get working version of gaussian diffusion with continuous time (only beta linear schedule for now, but will eventually contain alpha cosine schedule as well as parameterized, learned monotonic MLP)

README.md

@@ -6,10 +6,6 @@ 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>
 
-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>
 
 [![PyPI version](https://badge.fury.io/py/denoising-diffusion-pytorch.svg)](https://badge.fury.io/py/denoising-diffusion-pytorch)
@@ -123,23 +119,3 @@ 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}
-}
-```
-
-```bibtex
-@article{Karras2022ElucidatingTD,
-    title   = {Elucidating the Design Space of Diffusion-Based Generative Models},
-    author  = {Tero Karras and Miika Aittala and Timo Aila and Samuli Laine},
-    journal = {ArXiv},
-    year    = {2022},
-    volume  = {abs/2206.00364}
-}
-```

denoising_diffusion_pytorch/__init__.py

@@ -3,4 +3,3 @@ from denoising_diffusion_pytorch.denoising_diffusion_pytorch import GaussianDiff
 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
-from denoising_diffusion_pytorch.elucidated_diffusion import ElucidatedDiffusion

denoising_diffusion_pytorch/continuous_time_gaussian_diffusion.py

@@ -1,4 +1,3 @@
-import math
 import torch
 from torch import sqrt
 from torch import nn, einsum
@@ -6,8 +5,7 @@ import torch.nn.functional as F
 from torch.special import expm1
 
 from tqdm import tqdm
-from einops import rearrange, repeat, reduce
+from einops import rearrange, repeat
-from einops.layers.torch import Rearrange
 
 # helpers
 
@@ -35,24 +33,6 @@ 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
@@ -60,54 +40,17 @@ class MonotonicLinear(nn.Module):
 
 # 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)))
+    return -torch.log(expm1(1e-4 + 10 * (t ** 2)))
 
-def alpha_cosine_log_snr(t, s = 0.008):
+def alpha_cosine_log_snr(t):
-    return -log((torch.cos((t + s) / (1 + s) * math.pi * 0.5) ** -2) - 1, eps = 1e-5)
+    raise NotImplementedError
 
 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):
-
-    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
+        raise NotImplementedError
-        self.intercept = log_snr_max
+        # learned noise schedule, using learned monotonic MLP (weights kept positive) in the paper
-
-        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__(
@@ -116,17 +59,12 @@ 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 denoise_fn.learned_sinusoidal_cond
 
         self.denoise_fn = denoise_fn
 
@@ -137,36 +75,17 @@ 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):
@@ -185,6 +104,14 @@ 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)
@@ -192,21 +119,7 @@ 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()
 
-        alpha, sigma, alpha_next = map(sqrt, (squared_alpha, squared_sigma, squared_alpha_next))
+        model_mean = sqrt(squared_alpha_next / squared_alpha) * (x - c * sqrt(squared_sigma) * pred_noise)
-
-        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
@@ -214,16 +127,18 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
     # sampling related functions
 
     @torch.no_grad()
-    def p_sample(self, x, time, time_next):
+    def p_sample(self, x, time, time_next, eps = 2e-4):
         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
+
+        # no noise when time is below some epsilon
+        # not sure how important this is
+
+        time = repeat(time, ' -> b', b = batch)
+        nonzero_mask = (1 - (time < eps).float()).reshape(batch, *((1,) * (len(x.shape) - 1)))
+        return model_mean + nonzero_mask * sqrt(model_variance) * noise
 
     @torch.no_grad()
     def p_sample_loop(self, shape):
@@ -237,7 +152,6 @@ 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
 
@@ -266,17 +180,9 @@ 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)
 
-        losses = self.loss_fn(model_out, noise, reduction = 'none')
+        model_out = self.denoise_fn(x, log_snr * self.cond_scale)
-        losses = reduce(losses, 'b ... -> b', 'mean')
+        return self.loss_fn(model_out, noise)
-
-        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

@@ -7,7 +7,6 @@ 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
@@ -16,10 +15,7 @@ from torchvision import transforms, utils
 from PIL import Image
 
 from tqdm import tqdm
-from einops import rearrange, reduce
+from einops import rearrange
-from einops.layers.torch import Rearrange
-
-from ema_pytorch import EMA
 
 # helpers functions
 
@@ -52,6 +48,21 @@ def unnormalize_to_zero_to_one(t):
 
 # 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__()
@@ -60,14 +71,25 @@ class Residual(nn.Module):
     def forward(self, x, *args, **kwargs):
         return self.fn(x, *args, **kwargs) + x
 
-def Upsample(dim, dim_out = None):
+class SinusoidalPosEmb(nn.Module):
-    return nn.Sequential(
+    def __init__(self, dim):
-        nn.Upsample(scale_factor = 2, mode = 'nearest'),
+        super().__init__()
-        nn.Conv2d(dim, default(dim_out, dim), 3, padding = 1)
+        self.dim = dim
-    )
 
-def Downsample(dim, dim_out = None):
+    def forward(self, x):
-    return nn.Conv2d(dim, default(dim_out, dim), 4, 2, 1)
+        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)
+
+def Downsample(dim):
+    return nn.Conv2d(dim, dim, 4, 2, 1)
 
 class LayerNorm(nn.Module):
     def __init__(self, dim, eps = 1e-5):
@@ -91,39 +113,6 @@ 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):
@@ -167,7 +156,6 @@ class ResnetBlock(nn.Module):
         h = self.block1(x, scale_shift = scale_shift)
 
         h = self.block2(h)
-
         return h + self.res_conv(x)
 
 class LinearAttention(nn.Module):
@@ -231,10 +219,9 @@ 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
-        learned_sinusoidal_cond = False,
-        learned_sinusoidal_dim = 16
     ):
         super().__init__()
 
@@ -242,7 +229,7 @@ class Unet(nn.Module):
 
         self.channels = channels
 
-        init_dim = default(init_dim, dim)
+        init_dim = default(init_dim, dim // 3 * 2)
         self.init_conv = nn.Conv2d(channels, init_dim, 7, padding = 3)
 
         dims = [init_dim, *map(lambda m: dim * m, dim_mults)]
@@ -252,23 +239,17 @@ class Unet(nn.Module):
 
         # time embeddings
 
-        time_dim = dim * 4
+        if with_time_emb:
-
+            time_dim = dim * 4
-        self.learned_sinusoidal_cond = learned_sinusoidal_cond
+            self.time_mlp = nn.Sequential(
-
+                SinusoidalPosEmb(dim),
-        if learned_sinusoidal_cond:
+                nn.Linear(dim, time_dim),
-            sinu_pos_emb = LearnedSinusoidalPosEmb(learned_sinusoidal_dim)
+                nn.GELU(),
-            fourier_dim = learned_sinusoidal_dim + 1
+                nn.Linear(time_dim, time_dim)
+            )
         else:
-            sinu_pos_emb = SinusoidalPosEmb(dim)
+            time_dim = None
-            fourier_dim = dim
+            self.time_mlp = None
-
-        self.time_mlp = nn.Sequential(
-            sinu_pos_emb,
-            nn.Linear(fourier_dim, time_dim),
-            nn.GELU(),
-            nn.Linear(time_dim, time_dim)
-        )
 
         # layers
 
@@ -280,10 +261,10 @@ class Unet(nn.Module):
             is_last = ind >= (num_resolutions - 1)
 
             self.downs.append(nn.ModuleList([
-                block_klass(dim_in, dim_in, time_emb_dim = time_dim),
+                block_klass(dim_in, dim_out, time_emb_dim = time_dim),
-                block_klass(dim_in, dim_in, time_emb_dim = time_dim),
+                block_klass(dim_out, dim_out, time_emb_dim = time_dim),
-                Residual(PreNorm(dim_in, LinearAttention(dim_in))),
+                Residual(PreNorm(dim_out, LinearAttention(dim_out))),
-                Downsample(dim_in, dim_out) if not is_last else nn.Conv2d(dim_in, dim_out, 3, padding = 1)
+                Downsample(dim_out) if not is_last else nn.Identity()
             ]))
 
         mid_dim = dims[-1]
@@ -291,38 +272,36 @@ 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)):
+        for ind, (dim_in, dim_out) in enumerate(reversed(in_out[1:])):
-            is_last = ind == (len(in_out) - 1)
+            is_last = ind >= (num_resolutions - 1)
 
             self.ups.append(nn.ModuleList([
-                block_klass(dim_out + dim_in, dim_out, time_emb_dim = time_dim),
+                block_klass(dim_out * 2, dim_in, time_emb_dim = time_dim),
-                block_klass(dim_out + dim_in, dim_out, time_emb_dim = time_dim),
+                block_klass(dim_in, dim_in, time_emb_dim = time_dim),
-                Residual(PreNorm(dim_out, LinearAttention(dim_out))),
+                Residual(PreNorm(dim_in, LinearAttention(dim_in))),
-                Upsample(dim_out, dim_in) if not is_last else  nn.Conv2d(dim_out, dim_in, 3, padding = 1)
+                Upsample(dim_in) if not is_last else nn.Identity()
             ]))
 
         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.Sequential(
-        self.final_conv = nn.Conv2d(dim, self.out_dim, 1)
+            block_klass(dim, dim),
+            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)
+        t = self.time_mlp(time) if exists(self.time_mlp) else None
 
         h = []
 
         for block1, block2, attn, downsample in self.downs:
             x = block1(x, t)
-            h.append(x)
-
             x = block2(x, t)
             x = attn(x)
             h.append(x)
-
             x = downsample(x)
 
         x = self.mid_block1(x, t)
@@ -330,18 +309,12 @@ 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 = torch.cat((x, h.pop()), dim = 1)
             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
@@ -364,7 +337,7 @@ def cosine_beta_schedule(timesteps, s = 0.008):
     """
     steps = timesteps + 1
     x = torch.linspace(0, timesteps, steps, dtype = torch.float64)
-    alphas_cumprod = torch.cos(((x / timesteps) + s) / (1 + s) * math.pi * 0.5) ** 2
+    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])
     return torch.clip(betas, 0, 0.999)
@@ -379,9 +352,7 @@ class GaussianDiffusion(nn.Module):
         timesteps = 1000,
         loss_type = 'l1',
         objective = 'pred_noise',
-        beta_schedule = 'cosine',
+        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)
@@ -436,10 +407,6 @@ 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))
 
-        # 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 (
             extract(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t -
@@ -546,11 +513,8 @@ class GaussianDiffusion(nn.Module):
         else:
             raise ValueError(f'unknown objective {self.objective}')
 
-        loss = self.loss_fn(model_out, target, reduction = 'none')
+        loss = self.loss_fn(model_out, target)
-        loss = reduce(loss, 'b ... -> b (...)', 'mean')
+        return loss
-
-        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
@@ -563,7 +527,7 @@ class GaussianDiffusion(nn.Module):
 # dataset classes
 
 class Dataset(data.Dataset):
-    def __init__(self, folder, image_size, exts = ['jpg', 'jpeg', 'png'], augment_horizontal_flip = False):
+    def __init__(self, folder, image_size, exts = ['jpg', 'jpeg', 'png']):
         super().__init__()
         self.folder = folder
         self.image_size = image_size
@@ -571,7 +535,7 @@ class Dataset(data.Dataset):
 
         self.transform = transforms.Compose([
             transforms.Resize(image_size),
-            transforms.RandomHorizontalFlip() if augment_horizontal_flip else nn.Identity(),
+            transforms.RandomHorizontalFlip(),
             transforms.CenterCrop(image_size),
             transforms.ToTensor()
         ])
@@ -593,22 +557,22 @@ class Trainer(object):
         folder,
         *,
         ema_decay = 0.995,
+        image_size = 128,
         train_batch_size = 32,
         train_lr = 1e-4,
         train_num_steps = 100000,
         gradient_accumulate_every = 2,
         amp = False,
         step_start_ema = 2000,
-        ema_update_every = 10,
+        update_ema_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(diffusion_model, beta = ema_decay, update_every = ema_update_every)
+        self.ema = EMA(ema_decay)
+        self.ema_model = copy.deepcopy(self.model)
+        self.update_ema_every = update_ema_every
 
         self.step_start_ema = step_start_ema
         self.save_and_sample_every = save_and_sample_every
@@ -618,9 +582,9 @@ class Trainer(object):
         self.gradient_accumulate_every = gradient_accumulate_every
         self.train_num_steps = train_num_steps
 
-        self.ds = Dataset(folder, self.image_size, augment_horizontal_flip = augment_horizontal_flip)
+        self.ds = Dataset(folder, image_size)
-        self.dl = cycle(data.DataLoader(self.ds, batch_size = train_batch_size, shuffle = True, pin_memory = True, num_workers = cpu_count()))
+        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.opt = Adam(diffusion_model.parameters(), lr=train_lr)
 
         self.step = 0
 
@@ -630,11 +594,22 @@ 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.state_dict(),
+            'ema': self.ema_model.state_dict(),
             'scaler': self.scaler.state_dict()
         }
         torch.save(data, str(self.results_folder / f'model-{milestone}.pt'))
@@ -644,7 +619,7 @@ class Trainer(object):
 
         self.step = data['step']
         self.model.load_state_dict(data['model'])
-        self.ema.load_state_dict(data['ema'])
+        self.ema_model.load_state_dict(data['ema'])
         self.scaler.load_state_dict(data['scaler'])
 
     def train(self):
@@ -664,15 +639,15 @@ class Trainer(object):
                 self.scaler.update()
                 self.opt.zero_grad()
 
-                self.ema.update()
+                if self.step % self.update_ema_every == 0:
+                    self.step_ema()
 
                 if self.step != 0 and self.step % self.save_and_sample_every == 0:
-                    self.ema.ema_model.eval()
+                    self.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))
 
+                    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)
                     utils.save_image(all_images, str(self.results_folder / f'sample-{milestone}.png'), nrow = 6)
                     self.save(milestone)

denoising_diffusion_pytorch/elucidated_diffusion.py

@@ -1,217 +0,0 @@
-from math import sqrt
-import torch
-from torch import nn, einsum
-import torch.nn.functional as F
-
-from tqdm import tqdm
-from einops import rearrange, repeat, reduce
-
-# 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
-
-# tensor helpers
-
-def log(t, eps = 1e-20):
-    return torch.log(t.clamp(min = eps))
-
-# 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
-
-# main class
-
-class ElucidatedDiffusion(nn.Module):
-    def __init__(
-        self,
-        net,
-        *,
-        image_size,
-        channels = 3,
-        num_sample_steps = 32, # number of sampling steps
-        sigma_min = 0.002,     # min noise level
-        sigma_max = 80,        # max noise level
-        sigma_data = 0.5,      # standard deviation of data distribution
-        rho = 7,               # controls the sampling schedule
-        P_mean = -1.2,         # mean of log-normal distribution from which noise is drawn for training
-        P_std = 1.2,           # standard deviation of log-normal distribution from which noise is drawn for training
-        S_churn = 80,          # parameters for stochastic sampling - depends on dataset, Table 5 in apper
-        S_tmin = 0.05,
-        S_tmax = 50,
-        S_noise = 1.003,
-    ):
-        super().__init__()
-        assert net.learned_sinusoidal_cond
-
-        self.net = net
-
-        # image dimensions
-
-        self.channels = channels
-        self.image_size = image_size
-
-        # parameters
-
-        self.sigma_min = sigma_min
-        self.sigma_max = sigma_max
-        self.sigma_data = sigma_data
-
-        self.rho = rho
-
-        self.P_mean = P_mean
-        self.P_std = P_std
-
-        self.num_sample_steps = num_sample_steps  # otherwise known as N in the paper
-
-        self.S_churn = S_churn
-        self.S_tmin = S_tmin
-        self.S_tmax = S_tmax
-        self.S_noise = S_noise
-
-    @property
-    def device(self):
-        return next(self.net.parameters()).device
-
-    # derived preconditioning params - Table 1
-
-    def c_skip(self, sigma):
-        return (self.sigma_data ** 2) / (sigma ** 2 + self.sigma_data ** 2)
-
-    def c_out(self, sigma):
-        return sigma * self.sigma_data * (self.sigma_data ** 2 + sigma ** 2) ** -0.5
-
-    def c_in(self, sigma):
-        return 1 * (sigma ** 2 + self.sigma_data ** 2) ** -0.5
-
-    def c_noise(self, sigma):
-        return log(sigma) * 0.25
-
-    # noise distribution
-
-    def noise_distribution(self, batch_size):
-        return (self.P_mean + self.P_std * torch.randn((batch_size,), device = self.device)).exp()
-
-    def loss_weight(self, sigma):
-        return (sigma ** 2 + self.sigma_data ** 2) * (sigma * self.sigma_data) ** -2
-
-    # sample schedule
-    # equation (5) in the paper
-
-    def sample_schedule(self, num_sample_steps = None):
-        num_sample_steps = default(num_sample_steps, self.num_sample_steps)
-
-        N = num_sample_steps
-        inv_rho = 1 / self.rho
-
-        steps = torch.arange(num_sample_steps, device = self.device, dtype = torch.float32)
-        sigmas = (self.sigma_max ** inv_rho + steps / (N - 1) * (self.sigma_min ** inv_rho - self.sigma_max ** inv_rho)) ** self.rho
-
-        sigmas = F.pad(sigmas, (0, 1), value = 0.) # last step is sigma value of 0.
-        return sigmas
-
-    # preconditioned network output
-    # equation (7) in the paper
-
-    def preconditioned_network_forward(self, noised_images, sigma):
-        batch, device = noised_images.shape[0], noised_images.device
-
-        if isinstance(sigma, float):
-            sigma = torch.full((batch,), sigma, device = device)
-
-        padded_sigma = rearrange(sigma, 'b -> b 1 1 1')
-
-        net_out = self.net(
-            self.c_in(padded_sigma) * noised_images,
-            self.c_noise(sigma)
-        )
-
-        return self.c_skip(padded_sigma) * noised_images +  self.c_out(padded_sigma) * net_out
-
-    # sampling
-
-    @torch.no_grad()
-    def sample(self, batch_size = 16, num_sample_steps = None):
-        num_sample_steps = default(num_sample_steps, self.num_sample_steps)
-
-        shape = (batch_size, self.channels, self.image_size, self.image_size)
-
-        # get the schedule, which is returned as (sigma, gamma) tuple, and pair up with the next sigma and gamma
-
-        sigmas = self.sample_schedule(num_sample_steps)
-
-        gammas = torch.where(
-            (sigmas >= self.S_tmin) & (sigmas <= self.S_tmax),
-            min(self.S_churn / num_sample_steps, sqrt(2) - 1),
-            0.
-        )
-
-        sigmas_and_gammas = list(zip(sigmas[:-1], sigmas[1:], gammas[:-1]))
-
-        # images is noise at the beginning
-
-        init_sigma = sigmas[0]
-
-        images = init_sigma * torch.randn(shape, device = self.device)
-
-        # gradually denoise
-
-        for sigma, sigma_next, gamma in tqdm(sigmas_and_gammas, desc = 'sampling time step'):
-            sigma, sigma_next, gamma = map(lambda t: t.item(), (sigma, sigma_next, gamma))
-
-            eps = gamma * torch.randn(shape, device = self.device)
-
-            sigma_hat = sigma + gamma * sigma
-            images_hat = images + sqrt(sigma_hat ** 2 - sigma ** 2) * eps
-
-            model_output = self.preconditioned_network_forward(images_hat, sigma_hat)
-            denoised_over_sigma = (images_hat - model_output) / sigma_hat
-
-            images_next = images_hat + (sigma_next - sigma_hat) * denoised_over_sigma
-
-            # second order correction, if not the last timestep
-
-            if sigma_next != 0:
-                model_output_next = self.preconditioned_network_forward(images_next, sigma_next)
-                denoised_prime_over_sigma = (images_next - model_output_next) / sigma_next
-                images_next = images_hat + 0.5 * (sigma_next - sigma_hat) * (denoised_over_sigma + denoised_prime_over_sigma)
-
-            images = images_next
-
-        images = images.clamp(-1., 1.)
-        return unnormalize_to_zero_to_one(images)
-
-    # training
-
-    def forward(self, images):
-        batch_size, c, h, w, device, image_size, channels = *images.shape, images.device, self.image_size, self.channels
-
-        assert h == image_size and w == image_size, f'height and width of image must be {image_size}'
-        assert c == channels, 'mismatch of image channels'
-
-        images = normalize_to_neg_one_to_one(images)
-
-        sigmas = self.noise_distribution(batch_size)
-        padded_sigmas = rearrange(sigmas, 'b -> b 1 1 1')
-
-        noise = torch.randn_like(images)
-
-        noised_images = images + padded_sigmas * noise  # alphas are 1. in the paper
-
-        denoised = self.preconditioned_network_forward(noised_images, sigmas)
-
-        losses = F.mse_loss(denoised, images, reduction = 'none')
-        losses = reduce(losses, 'b ... -> b', 'mean')
-
-        losses = losses * self.loss_weight(sigmas)
-
-        return losses.mean()

setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'denoising-diffusion-pytorch',
   packages = find_packages(),
-  version = '0.23.0',
+  version = '0.16.2',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',
@@ -16,7 +16,6 @@ setup(
   ],
   install_requires=[
     'einops',
-    'ema-pytorch',
     'pillow',
     'torch',
     'torchvision',