pass parameter for Trainer to EMA properly

updated citation in README

README.md

@@ -4,7 +4,11 @@
 
 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>
+
+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://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>
 
@@ -34,7 +38,7 @@ diffusion = GaussianDiffusion(
     loss_type = 'l1'    # L1 or L2
 )
 
-training_images = torch.randn(8, 3, 128, 128) # your images need to be normalized from a range of -1 to +1
+training_images = torch.randn(8, 3, 128, 128) # images are normalized from 0 to 1
 loss = diffusion(training_images)
 loss.backward()
 # after a lot of training
@@ -64,7 +68,7 @@ trainer = Trainer(
     diffusion,
     'path/to/your/images',
     train_batch_size = 32,
-    train_lr = 2e-5,
+    train_lr = 1e-4,
     train_num_steps = 700000,         # total training steps
     gradient_accumulate_every = 2,    # gradient accumulation steps
     ema_decay = 0.995,                # exponential moving average decay
@@ -79,34 +83,53 @@ 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}
+@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},
 }
 ```
 
 ```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},
+@inproceedings{kingma2021on,
+    title       = {On Density Estimation with Diffusion Models},
+    author      = {Diederik P Kingma and Tim Salimans and Ben Poole and Jonathan Ho},
+    booktitle   = {Advances in Neural Information Processing Systems},
+    editor      = {A. Beygelzimer and Y. Dauphin and P. Liang and J. Wortman Vaughan},
+    year        = {2021},
+    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},
-    eprint  = {2201.03545},
-    archivePrefix = {arXiv},
-    primaryClass = {cs.CV}
+    volume  = {abs/2204.00227}
 }
 ```

denoising_diffusion_pytorch/__init__.py

@@ -1 +1,5 @@
 from denoising_diffusion_pytorch.denoising_diffusion_pytorch import GaussianDiffusion, Unet, Trainer
+
+from denoising_diffusion_pytorch.learned_gaussian_diffusion import LearnedGaussianDiffusion
+from denoising_diffusion_pytorch.continuous_time_gaussian_diffusion import ContinuousTimeGaussianDiffusion
+from denoising_diffusion_pytorch.weighted_objective_gaussian_diffusion import WeightedObjectiveGaussianDiffusion

denoising_diffusion_pytorch/continuous_time_gaussian_diffusion.py

@@ -0,0 +1,286 @@
+import torch
+from torch import sqrt
+from torch import nn, einsum
+import torch.nn.functional as F
+from torch.special import expm1
+
+from tqdm import tqdm
+from einops import rearrange, repeat, reduce
+from einops.layers.torch import Rearrange
+
+# helpers
+
+def exists(val):
+    return val is not None
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if callable(d) else d
+
+# normalization functions
+
+def normalize_to_neg_one_to_one(img):
+    return img * 2 - 1
+
+def unnormalize_to_zero_to_one(t):
+    return (t + 1) * 0.5
+
+# diffusion helpers
+
+def right_pad_dims_to(x, t):
+    padding_dims = x.ndim - t.ndim
+    if padding_dims <= 0:
+        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
+# @crowsonkb Katherine's repository also helped here https://github.com/crowsonkb/v-diffusion-jax/blob/master/diffusion/utils.py
+
+# log(snr) that approximates the original linear schedule
+
+def log(t, eps = 1e-20):
+    return torch.log(t.clamp(min = eps))
+
+def beta_linear_log_snr(t):
+    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):
+    """ 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__()
+        self.slope = log_snr_min - log_snr_max
+        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__(
+        self,
+        denoise_fn,
+        *,
+        image_size,
+        channels = 3,
+        loss_type = 'l1',
+        noise_schedule = 'linear',
+        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 denoise_fn.learned_sinusoidal_cond
+
+        self.denoise_fn = denoise_fn
+
+        # image dimensions
+
+        self.channels = channels
+        self.image_size = image_size
+
+        # continuous noise schedule related stuff
+
+        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):
+        return next(self.denoise_fn.parameters()).device
+
+    @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_mean_variance(self, x, time, time_next):
+        # reviewer found an error in the equation in the paper (missing sigma)
+        # following - https://openreview.net/forum?id=2LdBqxc1Yv&noteId=rIQgH0zKsRt
+
+        log_snr = self.log_snr(time)
+        log_snr_next = self.log_snr(time_next)
+        c = -expm1(log_snr - log_snr_next)
+
+        squared_alpha, squared_alpha_next = log_snr.sigmoid(), log_snr_next.sigmoid()
+        squared_sigma, squared_sigma_next = (-log_snr).sigmoid(), (-log_snr_next).sigmoid()
+
+        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
+
+    # sampling related functions
+
+    @torch.no_grad()
+    def p_sample(self, x, time, time_next):
+        batch, *_, device = *x.shape, x.device
+
+        model_mean, model_variance = self.p_mean_variance(x = x, time = time, time_next = time_next)
+
+        if time_next == 0:
+            return model_mean
+
+        noise = torch.randn_like(x)
+        return model_mean + sqrt(model_variance) * noise
+
+    @torch.no_grad()
+    def p_sample_loop(self, shape):
+        batch = shape[0]
+
+        img = torch.randn(shape, device = self.device)
+        steps = torch.linspace(1., 0., self.num_sample_steps + 1, device = self.device)
+
+        for i in tqdm(range(self.num_sample_steps), desc = 'sampling loop time step', total = self.num_sample_steps):
+            times = steps[i]
+            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
+
+    @torch.no_grad()
+    def sample(self, batch_size = 16):
+        return self.p_sample_loop((batch_size, self.channels, self.image_size, self.image_size))
+
+    # training related functions - noise prediction
+
+    def q_sample(self, x_start, times, noise = None):
+        noise = default(noise, lambda: torch.randn_like(x_start))
+
+        log_snr = self.log_snr(times)
+
+        log_snr_padded = right_pad_dims_to(x_start, log_snr)
+        alpha, sigma = sqrt(log_snr_padded.sigmoid()), sqrt((-log_snr_padded).sigmoid())
+        x_noised =  x_start * alpha + noise * sigma
+
+        return x_noised, log_snr
+
+    def random_times(self, batch_size):
+        # times are now uniform from 0 to 1
+        return torch.zeros((batch_size,), device = self.device).float().uniform_(0, 1)
+
+    def p_losses(self, x_start, times, noise = None):
+        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)
+
+        losses = self.loss_fn(model_out, noise, reduction = 'none')
+        losses = reduce(losses, 'b ... -> b', 'mean')
+
+        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
+        assert h == img_size and w == img_size, f'height and width of image must be {img_size}'
+
+        times = self.random_times(b)
+        img = normalize_to_neg_one_to_one(img)
+        return self.p_losses(img, times, *args, **kwargs)

denoising_diffusion_pytorch/denoising_diffusion_pytorch.py

@@ -7,6 +7,7 @@ from inspect import isfunction
 from functools import partial
 
 from torch.utils import data
+from multiprocessing import cpu_count
 from torch.cuda.amp import autocast, GradScaler
 
 from pathlib import Path
@@ -15,7 +16,10 @@ from torchvision import transforms, utils
 from PIL import Image
 
 from tqdm import tqdm
-from einops import rearrange
+from einops import rearrange, reduce
+from einops.layers.torch import Rearrange
+
+from ema_pytorch import EMA
 
 # helpers functions
 
@@ -40,23 +44,14 @@ 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():
-    def __init__(self, beta):
-        super().__init__()
-        self.beta = beta
-
-    def update_model_average(self, ma_model, current_model):
-        for current_params, ma_params in zip(current_model.parameters(), ma_model.parameters()):
-            old_weight, up_weight = ma_params.data, current_params.data
-            ma_params.data = self.update_average(old_weight, up_weight)
-
-    def update_average(self, old, new):
-        if old is None:
-            return new
-        return old * self.beta + (1 - self.beta) * new
-
 class Residual(nn.Module):
     def __init__(self, fn):
         super().__init__()
@@ -65,20 +60,6 @@ class Residual(nn.Module):
     def forward(self, x, *args, **kwargs):
         return self.fn(x, *args, **kwargs) + x
 
-class SinusoidalPosEmb(nn.Module):
-    def __init__(self, dim):
-        super().__init__()
-        self.dim = dim
-
-    def forward(self, x):
-        device = x.device
-        half_dim = self.dim // 2
-        emb = math.log(10000) / (half_dim - 1)
-        emb = torch.exp(torch.arange(half_dim, device=device) * -emb)
-        emb = x[:, None] * emb[None, :]
-        emb = torch.cat((emb.sin(), emb.cos()), dim=-1)
-        return emb
-
 def Upsample(dim):
     return nn.ConvTranspose2d(dim, dim, 4, 2, 1)
 
@@ -107,72 +88,83 @@ class PreNorm(nn.Module):
         x = self.norm(x)
         return self.fn(x)
 
+# sinusoidal positional embeds
+
+class SinusoidalPosEmb(nn.Module):
+    def __init__(self, dim):
+        super().__init__()
+        self.dim = dim
+
+    def forward(self, x):
+        device = x.device
+        half_dim = self.dim // 2
+        emb = math.log(10000) / (half_dim - 1)
+        emb = torch.exp(torch.arange(half_dim, device=device) * -emb)
+        emb = x[:, None] * emb[None, :]
+        emb = torch.cat((emb.sin(), emb.cos()), dim=-1)
+        return emb
+
+class LearnedSinusoidalPosEmb(nn.Module):
+    """ following @crowsonkb 's lead with learned sinusoidal pos emb """
+    """ https://github.com/crowsonkb/v-diffusion-jax/blob/master/diffusion/models/danbooru_128.py#L8 """
+
+    def __init__(self, dim):
+        super().__init__()
+        assert (dim % 2) == 0
+        half_dim = dim // 2
+        self.weights = nn.Parameter(torch.randn(half_dim))
+
+    def forward(self, x):
+        x = rearrange(x, 'b -> b 1')
+        freqs = x * rearrange(self.weights, 'd -> 1 d') * 2 * math.pi
+        fouriered = torch.cat((freqs.sin(), freqs.cos()), dim = -1)
+        fouriered = torch.cat((x, fouriered), dim = -1)
+        return fouriered
+
 # building block modules
 
 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)
+        self.proj = nn.Conv2d(dim, dim_out, 3, padding = 1)
+        self.norm = nn.GroupNorm(groups, dim_out)
+        self.act = nn.SiLU()
+
+    def forward(self, x, scale_shift = None):
+        x = self.proj(x)
+        x = self.norm(x)
+
+        if exists(scale_shift):
+            scale, shift = scale_shift
+            x = x * (scale + 1) + shift
+
+        x = self.act(x)
+        return x
 
 class ResnetBlock(nn.Module):
     def __init__(self, dim, dim_out, *, time_emb_dim = None, groups = 8):
         super().__init__()
         self.mlp = nn.Sequential(
             nn.SiLU(),
-            nn.Linear(time_emb_dim, dim_out)
+            nn.Linear(time_emb_dim, dim_out * 2)
         ) if exists(time_emb_dim) else None
 
-        self.block1 = Block(dim, dim_out)
-        self.block2 = Block(dim_out, dim_out)
+        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.block1(x)
 
+        scale_shift = None
         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
+            time_emb = rearrange(time_emb, 'b c -> b c 1 1')
+            scale_shift = time_emb.chunk(2, dim = 1)
+
+        h = self.block1(x, scale_shift = scale_shift)
 
         h = self.block2(h)
-        return h + self.res_conv(x)
 
-class ConvNextBlock(nn.Module):
-    """ https://arxiv.org/abs/2201.03545 """
-
-    def __init__(self, dim, dim_out, *, time_emb_dim = None, mult = 2, norm = True):
-        super().__init__()
-        self.mlp = nn.Sequential(
-            nn.GELU(),
-            nn.Linear(time_emb_dim, dim)
-        ) 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(),
-            LayerNorm(dim_out * mult),
-            nn.Conv2d(dim_out * mult, dim_out, 3, padding = 1)
-        )
-
-        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)
-
-        if exists(self.mlp) and exists(time_emb):
-            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')
-
-        h = self.net(h)
         return h + self.res_conv(x)
 
 class LinearAttention(nn.Module):
@@ -236,9 +228,10 @@ class Unet(nn.Module):
         out_dim = None,
         dim_mults=(1, 2, 4, 8),
         channels = 3,
-        with_time_emb = True,
-        use_convnext = False,
-        resnet_block_groups = 8
+        resnet_block_groups = 8,
+        learned_variance = False,
+        learned_sinusoidal_cond = False,
+        learned_sinusoidal_dim = 16
     ):
         super().__init__()
 
@@ -246,32 +239,33 @@ class Unet(nn.Module):
 
         self.channels = channels
 
-        init_dim = default(init_dim, dim // 3 * 2)
+        init_dim = default(init_dim, dim)
         self.init_conv = nn.Conv2d(channels, init_dim, 7, padding = 3)
 
         dims = [init_dim, *map(lambda m: dim * m, dim_mults)]
         in_out = list(zip(dims[:-1], dims[1:]))
 
-        # resnet or convnext
-
-        if use_convnext:
-            block_klass = ConvNextBlock
-        else:
-            block_klass = partial(ResnetBlock, groups = resnet_block_groups)
+        block_klass = partial(ResnetBlock, groups = resnet_block_groups)
 
         # time embeddings
 
-        if with_time_emb:
-            time_dim = dim * 4
-            self.time_mlp = nn.Sequential(
-                SinusoidalPosEmb(dim),
-                nn.Linear(dim, time_dim),
-                nn.GELU(),
-                nn.Linear(time_dim, time_dim)
-            )
+        time_dim = dim * 4
+
+        self.learned_sinusoidal_cond = learned_sinusoidal_cond
+
+        if learned_sinusoidal_cond:
+            sinu_pos_emb = LearnedSinusoidalPosEmb(learned_sinusoidal_dim)
+            fourier_dim = learned_sinusoidal_dim + 1
         else:
-            time_dim = None
-            self.time_mlp = None
+            sinu_pos_emb = SinusoidalPosEmb(dim)
+            fourier_dim = dim
+
+        self.time_mlp = nn.Sequential(
+            sinu_pos_emb,
+            nn.Linear(fourier_dim, time_dim),
+            nn.GELU(),
+            nn.Linear(time_dim, time_dim)
+        )
 
         # layers
 
@@ -294,8 +288,8 @@ class Unet(nn.Module):
         self.mid_attn = Residual(PreNorm(mid_dim, Attention(mid_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)
+        for ind, (dim_in, dim_out) in enumerate(reversed(in_out)):
+            is_last = ind == (len(in_out) - 1)
 
             self.ups.append(nn.ModuleList([
                 block_klass(dim_out * 2, dim_in, time_emb_dim = time_dim),
@@ -304,16 +298,17 @@ class Unet(nn.Module):
                 Upsample(dim_in) if not is_last else nn.Identity()
             ]))
 
-        out_dim = default(out_dim, channels)
-        self.final_conv = nn.Sequential(
-            block_klass(dim, dim),
-            nn.Conv2d(dim, out_dim, 1)
-        )
+        default_out_dim = channels * (1 if not learned_variance else 2)
+        self.out_dim = default(out_dim, default_out_dim)
+
+        self.final_res_block = block_klass(dim * 2, dim, time_emb_dim = time_dim)
+        self.final_conv = nn.Conv2d(dim, self.out_dim, 1)
 
     def forward(self, x, time):
         x = self.init_conv(x)
+        r = x.clone()
 
-        t = self.time_mlp(time) if exists(self.time_mlp) else None
+        t = self.time_mlp(time)
 
         h = []
 
@@ -329,12 +324,15 @@ class Unet(nn.Module):
         x = self.mid_block2(x, t)
 
         for block1, block2, attn, upsample in self.ups:
-            x = torch.cat((x, h.pop()), dim=1)
+            x = torch.cat((x, h.pop()), dim = 1)
             x = block1(x, t)
             x = block2(x, t)
             x = attn(x)
             x = upsample(x)
 
+        x = torch.cat((x, r), dim = 1)
+
+        x = self.final_res_block(x, t)
         return self.final_conv(x)
 
 # gaussian diffusion trainer class
@@ -344,10 +342,11 @@ def extract(a, t, x_shape):
     out = a.gather(-1, t)
     return out.reshape(b, *((1,) * (len(x_shape) - 1)))
 
-def noise_like(shape, device, repeat=False):
-    repeat_noise = lambda: torch.randn((1, *shape[1:]), device=device).repeat(shape[0], *((1,) * (len(shape) - 1)))
-    noise = lambda: torch.randn(shape, device=device)
-    return repeat_noise() if repeat else noise()
+def linear_beta_schedule(timesteps):
+    scale = 1000 / timesteps
+    beta_start = scale * 0.0001
+    beta_end = scale * 0.02
+    return torch.linspace(beta_start, beta_end, timesteps, dtype = torch.float64)
 
 def cosine_beta_schedule(timesteps, s = 0.008):
     """
@@ -355,7 +354,7 @@ 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])
@@ -369,14 +368,26 @@ class GaussianDiffusion(nn.Module):
         image_size,
         channels = 3,
         timesteps = 1000,
-        loss_type = 'l1'
+        loss_type = 'l1',
+        objective = 'pred_noise',
+        beta_schedule = 'cosine',
+        p2_loss_weight_gamma = 0., # p2 loss weight, from https://arxiv.org/abs/2204.00227 - 0 is equivalent to weight of 1 across time - 1. is recommended
+        p2_loss_weight_k = 1
     ):
         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
+        self.objective = objective
 
-        betas = cosine_beta_schedule(timesteps)
+        if beta_schedule == 'linear':
+            betas = linear_beta_schedule(timesteps)
+        elif beta_schedule == 'cosine':
+            betas = cosine_beta_schedule(timesteps)
+        else:
+            raise ValueError(f'unknown beta schedule {beta_schedule}')
 
         alphas = 1. - betas
         alphas_cumprod = torch.cumprod(alphas, axis=0)
@@ -386,17 +397,21 @@ 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)
 
@@ -404,19 +419,17 @@ 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
-        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
+        # calculate p2 reweighting
+
+        register_buffer('p2_loss_weight', (p2_loss_weight_k + alphas_cumprod / (1 - alphas_cumprod)) ** -p2_loss_weight_gamma)
 
     def predict_start_from_noise(self, x_t, t, noise):
         return (
@@ -434,19 +447,26 @@ 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()
-    def p_sample(self, x, t, clip_denoised=True, repeat_noise=False):
+    def p_sample(self, x, t, clip_denoised=True):
         b, *_, device = *x.shape, x.device
         model_mean, _, model_log_variance = self.p_mean_variance(x=x, t=t, clip_denoised=clip_denoised)
-        noise = noise_like(x.shape, device, repeat_noise)
+        noise = torch.randn_like(x)
         # no noise when t == 0
         nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
         return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
@@ -460,6 +480,8 @@ class GaussianDiffusion(nn.Module):
 
         for i in tqdm(reversed(range(0, self.num_timesteps)), desc='sampling loop time step', total=self.num_timesteps):
             img = self.p_sample(img, torch.full((b,), i, device=device, dtype=torch.long))
+
+        img = unnormalize_to_zero_to_one(img)
         return img
 
     @torch.no_grad()
@@ -492,32 +514,47 @@ 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))
 
-        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)
 
-        if self.loss_type == 'l1':
-            loss = (noise - x_recon).abs().mean()
-        elif self.loss_type == 'l2':
-            loss = F.mse_loss(noise, x_recon)
+        if self.objective == 'pred_noise':
+            target = noise
+        elif self.objective == 'pred_x0':
+            target = x_start
         else:
-            raise NotImplementedError()
+            raise ValueError(f'unknown objective {self.objective}')
 
-        return loss
+        loss = self.loss_fn(model_out, target, reduction = 'none')
+        loss = reduce(loss, 'b ... -> b (...)', 'mean')
 
-    def forward(self, x, *args, **kwargs):
-        b, c, h, w, device, img_size, = *x.shape, x.device, self.image_size
+        loss = loss * extract(self.p2_loss_weight, t, loss.shape)
+        return loss.mean()
+
+    def forward(self, img, *args, **kwargs):
+        b, c, h, w, device, img_size, = *img.shape, img.device, self.image_size
         assert h == img_size and w == img_size, f'height and width of image must be {img_size}'
         t = torch.randint(0, self.num_timesteps, (b,), device=device).long()
-        return self.p_losses(x, t, *args, **kwargs)
+
+        img = normalize_to_neg_one_to_one(img)
+        return self.p_losses(img, t, *args, **kwargs)
 
 # dataset classes
 
 class Dataset(data.Dataset):
-    def __init__(self, folder, image_size, exts = ['jpg', 'jpeg', 'png']):
+    def __init__(self, folder, image_size, exts = ['jpg', 'jpeg', 'png'], augment_horizontal_flip = False):
         super().__init__()
         self.folder = folder
         self.image_size = image_size
@@ -525,10 +562,9 @@ class Dataset(data.Dataset):
 
         self.transform = transforms.Compose([
             transforms.Resize(image_size),
-            transforms.RandomHorizontalFlip(),
+            transforms.RandomHorizontalFlip() if augment_horizontal_flip else nn.Identity(),
             transforms.CenterCrop(image_size),
-            transforms.ToTensor(),
-            transforms.Lambda(lambda t: (t * 2) - 1)
+            transforms.ToTensor()
         ])
 
     def __len__(self):
@@ -548,22 +584,22 @@ class Trainer(object):
         folder,
         *,
         ema_decay = 0.995,
-        image_size = 128,
         train_batch_size = 32,
-        train_lr = 2e-5,
+        train_lr = 1e-4,
         train_num_steps = 100000,
         gradient_accumulate_every = 2,
         amp = False,
         step_start_ema = 2000,
-        update_ema_every = 10,
+        ema_update_every = 10,
         save_and_sample_every = 1000,
-        results_folder = './results'
+        results_folder = './results',
+        augment_horizontal_flip = True
     ):
         super().__init__()
+        self.image_size = diffusion_model.image_size
+
         self.model = diffusion_model
-        self.ema = EMA(ema_decay)
-        self.ema_model = copy.deepcopy(self.model)
-        self.update_ema_every = update_ema_every
+        self.ema = EMA(diffusion_model, beta = ema_decay, update_every = ema_update_every)
 
         self.step_start_ema = step_start_ema
         self.save_and_sample_every = save_and_sample_every
@@ -573,9 +609,9 @@ class Trainer(object):
         self.gradient_accumulate_every = gradient_accumulate_every
         self.train_num_steps = train_num_steps
 
-        self.ds = Dataset(folder, image_size)
-        self.dl = cycle(data.DataLoader(self.ds, batch_size = train_batch_size, shuffle=True, pin_memory=True))
-        self.opt = Adam(diffusion_model.parameters(), lr=train_lr)
+        self.ds = Dataset(folder, self.image_size, augment_horizontal_flip = augment_horizontal_flip)
+        self.dl = cycle(data.DataLoader(self.ds, batch_size = train_batch_size, shuffle = True, pin_memory = True, num_workers = cpu_count()))
+        self.opt = Adam(diffusion_model.parameters(), lr = train_lr)
 
         self.step = 0
 
@@ -585,22 +621,11 @@ class Trainer(object):
         self.results_folder = Path(results_folder)
         self.results_folder.mkdir(exist_ok = True)
 
-        self.reset_parameters()
-
-    def reset_parameters(self):
-        self.ema_model.load_state_dict(self.model.state_dict())
-
-    def step_ema(self):
-        if self.step < self.step_start_ema:
-            self.reset_parameters()
-            return
-        self.ema.update_model_average(self.ema_model, self.model)
-
     def save(self, milestone):
         data = {
             'step': self.step,
             'model': self.model.state_dict(),
-            'ema': self.ema_model.state_dict(),
+            'ema': self.ema.state_dict(),
             'scaler': self.scaler.state_dict()
         }
         torch.save(data, str(self.results_folder / f'model-{milestone}.pt'))
@@ -610,36 +635,40 @@ class Trainer(object):
 
         self.step = data['step']
         self.model.load_state_dict(data['model'])
-        self.ema_model.load_state_dict(data['ema'])
+        self.ema.load_state_dict(data['ema'])
         self.scaler.load_state_dict(data['scaler'])
 
     def train(self):
-        while self.step < self.train_num_steps:
-            for i in range(self.gradient_accumulate_every):
-                data = next(self.dl).cuda()
+        with tqdm(initial = self.step, total = self.train_num_steps) as pbar:
 
-                with autocast(enabled = self.amp):
-                    loss = self.model(data)
-                    self.scaler.scale(loss / self.gradient_accumulate_every).backward()
+            while self.step < self.train_num_steps:
+                for i in range(self.gradient_accumulate_every):
+                    data = next(self.dl).cuda()
 
-                print(f'{self.step}: {loss.item()}')
+                    with autocast(enabled = self.amp):
+                        loss = self.model(data)
+                        self.scaler.scale(loss / self.gradient_accumulate_every).backward()
 
-            self.scaler.step(self.opt)
-            self.scaler.update()
-            self.opt.zero_grad()
+                    pbar.set_description(f'loss: {loss.item():.4f}')
 
-            if self.step % self.update_ema_every == 0:
-                self.step_ema()
+                self.scaler.step(self.opt)
+                self.scaler.update()
+                self.opt.zero_grad()
 
-            if self.step != 0 and self.step % self.save_and_sample_every == 0:
-                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
-                utils.save_image(all_images, str(self.results_folder / f'sample-{milestone}.png'), nrow = 6)
-                self.save(milestone)
+                self.ema.update()
 
-            self.step += 1
+                if self.step != 0 and self.step % self.save_and_sample_every == 0:
+                    self.ema.ema_model.eval()
+                    with torch.no_grad():
+                        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.ema_model.sample(batch_size=n), batches))
 
-        print('training completed')
+                    all_images = torch.cat(all_images_list, dim=0)
+                    utils.save_image(all_images, str(self.results_folder / f'sample-{milestone}.png'), nrow = 6)
+                    self.save(milestone)
+
+                self.step += 1
+                pbar.update(1)
+
+        print('training complete')

denoising_diffusion_pytorch/learned_gaussian_diffusion.py

@@ -0,0 +1,132 @@
+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 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(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
+
+        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 = 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))
+
+        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

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
+
+# 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

@@ -3,18 +3,20 @@ from setuptools import setup, find_packages
 setup(
   name = 'denoising-diffusion-pytorch',
   packages = find_packages(),
-  version = '0.11.0',
+  version = '0.21.1',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',
   author_email = 'lucidrains@gmail.com',
   url = 'https://github.com/lucidrains/denoising-diffusion-pytorch',
+  long_description_content_type = 'text/markdown',
   keywords = [
     'artificial intelligence',
     'generative models'
   ],
   install_requires=[
     'einops',
+    'ema-pytorch',
     'pillow',
     'torch',
     'torchvision',