add p2 loss weighting to SNR version of denoising diffusion, brought up by @Mut1nyJD, paper is https://arxiv.org/abs/2204.00227

README.md

@@ -6,6 +6,8 @@ Implementation of <a href="https://arxiv.org/abs/2006.11239">Denoising Diffusion
 
 This implementation was transcribed from the official Tensorflow version <a href="https://github.com/hojonathanho/diffusion">here</a>
 
+<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>
 
 [![PyPI version](https://badge.fury.io/py/denoising-diffusion-pytorch.svg)](https://badge.fury.io/py/denoising-diffusion-pytorch)
@@ -119,3 +121,13 @@ Samples and model checkpoints will be logged to `./results` periodically
     url         = {https://openreview.net/forum?id=2LdBqxc1Yv}
 }
 ```
+
+```bibtex
+@article{Choi2022PerceptionPT,
+    title   = {Perception Prioritized Training of Diffusion Models},
+    author  = {Jooyoung Choi and Jungbeom Lee and Chaehun Shin and Sungwon Kim and Hyunwoo J. Kim and Sung-Hoon Yoon},
+    journal = {ArXiv},
+    year    = {2022},
+    volume  = {abs/2204.00227}
+}
+```

denoising_diffusion_pytorch/continuous_time_gaussian_diffusion.py

@@ -6,6 +6,7 @@ from torch.special import expm1
 
 from tqdm import tqdm
 from einops import rearrange, repeat
+from einops.layers.torch import Rearrange
 
 # helpers
 
@@ -33,6 +34,24 @@ def right_pad_dims_to(x, t):
         return t
     return t.view(*t.shape, *((1,) * padding_dims))
 
+# neural net helpers
+
+class Residual(nn.Module):
+    def __init__(self, fn):
+        super().__init__()
+        self.fn = fn
+
+    def forward(self, x):
+        return x + self.fn(x)
+
+class MonotonicLinear(nn.Module):
+    def __init__(self, *args, **kwargs):
+        super().__init__()
+        self.net = nn.Linear(*args, **kwargs)
+
+    def forward(self, x):
+        return F.linear(x, self.net.weight.abs(), self.net.bias.abs())
+
 # continuous schedules
 
 # equations are taken from https://openreview.net/attachment?id=2LdBqxc1Yv&name=supplementary_material
@@ -40,17 +59,54 @@ def right_pad_dims_to(x, t):
 
 # log(snr) that approximates the original linear schedule
 
-def beta_linear_log_snr(t):
+def log(t, eps = 1e-20):
-    return -torch.log(expm1(1e-4 + 10 * (t ** 2)))
+    return torch.log(t.clamp(min = eps))
 
-def alpha_cosine_log_snr(t):
+def beta_linear_log_snr(t):
-    raise NotImplementedError
+    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)
 
 class learned_noise_schedule(nn.Module):
-    def __init__(self):
+    """ described in section H and then I.2 of the supplementary material for variational ddpm paper """
+
+    def __init__(
+        self,
+        *,
+        log_snr_max,
+        log_snr_min,
+        hidden_dim = 1024,
+        frac_gradient = 1.
+    ):
         super().__init__()
-        raise NotImplementedError
+        self.slope = log_snr_min - log_snr_max
-        # learned noise schedule, using learned monotonic MLP (weights kept positive) in the paper
+        self.intercept = log_snr_max
+
+        self.net = nn.Sequential(
+            Rearrange('... -> ... 1'),
+            MonotonicLinear(1, 1),
+            Residual(nn.Sequential(
+                MonotonicLinear(1, hidden_dim),
+                nn.Sigmoid(),
+                MonotonicLinear(hidden_dim, 1)
+            )),
+            Rearrange('... 1 -> ...'),
+        )
+
+        self.frac_gradient = frac_gradient
+
+    def forward(self, x):
+        frac_gradient = self.frac_gradient
+        device = x.device
+
+        out_zero = self.net(torch.zeros_like(x))
+        out_one =  self.net(torch.ones_like(x))
+
+        x = self.net(x)
+
+        normed = self.slope * ((x - out_zero) / (out_one - out_zero)) + self.intercept
+        return normed * frac_gradient + normed.detach() * (1 - frac_gradient)
 
 class ContinuousTimeGaussianDiffusion(nn.Module):
     def __init__(
@@ -59,12 +115,17 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
         *,
         image_size,
         channels = 3,
-        cond_scale = 500,
         loss_type = 'l1',
         noise_schedule = 'linear',
-        num_sample_steps = 500
+        num_sample_steps = 500,
+        clip_sample_denoised = True,
+        learned_schedule_net_hidden_dim = 1024,
+        learned_noise_schedule_frac_gradient = 1.,   # between 0 and 1, determines what percentage of gradients go back, so one can update the learned noise schedule more slowly
+        p2_loss_weight_gamma = 0.,                   # p2 loss weight, from https://arxiv.org/abs/2204.00227 - 0 is equivalent to weight of 1 across time
+        p2_loss_weight_k = 1
     ):
         super().__init__()
+        assert not denoise_fn.sinusoidal_cond_mlp
 
         self.denoise_fn = denoise_fn
 
@@ -75,17 +136,36 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
 
         # continuous noise schedule related stuff
 
-        self.cond_scale = cond_scale # the log(snr) will be scaled by this value
         self.loss_type = loss_type
 
         if noise_schedule == 'linear':
             self.log_snr = beta_linear_log_snr
+        elif noise_schedule == 'cosine':
+            self.log_snr = alpha_cosine_log_snr
+        elif noise_schedule == 'learned':
+            log_snr_max, log_snr_min = [beta_linear_log_snr(torch.tensor([time])).item() for time in (0., 1.)]
+
+            self.log_snr = learned_noise_schedule(
+                log_snr_max = log_snr_max,
+                log_snr_min = log_snr_min,
+                hidden_dim = learned_schedule_net_hidden_dim,
+                frac_gradient = learned_noise_schedule_frac_gradient
+            )
         else:
             raise ValueError(f'unknown noise schedule {noise_schedule}')
 
         # sampling
 
         self.num_sample_steps = num_sample_steps
+        self.clip_sample_denoised = clip_sample_denoised
+
+        # p2 loss weight
+        # proposed https://arxiv.org/abs/2204.00227
+
+        assert p2_loss_weight_gamma <= 2, 'in paper, they noticed any gamma greater than 2 is harmful'
+
+        self.p2_loss_weight_gamma = p2_loss_weight_gamma  # recommended to be 0.5 or 1
+        self.p2_loss_weight_k = p2_loss_weight_k
 
     @property
     def device(self):
@@ -104,14 +184,6 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
         # reviewer found an error in the equation in the paper (missing sigma)
         # following - https://openreview.net/forum?id=2LdBqxc1Yv&noteId=rIQgH0zKsRt
 
-        # todo - derive x_start from the posterior mean and do dynamic thresholding
-        # assumed that is what is going on in Imagen
-
-        batch = x.shape[0]
-        batch_time = repeat(time, ' -> b', b = batch)
-
-        pred_noise = self.denoise_fn(x, batch_time * self.cond_scale)
-
         log_snr = self.log_snr(time)
         log_snr_next = self.log_snr(time_next)
         c = -expm1(log_snr - log_snr_next)
@@ -119,7 +191,21 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
         squared_alpha, squared_alpha_next = log_snr.sigmoid(), log_snr_next.sigmoid()
         squared_sigma, squared_sigma_next = (-log_snr).sigmoid(), (-log_snr_next).sigmoid()
 
-        model_mean = sqrt(squared_alpha_next / squared_alpha) * (x - c * sqrt(squared_sigma) * pred_noise)
+        alpha, sigma, alpha_next = map(sqrt, (squared_alpha, squared_sigma, squared_alpha_next))
+
+        batch_log_snr = repeat(log_snr, ' -> b', b = x.shape[0])
+        pred_noise = self.denoise_fn(x, batch_log_snr)
+
+        if self.clip_sample_denoised:
+            x_start = (x - sigma * pred_noise) / alpha
+
+            # in Imagen, this was changed to dynamic thresholding
+            x_start.clamp_(-1., 1.)
+
+            model_mean = alpha_next * (x * (1 - c) / alpha + c * x_start)
+        else:
+            model_mean = alpha_next / alpha * (x - c * sigma * pred_noise)
+
         posterior_variance = squared_sigma_next * c
 
         return model_mean, posterior_variance
@@ -150,6 +236,7 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
             times_next = steps[i + 1]
             img = self.p_sample(img, times, times_next)
 
+        img.clamp_(-1., 1.)
         img = unnormalize_to_zero_to_one(img)
         return img
 
@@ -178,9 +265,17 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
         noise = default(noise, lambda: torch.randn_like(x_start))
 
         x, log_snr = self.q_sample(x_start = x_start, times = times, noise = noise)
+        model_out = self.denoise_fn(x, log_snr)
 
-        model_out = self.denoise_fn(x, log_snr * self.cond_scale)
+        losses = self.loss_fn(model_out, noise, reduction = 'none')
-        return self.loss_fn(model_out, noise)
+        losses = losses.mean(dim = tuple(range(1, losses.ndim)))
+
+        if self.p2_loss_weight_gamma >= 0:
+            # following eq 8. in https://arxiv.org/abs/2204.00227
+            loss_weight = (self.p2_loss_weight_k + log_snr.exp()) ** -self.p2_loss_weight_gamma
+            losses = losses * loss_weight
+
+        return losses.mean()
 
     def forward(self, img, *args, **kwargs):
         b, c, h, w, device, img_size, = *img.shape, img.device, self.image_size

denoising_diffusion_pytorch/denoising_diffusion_pytorch.py

@@ -16,6 +16,7 @@ from PIL import Image
 
 from tqdm import tqdm
 from einops import rearrange
+from einops.layers.torch import Rearrange
 
 # helpers functions
 
@@ -211,6 +212,18 @@ class Attention(nn.Module):
 
 # model
 
+def MLP(dim_in, dim_hidden):
+    return nn.Sequential(
+        Rearrange('... -> ... 1'),
+        nn.Linear(1, dim_hidden),
+        nn.GELU(),
+        nn.LayerNorm(dim_hidden),
+        nn.Linear(dim_hidden, dim_hidden),
+        nn.GELU(),
+        nn.LayerNorm(dim_hidden),
+        nn.Linear(dim_hidden, dim_hidden)
+    )
+
 class Unet(nn.Module):
     def __init__(
         self,
@@ -219,9 +232,9 @@ class Unet(nn.Module):
         out_dim = None,
         dim_mults=(1, 2, 4, 8),
         channels = 3,
-        with_time_emb = True,
         resnet_block_groups = 8,
-        learned_variance = False
+        learned_variance = False,
+        sinusoidal_cond_mlp = True
     ):
         super().__init__()
 
@@ -239,8 +252,11 @@ class Unet(nn.Module):
 
         # time embeddings
 
-        if with_time_emb:
+        time_dim = dim * 4
-            time_dim = dim * 4
+
+        self.sinusoidal_cond_mlp = sinusoidal_cond_mlp
+
+        if sinusoidal_cond_mlp:
             self.time_mlp = nn.Sequential(
                 SinusoidalPosEmb(dim),
                 nn.Linear(dim, time_dim),
@@ -248,8 +264,7 @@ class Unet(nn.Module):
                 nn.Linear(time_dim, time_dim)
             )
         else:
-            time_dim = None
+            self.time_mlp = MLP(1, time_dim)
-            self.time_mlp = None
 
         # layers
 
@@ -292,8 +307,7 @@ class Unet(nn.Module):
 
     def forward(self, x, time):
         x = self.init_conv(x)
-
+        t = self.time_mlp(time)
-        t = self.time_mlp(time) if exists(self.time_mlp) else None
 
         h = []
 

setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'denoising-diffusion-pytorch',
   packages = find_packages(),
-  version = '0.16.4',
+  version = '0.18.0',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',