make sure to clip when sampling from gaussian diffusion with learned variance

updated citation in README

README.md

@@ -4,7 +4,7 @@
 
 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.
+This implementation was transcribed from the official Tensorflow version <a href="https://github.com/hojonathanho/diffusion">here</a>
 
 <img src="./sample.png" width="500px"><img>
 
@@ -34,7 +34,7 @@ diffusion = GaussianDiffusion(
     loss_type = 'l1'    # L1 or L2
 )
 
-training_images = torch.randn(8, 3, 128, 128)
+training_images = torch.randn(8, 3, 128, 128) # your images need to be normalized from a range of -1 to +1
 loss = diffusion(training_images)
 loss.backward()
 # after a lot of training
@@ -79,34 +79,32 @@ Samples and model checkpoints will be logged to `./results` periodically
 ## Citations
 
 ```bibtex
-@misc{ho2020denoising,
-    title   = {Denoising Diffusion Probabilistic Models},
-    author  = {Jonathan Ho and Ajay Jain and Pieter Abbeel},
-    year    = {2020},
-    eprint  = {2006.11239},
-    archivePrefix = {arXiv},
-    primaryClass = {cs.LG}
+@inproceedings{NEURIPS2020_4c5bcfec,
+    author      = {Ho, Jonathan and Jain, Ajay and Abbeel, Pieter},
+    booktitle   = {Advances in Neural Information Processing Systems},
+    editor      = {H. Larochelle and M. Ranzato and R. Hadsell and M.F. Balcan and H. Lin},
+    pages       = {6840--6851},
+    publisher   = {Curran Associates, Inc.},
+    title       = {Denoising Diffusion Probabilistic Models},
+    url         = {https://proceedings.neurips.cc/paper/2020/file/4c5bcfec8584af0d967f1ab10179ca4b-Paper.pdf},
+    volume      = {33},
+    year        = {2020}
 }
 ```
 
 ```bibtex
-@inproceedings{anonymous2021improved,
-    title   = {Improved Denoising Diffusion Probabilistic Models},
-    author  = {Anonymous},
-    booktitle = {Submitted to International Conference on Learning Representations},
-    year    = {2021},
-    url     = {https://openreview.net/forum?id=-NEXDKk8gZ},
-    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}
+@InProceedings{pmlr-v139-nichol21a,
+    title       = {Improved Denoising Diffusion Probabilistic Models},
+    author      = {Nichol, Alexander Quinn and Dhariwal, Prafulla},
+    booktitle   = {Proceedings of the 38th International Conference on Machine Learning},
+    pages       = {8162--8171},
+    year        = {2021},
+    editor      = {Meila, Marina and Zhang, Tong},
+    volume      = {139},
+    series      = {Proceedings of Machine Learning Research},
+    month       = {18--24 Jul},
+    publisher   = {PMLR},
+    pdf         = {http://proceedings.mlr.press/v139/nichol21a/nichol21a.pdf},
+    url         = {https://proceedings.mlr.press/v139/nichol21a.html},
 }
 ```

denoising_diffusion_pytorch/__init__.py

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

denoising_diffusion_pytorch/denoising_diffusion_pytorch.py

@@ -40,6 +40,12 @@ def num_to_groups(num, divisor):
         arr.append(remainder)
     return arr
 
+def normalize_to_neg_one_to_one(img):
+    return img * 2 - 1
+
+def unnormalize_to_zero_to_one(t):
+    return (t + 1) * 0.5
+
 # small helper modules
 
 class EMA():
@@ -109,36 +115,37 @@ class PreNorm(nn.Module):
 
 # building block modules
 
-class ConvNextBlock(nn.Module):
-    """ https://arxiv.org/abs/2201.03545 """
+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),
+            nn.SiLU()
+        )
+    def forward(self, x):
+        return self.block(x)
 
-    def __init__(self, dim, dim_out, *, time_emb_dim = None, mult = 2, norm = True):
+class ResnetBlock(nn.Module):
+    def __init__(self, dim, dim_out, *, time_emb_dim = None, groups = 8):
         super().__init__()
         self.mlp = nn.Sequential(
-            nn.GELU(),
-            nn.Linear(time_emb_dim, dim)
+            nn.SiLU(),
+            nn.Linear(time_emb_dim, dim_out)
         ) 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, groups = groups)
+        self.block2 = Block(dim_out, dim_out, groups = groups)
         self.res_conv = nn.Conv2d(dim, dim_out, 1) if dim != dim_out else nn.Identity()
 
     def forward(self, x, time_emb = None):
-        h = self.ds_conv(x)
+        h = self.block1(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')
+        if exists(self.mlp) and exists(time_emb):
+            time_emb = self.mlp(time_emb)
+            h = rearrange(time_emb, 'b c -> b c 1 1') + h
 
-        h = self.net(h)
+        h = self.block2(h)
         return h + self.res_conv(x)
 
 class LinearAttention(nn.Module):
@@ -148,15 +155,21 @@ class LinearAttention(nn.Module):
         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)
+
+        self.to_out = nn.Sequential(
+            nn.Conv2d(hidden_dim, dim, 1),
+            LayerNorm(dim)
+        )
 
     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
 
+        q = q.softmax(dim = -2)
         k = k.softmax(dim = -1)
+
+        q = q * self.scale
         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)
@@ -196,9 +209,14 @@ class Unet(nn.Module):
         out_dim = None,
         dim_mults=(1, 2, 4, 8),
         channels = 3,
-        with_time_emb = True
+        with_time_emb = True,
+        resnet_block_groups = 8,
+        learned_variance = False
     ):
         super().__init__()
+
+        # determine dimensions
+
         self.channels = channels
 
         init_dim = default(init_dim, dim // 3 * 2)
@@ -207,6 +225,10 @@ class Unet(nn.Module):
         dims = [init_dim, *map(lambda m: dim * m, dim_mults)]
         in_out = list(zip(dims[:-1], dims[1:]))
 
+        block_klass = partial(ResnetBlock, groups = resnet_block_groups)
+
+        # time embeddings
+
         if with_time_emb:
             time_dim = dim * 4
             self.time_mlp = nn.Sequential(
@@ -219,6 +241,8 @@ class Unet(nn.Module):
             time_dim = None
             self.time_mlp = None
 
+        # layers
+
         self.downs = nn.ModuleList([])
         self.ups = nn.ModuleList([])
         num_resolutions = len(in_out)
@@ -227,31 +251,33 @@ class Unet(nn.Module):
             is_last = ind >= (num_resolutions - 1)
 
             self.downs.append(nn.ModuleList([
-                ConvNextBlock(dim_in, dim_out, time_emb_dim = time_dim, norm = ind != 0),
-                ConvNextBlock(dim_out, dim_out, time_emb_dim = time_dim),
+                block_klass(dim_in, dim_out, time_emb_dim = time_dim),
+                block_klass(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 = ConvNextBlock(mid_dim, mid_dim, time_emb_dim = time_dim)
+        self.mid_block1 = block_klass(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)
+        self.mid_block2 = block_klass(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([
-                ConvNextBlock(dim_out * 2, dim_in, time_emb_dim = time_dim),
-                ConvNextBlock(dim_in, dim_in, time_emb_dim = time_dim),
+                block_klass(dim_out * 2, dim_in, time_emb_dim = time_dim),
+                block_klass(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)
+        default_out_dim = channels * (1 if not learned_variance else 2)
+        self.out_dim = default(out_dim, default_out_dim)
+
         self.final_conv = nn.Sequential(
-            ConvNextBlock(dim, dim),
-            nn.Conv2d(dim, out_dim, 1)
+            block_klass(dim, dim),
+            nn.Conv2d(dim, self.out_dim, 1)
         )
 
     def forward(self, x, time):
@@ -261,9 +287,9 @@ class Unet(nn.Module):
 
         h = []
 
-        for convnext, convnext2, attn, downsample in self.downs:
-            x = convnext(x, t)
-            x = convnext2(x, t)
+        for block1, block2, attn, downsample in self.downs:
+            x = block1(x, t)
+            x = block2(x, t)
             x = attn(x)
             h.append(x)
             x = downsample(x)
@@ -272,10 +298,10 @@ class Unet(nn.Module):
         x = self.mid_attn(x)
         x = self.mid_block2(x, t)
 
-        for convnext, convnext2, attn, upsample in self.ups:
+        for block1, block2, attn, upsample in self.ups:
             x = torch.cat((x, h.pop()), dim=1)
-            x = convnext(x, t)
-            x = convnext2(x, t)
+            x = block1(x, t)
+            x = block2(x, t)
             x = attn(x)
             x = upsample(x)
 
@@ -299,7 +325,7 @@ def cosine_beta_schedule(timesteps, s = 0.008):
     as proposed in https://openreview.net/forum?id=-NEXDKk8gZ
     """
     steps = timesteps + 1
-    x = torch.linspace(0, timesteps, steps)
+    x = torch.linspace(0, timesteps, steps, dtype = torch.float64)
     alphas_cumprod = torch.cos(((x / timesteps) + s) / (1 + s) * torch.pi * 0.5) ** 2
     alphas_cumprod = alphas_cumprod / alphas_cumprod[0]
     betas = 1 - (alphas_cumprod[1:] / alphas_cumprod[:-1])
@@ -316,6 +342,8 @@ class GaussianDiffusion(nn.Module):
         loss_type = 'l1'
     ):
         super().__init__()
+        assert not (type(self) == GaussianDiffusion and denoise_fn.channels != denoise_fn.out_dim)
+
         self.channels = channels
         self.image_size = image_size
         self.denoise_fn = denoise_fn
@@ -330,17 +358,21 @@ class GaussianDiffusion(nn.Module):
         self.num_timesteps = int(timesteps)
         self.loss_type = loss_type
 
-        self.register_buffer('betas', betas)
-        self.register_buffer('alphas_cumprod', alphas_cumprod)
-        self.register_buffer('alphas_cumprod_prev', alphas_cumprod_prev)
+        # helper function to register buffer from float64 to float32
+
+        register_buffer = lambda name, val: self.register_buffer(name, val.to(torch.float32))
+
+        register_buffer('betas', betas)
+        register_buffer('alphas_cumprod', alphas_cumprod)
+        register_buffer('alphas_cumprod_prev', alphas_cumprod_prev)
 
         # calculations for diffusion q(x_t | x_{t-1}) and others
 
-        self.register_buffer('sqrt_alphas_cumprod', torch.sqrt(alphas_cumprod))
-        self.register_buffer('sqrt_one_minus_alphas_cumprod', torch.sqrt(1. - alphas_cumprod))
-        self.register_buffer('log_one_minus_alphas_cumprod', torch.log(1. - alphas_cumprod))
-        self.register_buffer('sqrt_recip_alphas_cumprod', torch.sqrt(1. / alphas_cumprod))
-        self.register_buffer('sqrt_recipm1_alphas_cumprod', torch.sqrt(1. / alphas_cumprod - 1))
+        register_buffer('sqrt_alphas_cumprod', torch.sqrt(alphas_cumprod))
+        register_buffer('sqrt_one_minus_alphas_cumprod', torch.sqrt(1. - alphas_cumprod))
+        register_buffer('log_one_minus_alphas_cumprod', torch.log(1. - alphas_cumprod))
+        register_buffer('sqrt_recip_alphas_cumprod', torch.sqrt(1. / alphas_cumprod))
+        register_buffer('sqrt_recipm1_alphas_cumprod', torch.sqrt(1. / alphas_cumprod - 1))
 
         # calculations for posterior q(x_{t-1} | x_t, x_0)
 
@@ -348,13 +380,13 @@ class GaussianDiffusion(nn.Module):
 
         # above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t)
 
-        self.register_buffer('posterior_variance', posterior_variance)
+        register_buffer('posterior_variance', posterior_variance)
 
         # below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain
 
-        self.register_buffer('posterior_log_variance_clipped', torch.log(posterior_variance.clamp(min =1e-20)))
-        self.register_buffer('posterior_mean_coef1', betas * torch.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod))
-        self.register_buffer('posterior_mean_coef2', (1. - alphas_cumprod_prev) * torch.sqrt(alphas) / (1. - alphas_cumprod))
+        register_buffer('posterior_log_variance_clipped', torch.log(posterior_variance.clamp(min =1e-20)))
+        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
@@ -436,6 +468,15 @@ class GaussianDiffusion(nn.Module):
             extract(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise
         )
 
+    @property
+    def loss_fn(self):
+        if self.loss_type == 'l1':
+            return F.l1_loss
+        elif self.loss_type == 'l2':
+            return F.mse_loss
+        else:
+            raise ValueError(f'invalid loss type {self.loss_type}')
+
     def p_losses(self, x_start, t, noise = None):
         b, c, h, w = x_start.shape
         noise = default(noise, lambda: torch.randn_like(x_start))
@@ -443,13 +484,7 @@ class GaussianDiffusion(nn.Module):
         x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
         x_recon = self.denoise_fn(x_noisy, t)
 
-        if self.loss_type == 'l1':
-            loss = (noise - x_recon).abs().mean()
-        elif self.loss_type == 'l2':
-            loss = F.mse_loss(noise, x_recon)
-        else:
-            raise NotImplementedError()
-
+        loss = self.loss_fn(noise, x_recon)
         return loss
 
     def forward(self, x, *args, **kwargs):
@@ -472,7 +507,7 @@ class Dataset(data.Dataset):
             transforms.RandomHorizontalFlip(),
             transforms.CenterCrop(image_size),
             transforms.ToTensor(),
-            transforms.Lambda(lambda t: (t * 2) - 1)
+            transforms.Lambda(normalize_to_neg_one_to_one)
         ])
 
     def __len__(self):
@@ -576,11 +611,13 @@ class Trainer(object):
                 self.step_ema()
 
             if self.step != 0 and self.step % self.save_and_sample_every == 0:
+                self.ema_model.eval()
+
                 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)
-                all_images = (all_images + 1) * 0.5
+                all_images = unnormalize_to_zero_to_one(all_images)
                 utils.save_image(all_images, str(self.results_folder / f'sample-{milestone}.png'), nrow = 6)
                 self.save(milestone)
 

denoising_diffusion_pytorch/learned_gaussian_diffusion.py

@@ -0,0 +1,149 @@
+import torch
+from math import pi, sqrt, log as ln
+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
+
+# constants
+
+NAT = 1. / ln(2)
+
+# 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
+
+# tensor helpers
+
+def log(t, eps = 1e-12):
+    return torch.log(t.clamp(min = eps))
+
+def meanflat(x):
+    return x.mean(dim = tuple(range(1, len(x.shape))))
+
+def normal_kl(mean1, logvar1, mean2, logvar2):
+    """
+    KL divergence between normal distributions parameterized by mean and log-variance.
+    """
+    return 0.5 * (-1.0 + logvar2 - logvar1 + torch.exp(logvar1 - logvar2) + ((mean1 - mean2) ** 2) * torch.exp(-logvar2))
+
+def approx_standard_normal_cdf(x):
+    return 0.5 * (1.0 + torch.tanh(sqrt(2.0 / pi) * (x + 0.044715 * (x ** 3))))
+
+def discretized_gaussian_log_likelihood(x, *, means, log_scales, thres = 0.999):
+    assert x.shape == means.shape == log_scales.shape
+
+    centered_x = x - means
+    inv_stdv = torch.exp(-log_scales)
+    plus_in = inv_stdv * (centered_x + 1. / 255.)
+    cdf_plus = approx_standard_normal_cdf(plus_in)
+    min_in = inv_stdv * (centered_x - 1. / 255.)
+    cdf_min = approx_standard_normal_cdf(min_in)
+    log_cdf_plus = log(cdf_plus)
+    log_one_minus_cdf_min = log(1. - cdf_min)
+    cdf_delta = cdf_plus - cdf_min
+
+    log_probs = torch.where(x < -thres,
+        log_cdf_plus,
+        torch.where(x > thres,
+            log_one_minus_cdf_min,
+            log(cdf_delta)))
+
+    return log_probs
+
+# https://arxiv.org/abs/2102.09672
+
+# i thought the results were questionable, if one were to focus only on FID
+# but may as well get this in here for others to try, as GLIDE is using it (and DALL-E2 first stage of cascade)
+# gaussian diffusion for learned variance + hybrid eps simple + vb loss
+
+class LearnedGaussianDiffusion(GaussianDiffusion):
+    def __init__(
+        self,
+        denoise_fn,
+        vb_loss_weight = 0.001,  # lambda was 0.001 in the paper
+        *args,
+        **kwargs
+    ):
+        super().__init__(denoise_fn, *args, **kwargs)
+        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)
+
+        min_log = extract(self.posterior_log_variance_clipped, t, x.shape)
+        max_log = extract(torch.log(self.betas), t, x.shape)
+        var_interp_frac = unnormalize_to_zero_to_one(var_interp_frac_unnormalized)
+
+        model_log_variance = var_interp_frac * max_log + (1 - var_interp_frac) * min_log
+        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
+
+    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
+
+        model_output = self.denoise_fn(x_t, t)
+
+        # 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)
+        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)
+        kl = meanflat(kl) * NAT
+
+        decoder_nll = -discretized_gaussian_log_likelihood(x_start, means = 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))
+
+        vb_losses = torch.where(t == 0, decoder_nll, kl)
+
+        # simple loss - predicting noise, x0, or x_prev
+
+        pred_noise, _ = model_output.chunk(2, dim = 1)
+
+        simple_losses = self.loss_fn(pred_noise, noise)
+
+        return simple_losses + vb_losses.mean() * self.vb_loss_weight

setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'denoising-diffusion-pytorch',
   packages = find_packages(),
-  version = '0.10.1',
+  version = '0.14.1',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',