add gaussian diffusion where model predicts both noise and x_start, with a learned weighting between the two (experimental)

denoising_diffusion_pytorch/__init__.py

@@ -1,2 +1,4 @@
 from denoising_diffusion_pytorch.denoising_diffusion_pytorch import GaussianDiffusion, Unet, Trainer
+
 from denoising_diffusion_pytorch.learned_gaussian_diffusion import LearnedGaussianDiffusion
+from denoising_diffusion_pytorch.weighted_objective_gaussian_diffusion import WeightedObjectiveGaussianDiffusion

denoising_diffusion_pytorch/denoising_diffusion_pytorch.py

@@ -339,7 +339,8 @@ class GaussianDiffusion(nn.Module):
         image_size,
         channels = 3,
         timesteps = 1000,
-        loss_type = 'l1'
+        loss_type = 'l1',
+        objective = 'pred_noise'
     ):
         super().__init__()
         assert not (type(self) == GaussianDiffusion and denoise_fn.channels != denoise_fn.out_dim)
@@ -347,6 +348,7 @@ class GaussianDiffusion(nn.Module):
         self.channels = channels
         self.image_size = image_size
         self.denoise_fn = denoise_fn
+        self.objective = objective
 
         betas = cosine_beta_schedule(timesteps)
 
@@ -388,12 +390,6 @@ class GaussianDiffusion(nn.Module):
         register_buffer('posterior_mean_coef1', betas * torch.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod))
         register_buffer('posterior_mean_coef2', (1. - alphas_cumprod_prev) * torch.sqrt(alphas) / (1. - alphas_cumprod))
 
-    def q_mean_variance(self, x_start, t):
-        mean = extract(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start
-        variance = extract(1. - self.alphas_cumprod, t, x_start.shape)
-        log_variance = extract(self.log_one_minus_alphas_cumprod, t, x_start.shape)
-        return mean, variance, log_variance
-
     def predict_start_from_noise(self, x_t, t, noise):
         return (
             extract(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t -
@@ -410,12 +406,19 @@ class GaussianDiffusion(nn.Module):
         return posterior_mean, posterior_variance, posterior_log_variance_clipped
 
     def p_mean_variance(self, x, t, clip_denoised: bool):
-        x_recon = self.predict_start_from_noise(x, t=t, noise=self.denoise_fn(x, t))
+        model_output = self.denoise_fn(x, t)
+
+        if self.objective == 'pred_noise':
+            x_start = self.predict_start_from_noise(x, t = t, noise = model_output)
+        elif self.objective == 'pred_x0':
+            x_start = model_output
+        else:
+            raise ValueError(f'unknown objective {self.objective}')
 
         if clip_denoised:
-            x_recon.clamp_(-1., 1.)
+            x_start.clamp_(-1., 1.)
 
-        model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
+        model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start = x_start, x_t = x, t = t)
         return model_mean, posterior_variance, posterior_log_variance
 
     @torch.no_grad()
@@ -481,10 +484,17 @@ class GaussianDiffusion(nn.Module):
         b, c, h, w = x_start.shape
         noise = default(noise, lambda: torch.randn_like(x_start))
 
-        x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
-        x_recon = self.denoise_fn(x_noisy, t)
+        x = self.q_sample(x_start=x_start, t=t, noise=noise)
+        model_out = self.denoise_fn(x, t)
 
-        loss = self.loss_fn(noise, x_recon)
+        if self.objective == 'pred_noise':
+            target = noise
+        elif self.objective == 'pred_x0':
+            target = x_start
+        else:
+            raise ValueError(f'unknown objective {self.objective}')
+
+        loss = self.loss_fn(model_out, target)
         return loss
 
     def forward(self, x, *args, **kwargs):

denoising_diffusion_pytorch/learned_gaussian_diffusion.py

@@ -76,25 +76,6 @@ class LearnedGaussianDiffusion(GaussianDiffusion):
         assert denoise_fn.out_dim == (denoise_fn.channels * 2), 'dimension out of unet must be twice the number of channels for learned variance - you can also set the `learned_variance` keyword argument on the Unet to be `True`'
         self.vb_loss_weight = vb_loss_weight
 
-    def q_posterior_mean_variance(self, x_start, x_t, t):
-        """
-        Compute the mean and variance of the diffusion posterior q(x_{t-1} | x_t, x_0)
-        """
-        posterior_mean = (
-            extract(self.posterior_mean_coef1, t, x_t.shape) * x_start +
-            extract(self.posterior_mean_coef2, t, x_t.shape) * x_t
-        )
-        posterior_variance = extract(self.posterior_variance, t, x_t.shape)
-        posterior_log_variance_clipped = extract(self.posterior_log_variance_clipped, t, x_t.shape)
-        return posterior_mean, posterior_variance, posterior_log_variance_clipped
-
-    def predict_xstart_from_xprev(self, x_t, t, xprev):
-        # (xprev - coef2*x_t) / coef1
-        return (
-            extract(1. / self.posterior_mean_coef1, t, x_t.shape) * xprev -
-            extract(self.posterior_mean_coef2 / self.posterior_mean_coef1, t, x_t.shape) * x_t
-        )
-
     def p_mean_variance(self, *, x, t, clip_denoised, model_output = None):
         model_output = default(model_output, lambda: self.denoise_fn(x, t))
         pred_noise, var_interp_frac_unnormalized = model_output.chunk(2, dim = 1)
@@ -107,6 +88,10 @@ class LearnedGaussianDiffusion(GaussianDiffusion):
         model_variance = model_log_variance.exp()
 
         x_start = self.predict_start_from_noise(x, t, pred_noise)
+
+        if clip_denoised:
+            x_start.clamp_(-1., 1.)
+
         model_mean, _, _ = self.q_posterior(x_start, x, t)
 
         return model_mean, model_variance, model_log_variance
@@ -121,15 +106,17 @@ class LearnedGaussianDiffusion(GaussianDiffusion):
 
         # calculating kl loss for learned variance (interpolation)
 
-        true_mean, _, true_log_variance_clipped = self.q_posterior_mean_variance(x_start = x_start, x_t = x_t, t = t)
+        true_mean, _, true_log_variance_clipped = self.q_posterior(x_start = x_start, x_t = x_t, t = t)
         model_mean, _, model_log_variance = self.p_mean_variance(x = x_t, t = t, clip_denoised = clip_denoised, model_output = model_output)
 
         # kl loss with detached model predicted mean, for stability reasons as in paper
 
-        kl = normal_kl(true_mean, true_log_variance_clipped, model_mean.detach(), model_log_variance)
+        detached_model_mean = model_mean.detach()
+
+        kl = normal_kl(true_mean, true_log_variance_clipped, detached_model_mean, model_log_variance)
         kl = meanflat(kl) * NAT
 
-        decoder_nll = -discretized_gaussian_log_likelihood(x_start, means = model_mean, log_scales = 0.5 * model_log_variance)
+        decoder_nll = -discretized_gaussian_log_likelihood(x_start, means = detached_model_mean, log_scales = 0.5 * model_log_variance)
         decoder_nll = meanflat(decoder_nll) * NAT
 
         # at the first timestep return the decoder NLL, otherwise return KL(q(x_{t-1}|x_t,x_0) || p(x_{t-1}|x_t))

denoising_diffusion_pytorch/weighted_objective_gaussian_diffusion.py

@@ -0,0 +1,80 @@
+import torch
+from inspect import isfunction
+from torch import nn, einsum
+from einops import rearrange
+
+from denoising_diffusion_pytorch.denoising_diffusion_pytorch import GaussianDiffusion, extract, unnormalize_to_zero_to_one
+
+# helper functions
+
+def exists(x):
+    return x is not None
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+# some improvisation on my end
+# where i have the model learn to both predict noise and x0
+# and learn the weighted sum for each depending on time step
+
+class WeightedObjectiveGaussianDiffusion(GaussianDiffusion):
+    def __init__(
+        self,
+        denoise_fn,
+        *args,
+        pred_noise_loss_weight = 0.1,
+        pred_x_start_loss_weight = 0.1,
+        **kwargs
+    ):
+        super().__init__(denoise_fn, *args, **kwargs)
+        channels = denoise_fn.channels
+        assert denoise_fn.out_dim == (channels * 2 + 2), 'dimension out (out_dim) of unet must be twice the number of channels + 2 (for the softmax weighted sum) - for channels of 3, this should be (3 * 2) + 2 = 8'
+
+        self.split_dims = (channels, channels, 2)
+        self.pred_noise_loss_weight = pred_noise_loss_weight
+        self.pred_x_start_loss_weight = pred_x_start_loss_weight
+
+    def p_mean_variance(self, *, x, t, clip_denoised, model_output = None):
+        model_output = self.denoise_fn(x, t)
+
+        pred_noise, pred_x_start, weights = model_output.split(self.split_dims, dim = 1)
+        normalized_weights = weights.softmax(dim = 1)
+
+        x_start_from_noise = self.predict_start_from_noise(x, t = t, noise = pred_noise)
+        
+        x_starts = torch.stack((x_start_from_noise, pred_x_start), dim = 1)
+        weighted_x_start = einsum('b j h w, b j c h w -> b c h w', normalized_weights, x_starts)
+
+        if clip_denoised:
+            weighted_x_start.clamp_(-1., 1.)
+
+        model_mean, model_variance, model_log_variance = self.q_posterior(weighted_x_start, x, t)
+
+        return model_mean, model_variance, model_log_variance
+
+    def p_losses(self, x_start, t, noise = None, clip_denoised = False):
+        noise = default(noise, lambda: torch.randn_like(x_start))
+        x_t = self.q_sample(x_start = x_start, t = t, noise = noise)
+
+        model_output = self.denoise_fn(x_t, t)
+        pred_noise, pred_x_start, weights = model_output.split(self.split_dims, dim = 1)
+
+        # get loss for predicted noise and x_start
+        # with the loss weight given at initialization
+
+        noise_loss = self.loss_fn(noise, pred_noise) * self.pred_noise_loss_weight
+        x_start_loss = self.loss_fn(x_start, pred_x_start) * self.pred_x_start_loss_weight
+
+        # calculate x_start from predicted noise
+        # then do a weighted sum of the x_start prediction, weights also predicted by the model (softmax normalized)
+
+        x_start_from_pred_noise = self.predict_start_from_noise(x_t, t, pred_noise)
+        x_start_from_pred_noise = x_start_from_pred_noise.clamp(-2., 2.)
+        weighted_x_start = einsum('b j h w, b j c h w -> b c h w', weights.softmax(dim = 1), torch.stack((x_start_from_pred_noise, pred_x_start), dim = 1))
+
+        # main loss to x_start with the weighted one
+
+        weighted_x_start_loss = self.loss_fn(x_start, weighted_x_start)
+        return weighted_x_start_loss + x_start_loss + noise_loss

setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'denoising-diffusion-pytorch',
   packages = find_packages(),
-  version = '0.14.0',
+  version = '0.15.1',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',