | <a href="https://www.youtube.com/watch?v=344w5h24-h8">AI Coffeebreak with Letitia</a> | <a href="https://www.youtube.com/watch?v=HoKDTa5jHvg">Outlier</a>
 
+<a href="https://github.com/yiyixuxu/denoising-diffusion-flax">Flax implementation</a> from <a href="https://github.com/yiyixuxu">YiYi Xu</a>
+
 <a href="https://huggingface.co/blog/annotated-diffusion">Annotated code</a> by Research Scientists / Engineers from <a href="https://huggingface.co/">🤗 Huggingface</a>
 
+Update: Turns out none of the technicalities really matters at all | <a href="https://arxiv.org/abs/2208.09392">"Cold Diffusion" paper</a>
+
 <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)

denoising_diffusion_pytorch/denoising_diffusion_pytorch.py

@@ -37,6 +37,9 @@ def default(val, d):
         return val
     return d() if callable(d) else d
 
+def identity(t, *args, **kwargs):
+    return t
+
 def cycle(dl):
     while True:
         for data in dl:
@@ -53,14 +56,11 @@ def num_to_groups(num, divisor):
         arr.append(remainder)
     return arr
 
-def convert_image_to(img_type, image):
+def convert_image_to_fn(img_type, image):
     if image.mode != img_type:
         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):
@@ -97,16 +97,11 @@ class WeightStandardizedConv2d(nn.Conv2d):
         eps = 1e-5 if x.dtype == torch.float32 else 1e-3
 
         weight = self.weight
-        flattened_weights = rearrange(weight, 'o ... -> o (...)')
-
         mean = reduce(weight, 'o ... -> o 1 1 1', 'mean')
+        var = reduce(weight, 'o ... -> o 1 1 1', partial(torch.var, unbiased = False))
+        normalized_weight = (weight - mean) * (var + eps).rsqrt()
 
-        var = torch.var(flattened_weights, dim = -1, unbiased = False)
+        return F.conv2d(x, normalized_weight, self.bias, self.stride, self.padding, self.dilation, self.groups)
-        var = rearrange(var, 'o -> o 1 1 1')
-
-        weight = (weight - mean) * (var + eps).rsqrt()
-
-        return F.conv2d(x, weight, self.bias, self.stride, self.padding, self.dilation, self.groups)
 
 class LayerNorm(nn.Module):
     def __init__(self, dim):
@@ -241,9 +236,10 @@ class LinearAttention(nn.Module):
 class Attention(nn.Module):
     def __init__(self, dim, heads = 4, dim_head = 32, scale = 10):
         super().__init__()
-        self.scale = scale
+        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)
 
@@ -252,11 +248,12 @@ class Attention(nn.Module):
         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, k = map(l2norm, (q, k))
+        q = q * self.scale
 
-        sim = einsum('b h d i, b h d j -> b h i j', q, k) * self.scale
+        sim = einsum('b h d i, b h d j -> b h i j', q, k)
         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)
 
@@ -432,6 +429,7 @@ class GaussianDiffusion(nn.Module):
     ):
         super().__init__()
         assert not (type(self) == GaussianDiffusion and model.channels != model.out_dim)
+        assert not model.learned_sinusoidal_cond
 
         self.model = model
         self.channels = self.model.channels
@@ -451,7 +449,7 @@ class GaussianDiffusion(nn.Module):
             raise ValueError(f'unknown beta schedule {beta_schedule}')
 
         alphas = 1. - betas
-        alphas_cumprod = torch.cumprod(alphas, axis=0)
+        alphas_cumprod = torch.cumprod(alphas, dim=0)
         alphas_cumprod_prev = F.pad(alphas_cumprod[:-1], (1, 0), value = 1.)
 
         timesteps, = betas.shape
@@ -521,16 +519,19 @@ class GaussianDiffusion(nn.Module):
         posterior_log_variance_clipped = extract(self.posterior_log_variance_clipped, t, x_t.shape)
         return posterior_mean, posterior_variance, posterior_log_variance_clipped
 
-    def model_predictions(self, x, t, x_self_cond = None):
+    def model_predictions(self, x, t, x_self_cond = None, clip_x_start = False):
         model_output = self.model(x, t, x_self_cond)
+        maybe_clip = partial(torch.clamp, min = -1., max = 1.) if clip_x_start else identity
 
         if self.objective == 'pred_noise':
             pred_noise = model_output
-            x_start = self.predict_start_from_noise(x, t, model_output)
+            x_start = self.predict_start_from_noise(x, t, pred_noise)
+            x_start = maybe_clip(x_start)
 
         elif self.objective == 'pred_x0':
-            pred_noise = self.predict_noise_from_start(x, t, model_output)
             x_start = model_output
+            x_start = maybe_clip(x_start)
+            pred_noise = self.predict_noise_from_start(x, t, x_start)
 
         return ModelPrediction(pred_noise, x_start)
 
@@ -572,31 +573,30 @@ class GaussianDiffusion(nn.Module):
     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 = torch.linspace(-1, total_timesteps - 1, steps=sampling_timesteps + 1)   # [-1, 0, 1, 2, ..., T-1] when sampling_timesteps == total_timesteps
         times = list(reversed(times.int().tolist()))
-        time_pairs = list(zip(times[:-1], times[1:]))
+        time_pairs = list(zip(times[:-1], times[1:])) # [(T-1, T-2), (T-2, T-3), ..., (1, 0), (0, -1)]
 
         img = torch.randn(shape, device = device)
 
         x_start = None
 
         for time, time_next in tqdm(time_pairs, desc = 'sampling loop time step'):
-            alpha = self.alphas_cumprod_prev[time]
+            time_cond = torch.full((batch,), time, device=device, dtype=torch.long)
-            alpha_next = self.alphas_cumprod_prev[time_next]
-
-            time_cond = torch.full((batch,), time, device = device, dtype = torch.long)
-
             self_cond = x_start if self.self_condition else None
+            pred_noise, x_start, *_ = self.model_predictions(img, time_cond, self_cond, clip_x_start = clip_denoised)
 
-            pred_noise, x_start, *_ = self.model_predictions(img, time_cond, self_cond)
+            if time_next < 0:
+                img = x_start
+                continue
 
-            if clip_denoised:
+            alpha = self.alphas_cumprod[time]
-                x_start.clamp_(-1., 1.)
+            alpha_next = self.alphas_cumprod[time_next]
 
             sigma = eta * ((1 - alpha / alpha_next) * (1 - alpha_next) / (1 - alpha)).sqrt()
-            c = ((1 - alpha_next) - sigma ** 2).sqrt()
+            c = (1 - alpha_next - sigma ** 2).sqrt()
 
-            noise = torch.randn_like(img) if time_next > 0 else 0.
+            noise = torch.randn_like(img)
 
             img = x_start * alpha_next.sqrt() + \
                   c * pred_noise + \
@@ -703,7 +703,7 @@ class Dataset(Dataset):
         self.image_size = image_size
         self.paths = [p for ext in exts for p in Path(f'{folder}').glob(f'**/*.{ext}')]
 
-        maybe_convert_fn = partial(convert_image_to, convert_image_to) if exists(convert_image_to) else nn.Identity()
+        maybe_convert_fn = partial(convert_image_to_fn, convert_image_to) if exists(convert_image_to) else nn.Identity()
 
         self.transform = T.Compose([
             T.Lambda(maybe_convert_fn),
@@ -809,7 +809,10 @@ class Trainer(object):
         torch.save(data, str(self.results_folder / f'model-{milestone}.pt'))
 
     def load(self, milestone):
-        data = torch.load(str(self.results_folder / f'model-{milestone}.pt'))
+        accelerator = self.accelerator
+        device = accelerator.device
+
+        data = torch.load(str(self.results_folder / f'model-{milestone}.pt'), map_location=device)
 
         model = self.accelerator.unwrap_model(self.model)
         model.load_state_dict(data['model'])
@@ -841,6 +844,7 @@ class Trainer(object):
 
                     self.accelerator.backward(loss)
 
+                accelerator.clip_grad_norm_(self.model.parameters(), 1.0)
                 pbar.set_description(f'loss: {total_loss:.4f}')
 
                 accelerator.wait_for_everyone()
@@ -850,6 +854,7 @@ class Trainer(object):
 
                 accelerator.wait_for_everyone()
 
+                self.step += 1
                 if accelerator.is_main_process:
                     self.ema.to(device)
                     self.ema.update()
@@ -866,7 +871,6 @@ class Trainer(object):
                         utils.save_image(all_images, str(self.results_folder / f'sample-{milestone}.png'), nrow = int(math.sqrt(self.num_samples)))
                         self.save(milestone)
 
-                self.step += 1
                 pbar.update(1)
 
         accelerator.print('training complete')

setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'denoising-diffusion-pytorch',
   packages = find_packages(),
-  version = '0.27.3',
+  version = '0.28.0',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',
@@ -30,4 +30,4 @@ setup(
     'License :: OSI Approved :: MIT License',
     'Programming Language :: Python :: 3.6',
   ],
-)
+)