conditioning on final resnet block

README.md

@@ -6,6 +6,10 @@ 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)
@@ -119,3 +123,13 @@ Samples and model checkpoints will be logged to `./results` periodically
     url         = {https://openreview.net/forum?id=2LdBqxc1Yv}
 }
 ```
+
+```bibtex
+@article{Choi2022PerceptionPT,
+    title   = {Perception Prioritized Training of Diffusion Models},
+    author  = {Jooyoung Choi and Jungbeom Lee and Chaehun Shin and Sungwon Kim and Hyunwoo J. Kim and Sung-Hoon Yoon},
+    journal = {ArXiv},
+    year    = {2022},
+    volume  = {abs/2204.00227}
+}
+```

denoising_diffusion_pytorch/continuous_time_gaussian_diffusion.py

@@ -5,7 +5,8 @@ import torch.nn.functional as F
 from torch.special import expm1
 
 from tqdm import tqdm
-from einops import rearrange, repeat
+from einops import rearrange, repeat, reduce
+from einops.layers.torch import Rearrange
 
 # helpers
 
@@ -33,6 +34,24 @@ def right_pad_dims_to(x, t):
         return t
     return t.view(*t.shape, *((1,) * padding_dims))
 
+# neural net helpers
+
+class Residual(nn.Module):
+    def __init__(self, fn):
+        super().__init__()
+        self.fn = fn
+
+    def forward(self, x):
+        return x + self.fn(x)
+
+class MonotonicLinear(nn.Module):
+    def __init__(self, *args, **kwargs):
+        super().__init__()
+        self.net = nn.Linear(*args, **kwargs)
+
+    def forward(self, x):
+        return F.linear(x, self.net.weight.abs(), self.net.bias.abs())
+
 # continuous schedules
 
 # equations are taken from https://openreview.net/attachment?id=2LdBqxc1Yv&name=supplementary_material
@@ -40,17 +59,54 @@ def right_pad_dims_to(x, t):
 
 # log(snr) that approximates the original linear schedule
 
-def beta_linear_log_snr(t):
+def log(t, eps = 1e-20):
-    return -torch.log(expm1(1e-4 + 10 * (t ** 2)))
+    return torch.log(t.clamp(min = eps))
 
-def alpha_cosine_log_snr(t):
+def beta_linear_log_snr(t):
-    raise NotImplementedError
+    return -log(expm1(1e-4 + 10 * (t ** 2)))
+
+def alpha_cosine_log_snr(t, s = 0.008):
+    return -log((torch.cos((t + s) / (1 + s) * torch.pi * 0.5) ** -2) - 1, eps = 1e-5)
 
 class learned_noise_schedule(nn.Module):
-    def __init__(self):
+    """ described in section H and then I.2 of the supplementary material for variational ddpm paper """
+
+    def __init__(
+        self,
+        *,
+        log_snr_max,
+        log_snr_min,
+        hidden_dim = 1024,
+        frac_gradient = 1.
+    ):
         super().__init__()
-        raise NotImplementedError
+        self.slope = log_snr_min - log_snr_max
-        # learned noise schedule, using learned monotonic MLP (weights kept positive) in the paper
+        self.intercept = log_snr_max
+
+        self.net = nn.Sequential(
+            Rearrange('... -> ... 1'),
+            MonotonicLinear(1, 1),
+            Residual(nn.Sequential(
+                MonotonicLinear(1, hidden_dim),
+                nn.Sigmoid(),
+                MonotonicLinear(hidden_dim, 1)
+            )),
+            Rearrange('... 1 -> ...'),
+        )
+
+        self.frac_gradient = frac_gradient
+
+    def forward(self, x):
+        frac_gradient = self.frac_gradient
+        device = x.device
+
+        out_zero = self.net(torch.zeros_like(x))
+        out_one =  self.net(torch.ones_like(x))
+
+        x = self.net(x)
+
+        normed = self.slope * ((x - out_zero) / (out_one - out_zero)) + self.intercept
+        return normed * frac_gradient + normed.detach() * (1 - frac_gradient)
 
 class ContinuousTimeGaussianDiffusion(nn.Module):
     def __init__(
@@ -61,10 +117,15 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
         channels = 3,
         loss_type = 'l1',
         noise_schedule = 'linear',
-        num_sample_steps = 500
+        num_sample_steps = 500,
+        clip_sample_denoised = True,
+        learned_schedule_net_hidden_dim = 1024,
+        learned_noise_schedule_frac_gradient = 1.,   # between 0 and 1, determines what percentage of gradients go back, so one can update the learned noise schedule more slowly
+        p2_loss_weight_gamma = 0.,                   # p2 loss weight, from https://arxiv.org/abs/2204.00227 - 0 is equivalent to weight of 1 across time
+        p2_loss_weight_k = 1
     ):
         super().__init__()
-        assert not denoise_fn.sinusoidal_cond_mlp
+        assert denoise_fn.learned_sinusoidal_cond
 
         self.denoise_fn = denoise_fn
 
@@ -79,12 +140,32 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
 
         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):
@@ -103,9 +184,6 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
         # reviewer found an error in the equation in the paper (missing sigma)
         # following - https://openreview.net/forum?id=2LdBqxc1Yv&noteId=rIQgH0zKsRt
 
-        # todo - derive x_start from the posterior mean and do dynamic thresholding
-        # assumed that is what is going on in Imagen
-
         log_snr = self.log_snr(time)
         log_snr_next = self.log_snr(time_next)
         c = -expm1(log_snr - log_snr_next)
@@ -113,10 +191,21 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
         squared_alpha, squared_alpha_next = log_snr.sigmoid(), log_snr_next.sigmoid()
         squared_sigma, squared_sigma_next = (-log_snr).sigmoid(), (-log_snr_next).sigmoid()
 
+        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)
 
-        model_mean = sqrt(squared_alpha_next / squared_alpha) * (x - c * sqrt(squared_sigma) * pred_noise)
+        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
@@ -176,9 +265,17 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
         noise = default(noise, lambda: torch.randn_like(x_start))
 
         x, log_snr = self.q_sample(x_start = x_start, times = times, noise = noise)
-
         model_out = self.denoise_fn(x, log_snr)
-        return self.loss_fn(model_out, noise)
+
+        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

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 +16,7 @@ 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
 
 # helpers functions
@@ -72,20 +73,6 @@ 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)
 
@@ -114,6 +101,39 @@ 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):
@@ -157,6 +177,7 @@ 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):
@@ -212,18 +233,6 @@ class Attention(nn.Module):
 
 # model
 
-def MLP(dim_in, dim_hidden):
-    return nn.Sequential(
-        Rearrange('... -> ... 1'),
-        nn.Linear(1, dim_hidden),
-        nn.GELU(),
-        nn.LayerNorm(dim_hidden),
-        nn.Linear(dim_hidden, dim_hidden),
-        nn.GELU(),
-        nn.LayerNorm(dim_hidden),
-        nn.Linear(dim_hidden, dim_hidden)
-    )
-
 class Unet(nn.Module):
     def __init__(
         self,
@@ -234,7 +243,8 @@ class Unet(nn.Module):
         channels = 3,
         resnet_block_groups = 8,
         learned_variance = False,
-        sinusoidal_cond_mlp = True
+        learned_sinusoidal_cond = False,
+        learned_sinusoidal_dim = 16
     ):
         super().__init__()
 
@@ -242,7 +252,7 @@ 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)]
@@ -254,17 +264,21 @@ class Unet(nn.Module):
 
         time_dim = dim * 4
 
-        self.sinusoidal_cond_mlp = sinusoidal_cond_mlp
+        self.learned_sinusoidal_cond = learned_sinusoidal_cond
 
-        if sinusoidal_cond_mlp:
+        if learned_sinusoidal_cond:
-            self.time_mlp = nn.Sequential(
+            sinu_pos_emb = LearnedSinusoidalPosEmb(learned_sinusoidal_dim)
-                SinusoidalPosEmb(dim),
+            fourier_dim = learned_sinusoidal_dim + 1
-                nn.Linear(dim, time_dim),
-                nn.GELU(),
-                nn.Linear(time_dim, time_dim)
-            )
         else:
-            self.time_mlp = MLP(1, time_dim)
+            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
 
@@ -287,8 +301,8 @@ 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:])):
+        for ind, (dim_in, dim_out) in enumerate(reversed(in_out)):
-            is_last = ind >= (num_resolutions - 1)
+            is_last = ind == (len(in_out) - 1)
 
             self.ups.append(nn.ModuleList([
                 block_klass(dim_out * 2, dim_in, time_emb_dim = time_dim),
@@ -300,13 +314,13 @@ class Unet(nn.Module):
         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(
+        self.final_res_block = block_klass(dim * 2, dim, time_emb_dim = time_dim)
-            block_klass(dim, dim),
+        self.final_conv = nn.Conv2d(dim, self.out_dim, 1)
-            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)
 
         h = []
@@ -323,12 +337,15 @@ 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
@@ -366,7 +383,9 @@ 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)
@@ -421,6 +440,10 @@ 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 -
@@ -527,8 +550,11 @@ class GaussianDiffusion(nn.Module):
         else:
             raise ValueError(f'unknown objective {self.objective}')
 
-        loss = self.loss_fn(model_out, target)
+        loss = self.loss_fn(model_out, target, reduction = 'none')
-        return loss
+        loss = reduce(loss, 'b ... -> b (...)', 'mean')
+
+        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
@@ -541,7 +567,7 @@ class GaussianDiffusion(nn.Module):
 # 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
@@ -549,7 +575,7 @@ 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()
         ])
@@ -580,7 +606,8 @@ class Trainer(object):
         step_start_ema = 2000,
         update_ema_every = 10,
         save_and_sample_every = 1000,
-        results_folder = './results'
+        results_folder = './results',
+        augment_horizontal_flip = True
     ):
         super().__init__()
         self.model = diffusion_model
@@ -596,8 +623,8 @@ class Trainer(object):
         self.gradient_accumulate_every = gradient_accumulate_every
         self.train_num_steps = train_num_steps
 
-        self.ds = Dataset(folder, image_size)
+        self.ds = Dataset(folder, 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))
+        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

setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'denoising-diffusion-pytorch',
   packages = find_packages(),
-  version = '0.16.7',
+  version = '0.20.1',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',