fix sampling ddpm tqdm

0.27.0

README.md

@@ -1,4 +1,4 @@
-<img src="./denoising-diffusion.png" width="500px"></img>
+<img src="./images/denoising-diffusion.png" width="500px"></img>
 
 ## Denoising Diffusion Probabilistic Model, in Pytorch
 
@@ -10,7 +10,7 @@ Youtube AI Educators - <a href="https://www.youtube.com/watch?v=W-O7AZNzbzQ">Yan
 
 <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>
+<img src="./images/sample.png" width="500px"><img>
 
 [![PyPI version](https://badge.fury.io/py/denoising-diffusion-pytorch.svg)](https://badge.fury.io/py/denoising-diffusion-pytorch)
 
@@ -69,7 +69,7 @@ trainer = Trainer(
     diffusion,
     'path/to/your/images',
     train_batch_size = 32,
-    train_lr = 1e-4,
+    train_lr = 8e-5,
     train_num_steps = 700000,         # total training steps
     gradient_accumulate_every = 2,    # gradient accumulation steps
     ema_decay = 0.995,                # exponential moving average decay
@@ -170,3 +170,14 @@ $ accelerate launch train.py
     volume  = {abs/2010.02502}
 }
 ```
+
+```bibtex
+@misc{chen2022analog,
+    title   = {Analog Bits: Generating Discrete Data using Diffusion Models with Self-Conditioning},
+    author  = {Ting Chen and Ruixiang Zhang and Geoffrey Hinton},
+    year    = {2022},
+    eprint  = {2208.04202},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.CV}
+}
+```

denoising_diffusion_pytorch/continuous_time_gaussian_diffusion.py

@@ -127,6 +127,7 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
     ):
         super().__init__()
         assert model.learned_sinusoidal_cond
+        assert not model.self_condition, 'not supported yet'
 
         self.model = model
 

denoising_diffusion_pytorch/denoising_diffusion_pytorch.py

@@ -1,23 +1,23 @@
 import math
 import copy
+from pathlib import Path
+from random import random
+from functools import partial
+from collections import namedtuple
+from multiprocessing import cpu_count
+
 import torch
 from torch import nn, einsum
 import torch.nn.functional as F
-from inspect import isfunction
-from collections import namedtuple
-from functools import partial
-
 from torch.utils.data import Dataset, DataLoader
-from multiprocessing import cpu_count
 
-from pathlib import Path
 from torch.optim import Adam
 from torchvision import transforms as T, utils
-from PIL import Image
 
 from einops import rearrange, reduce
 from einops.layers.torch import Rearrange
 
+from PIL import Image
 from tqdm.auto import tqdm
 from ema_pytorch import EMA
 
@@ -35,7 +35,7 @@ def exists(x):
 def default(val, d):
     if exists(val):
         return val
-    return d() if isfunction(d) else d
+    return d() if callable(d) else d
 
 def cycle(dl):
     while True:
@@ -58,6 +58,9 @@ def convert_image_to(img_type, image):
         return image.convert(img_type)
     return image
 
+def l2norm(t):
+    return F.normalize(t, dim = -1)
+
 # normalization functions
 
 def normalize_to_neg_one_to_one(img):
@@ -86,16 +89,15 @@ def Downsample(dim, dim_out = None):
     return nn.Conv2d(dim, default(dim_out, dim), 4, 2, 1)
 
 class LayerNorm(nn.Module):
-    def __init__(self, dim, eps = 1e-5):
+    def __init__(self, dim):
         super().__init__()
-        self.eps = eps
         self.g = nn.Parameter(torch.ones(1, dim, 1, 1))
-        self.b = nn.Parameter(torch.zeros(1, dim, 1, 1))
 
     def forward(self, x):
+        eps = 1e-5 if x.dtype == torch.float32 else 1e-3
         var = torch.var(x, dim = 1, unbiased = False, keepdim = True)
         mean = torch.mean(x, dim = 1, keepdim = True)
-        return (x - mean) / (var + self.eps).sqrt() * self.g + self.b
+        return (x - mean) * (var + eps).rsqrt() * self.g
 
 class PreNorm(nn.Module):
     def __init__(self, dim, fn):
@@ -208,6 +210,8 @@ class LinearAttention(nn.Module):
         k = k.softmax(dim = -1)
 
         q = q * self.scale
+        v = v / (h * w)
+
         context = torch.einsum('b h d n, b h e n -> b h d e', k, v)
 
         out = torch.einsum('b h d e, b h d n -> b h e n', context, q)
@@ -215,9 +219,9 @@ class LinearAttention(nn.Module):
         return self.to_out(out)
 
 class Attention(nn.Module):
-    def __init__(self, dim, heads = 4, dim_head = 32):
+    def __init__(self, dim, heads = 4, dim_head = 32, scale = 16):
         super().__init__()
-        self.scale = dim_head ** -0.5
+        self.scale = scale
         self.heads = heads
         hidden_dim = dim_head * heads
         self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias = False)
@@ -227,10 +231,10 @@ class Attention(nn.Module):
         b, c, h, w = x.shape
         qkv = self.to_qkv(x).chunk(3, dim = 1)
         q, k, v = map(lambda t: rearrange(t, 'b (h c) x y -> b h c (x y)', h = self.heads), qkv)
-        q = q * self.scale
 
-        sim = einsum('b h d i, b h d j -> b h i j', q, k)
-        sim = sim - sim.amax(dim = -1, keepdim = True).detach()
+        q, k = map(l2norm, (q, k))
+
+        sim = einsum('b h d i, b h d j -> b h i j', q, k) * self.scale
         attn = sim.softmax(dim = -1)
 
         out = einsum('b h i j, b h d j -> b h i d', attn, v)
@@ -247,6 +251,7 @@ class Unet(nn.Module):
         out_dim = None,
         dim_mults=(1, 2, 4, 8),
         channels = 3,
+        self_condition = False,
         resnet_block_groups = 8,
         learned_variance = False,
         learned_sinusoidal_cond = False,
@@ -257,9 +262,11 @@ class Unet(nn.Module):
         # determine dimensions
 
         self.channels = channels
+        self.self_condition = self_condition
+        input_channels = channels * (2 if self_condition else 1)
 
         init_dim = default(init_dim, dim)
-        self.init_conv = nn.Conv2d(channels, init_dim, 7, padding = 3)
+        self.init_conv = nn.Conv2d(input_channels, init_dim, 7, padding = 3)
 
         dims = [init_dim, *map(lambda m: dim * m, dim_mults)]
         in_out = list(zip(dims[:-1], dims[1:]))
@@ -323,7 +330,11 @@ class Unet(nn.Module):
         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):
+    def forward(self, x, time, x_self_cond = None):
+        if self.self_condition:
+            x_self_cond = default(x_self_cond, lambda: torch.zeros_like(x))
+            x = torch.cat((x_self_cond, x), dim = 1)
+
         x = self.init_conv(x)
         r = x.clone()
 
@@ -391,7 +402,6 @@ class GaussianDiffusion(nn.Module):
         model,
         *,
         image_size,
-        channels = 3,
         timesteps = 1000,
         sampling_timesteps = None,
         loss_type = 'l1',
@@ -404,9 +414,12 @@ class GaussianDiffusion(nn.Module):
         super().__init__()
         assert not (type(self) == GaussianDiffusion and model.channels != model.out_dim)
 
-        self.channels = channels
-        self.image_size = image_size
         self.model = model
+        self.channels = self.model.channels
+        self.self_condition = self.model.self_condition
+
+        self.image_size = image_size
+
         self.objective = objective
 
         assert objective in {'pred_noise', 'pred_x0'}, 'objective must be either pred_noise (predict noise) or pred_x0 (predict image start)'
@@ -476,7 +489,7 @@ class GaussianDiffusion(nn.Module):
 
     def predict_noise_from_start(self, x_t, t, x0):
         return (
-            (x0 - extract(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t) / \
+            (extract(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - x0) / \
             extract(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape)
         )
 
@@ -489,8 +502,8 @@ class GaussianDiffusion(nn.Module):
         posterior_log_variance_clipped = extract(self.posterior_log_variance_clipped, t, x_t.shape)
         return posterior_mean, posterior_variance, posterior_log_variance_clipped
 
-    def model_predictions(self, x, t):
-        model_output = self.model(x, t)
+    def model_predictions(self, x, t, x_self_cond = None):
+        model_output = self.model(x, t, x_self_cond)
 
         if self.objective == 'pred_noise':
             pred_noise = model_output
@@ -502,23 +515,24 @@ class GaussianDiffusion(nn.Module):
 
         return ModelPrediction(pred_noise, x_start)
 
-    def p_mean_variance(self, x, t, clip_denoised: bool):
-        preds = self.model_predictions(x, t)
+    def p_mean_variance(self, x, t, x_self_cond = None, clip_denoised = True):
+        preds = self.model_predictions(x, t, x_self_cond)
         x_start = preds.pred_x_start
 
         if clip_denoised:
             x_start.clamp_(-1., 1.)
 
         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
+        return model_mean, posterior_variance, posterior_log_variance, x_start
 
     @torch.no_grad()
-    def p_sample(self, x, t: int, clip_denoised = True):
+    def p_sample(self, x, t: int, x_self_cond = None, clip_denoised = True):
         b, *_, device = *x.shape, x.device
         batched_times = torch.full((x.shape[0],), t, device = x.device, dtype = torch.long)
-        model_mean, _, model_log_variance = self.p_mean_variance(x = x, t = batched_times, clip_denoised = clip_denoised)
+        model_mean, _, model_log_variance, x_start = self.p_mean_variance(x = x, t = batched_times, x_self_cond = x_self_cond, clip_denoised = clip_denoised)
         noise = torch.randn_like(x) if t > 0 else 0. # no noise if t == 0
-        return model_mean + (0.5 * model_log_variance).exp() * noise
+        pred_img = model_mean + (0.5 * model_log_variance).exp() * noise
+        return pred_img, x_start
 
     @torch.no_grad()
     def p_sample_loop(self, shape):
@@ -526,8 +540,11 @@ class GaussianDiffusion(nn.Module):
 
         img = torch.randn(shape, device=device)
 
-        for t in tqdm(reversed(range(0, self.num_timesteps)), desc = 'sampling loop time step'):
-            img = self.p_sample(img, t)
+        x_start = None
+
+        for t in tqdm(reversed(range(0, self.num_timesteps)), desc = 'sampling loop time step', total = self.num_timesteps):
+            self_cond = x_start if self.self_condition else None
+            img, x_start = self.p_sample(img, t, self_cond)
 
         img = unnormalize_to_zero_to_one(img)
         return img
@@ -542,13 +559,17 @@ class GaussianDiffusion(nn.Module):
 
         img = torch.randn(shape, device = device)
 
+        x_start = None
+
         for time, time_next in tqdm(time_pairs, desc = 'sampling loop time step'):
             alpha = self.alphas_cumprod_prev[time]
             alpha_next = self.alphas_cumprod_prev[time_next]
 
             time_cond = torch.full((batch,), time, device = device, dtype = torch.long)
 
-            pred_noise, x_start, *_ = self.model_predictions(img, time_cond)
+            self_cond = x_start if self.self_condition else None
+
+            pred_noise, x_start, *_ = self.model_predictions(img, time_cond, self_cond)
 
             if clip_denoised:
                 x_start.clamp_(-1., 1.)
@@ -578,11 +599,11 @@ class GaussianDiffusion(nn.Module):
 
         assert x1.shape == x2.shape
 
-        t_batched = torch.stack([torch.tensor(t, device=device)] * b)
-        xt1, xt2 = map(lambda x: self.q_sample(x, t=t_batched), (x1, x2))
+        t_batched = torch.stack([torch.tensor(t, device = device)] * b)
+        xt1, xt2 = map(lambda x: self.q_sample(x, t = t_batched), (x1, x2))
 
         img = (1 - lam) * xt1 + lam * xt2
-        for i in tqdm(reversed(range(0, t)), desc='interpolation sample time step', total=t):
+        for i in tqdm(reversed(range(0, t)), desc = 'interpolation sample time step', total = t):
             img = self.p_sample(img, torch.full((b,), i, device=device, dtype=torch.long))
 
         return img
@@ -608,8 +629,23 @@ class GaussianDiffusion(nn.Module):
         b, c, h, w = x_start.shape
         noise = default(noise, lambda: torch.randn_like(x_start))
 
+        # noise sample
+
         x = self.q_sample(x_start = x_start, t = t, noise = noise)
-        model_out = self.model(x, t)
+
+        # if doing self-conditioning, 50% of the time, predict x_start from current set of times
+        # and condition with unet with that
+        # this technique will slow down training by 25%, but seems to lower FID significantly
+
+        x_self_cond = None
+        if self.self_condition and random() < 0.5:
+            with torch.no_grad():
+                x_self_cond = self.model_predictions(x, t).pred_x_start
+                x_self_cond.detach_()
+
+        # predict and take gradient step
+
+        model_out = self.model(x, t, x_self_cond)
 
         if self.objective == 'pred_noise':
             target = noise
@@ -681,6 +717,7 @@ class Trainer(object):
         train_num_steps = 100000,
         ema_update_every = 10,
         ema_decay = 0.995,
+        adam_betas = (0.9, 0.99),
         save_and_sample_every = 1000,
         num_samples = 25,
         results_folder = './results',
@@ -715,11 +752,12 @@ class Trainer(object):
         self.ds = Dataset(folder, self.image_size, augment_horizontal_flip = augment_horizontal_flip, convert_image_to = convert_image_to)
         dl = DataLoader(self.ds, batch_size = train_batch_size, shuffle = True, pin_memory = True, num_workers = cpu_count())
 
+        dl = self.accelerator.prepare(dl)
         self.dl = cycle(dl)
 
         # optimizer
 
-        self.opt = Adam(diffusion_model.parameters(), lr = train_lr)
+        self.opt = Adam(diffusion_model.parameters(), lr = train_lr, betas = adam_betas)
 
         # for logging results in a folder periodically
 
@@ -735,7 +773,7 @@ class Trainer(object):
 
         # prepare model, dataloader, optimizer with accelerator
 
-        self.model, self.dl, self.opt = self.accelerator.prepare(self.model, self.dl, self.opt)
+        self.model, self.opt = self.accelerator.prepare(self.model, self.opt)
 
     def save(self, milestone):
         if not self.accelerator.is_local_main_process:
@@ -772,14 +810,19 @@ class Trainer(object):
 
             while self.step < self.train_num_steps:
 
+                total_loss = 0.
+
                 for _ in range(self.gradient_accumulate_every):
                     data = next(self.dl).to(device)
 
                     with self.accelerator.autocast():
                         loss = self.model(data)
-                        self.accelerator.backward(loss / self.gradient_accumulate_every)
+                        loss = loss / self.gradient_accumulate_every
+                        total_loss += loss.item()
 
-                pbar.set_description(f'loss: {loss.item():.4f}')
+                    self.accelerator.backward(loss)
+
+                pbar.set_description(f'loss: {total_loss:.4f}')
 
                 accelerator.wait_for_everyone()
 

denoising_diffusion_pytorch/elucidated_diffusion.py

@@ -52,6 +52,7 @@ class ElucidatedDiffusion(nn.Module):
     ):
         super().__init__()
         assert net.learned_sinusoidal_cond
+        assert not net.self_condition, 'not supported yet'
 
         self.net = net
 

denoising_diffusion_pytorch/learned_gaussian_diffusion.py

@@ -77,6 +77,8 @@ class LearnedGaussianDiffusion(GaussianDiffusion):
     ):
         super().__init__(model, *args, **kwargs)
         assert model.out_dim == (model.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`'
+        assert not model.self_condition, 'not supported yet'
+
         self.vb_loss_weight = vb_loss_weight
 
     def model_predictions(self, x, t):

denoising_diffusion_pytorch/weighted_objective_gaussian_diffusion.py

@@ -31,6 +31,7 @@ class WeightedObjectiveGaussianDiffusion(GaussianDiffusion):
         super().__init__(model, *args, **kwargs)
         channels = model.channels
         assert model.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'
+        assert not model.self_condition, 'not supported yet'
         assert not self.is_ddim_sampling, 'ddim sampling cannot be used'
 
         self.split_dims = (channels, channels, 2)

setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'denoising-diffusion-pytorch',
   packages = find_packages(),
-  version = '0.25.2',
+  version = '0.27.1',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',