higher epsilon for fp16 in layernorm

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)
 
@@ -60,15 +60,16 @@ model = Unet(
 diffusion = GaussianDiffusion(
     model,
     image_size = 128,
-    timesteps = 1000,   # number of steps
-    loss_type = 'l1'    # L1 or L2
+    timesteps = 1000,           # number of steps
+    sampling_timesteps = 250,   # number of sampling timesteps (using ddim for faster inference [see citation for ddim paper])
+    loss_type = 'l1'            # L1 or L2
 ).cuda()
 
 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
@@ -159,3 +160,13 @@ $ accelerate launch train.py
     volume  = {abs/2206.00364}
 }
 ```
+
+```bibtex
+@article{Song2021DenoisingDI,
+    title   = {Denoising Diffusion Implicit Models},
+    author  = {Jiaming Song and Chenlin Meng and Stefano Ermon},
+    journal = {ArXiv},
+    year    = {2021},
+    volume  = {abs/2010.02502}
+}
+```

denoising_diffusion_pytorch/continuous_time_gaussian_diffusion.py

@@ -112,7 +112,7 @@ class learned_noise_schedule(nn.Module):
 class ContinuousTimeGaussianDiffusion(nn.Module):
     def __init__(
         self,
-        denoise_fn,
+        model,
         *,
         image_size,
         channels = 3,
@@ -126,9 +126,9 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
         p2_loss_weight_k = 1
     ):
         super().__init__()
-        assert denoise_fn.learned_sinusoidal_cond
+        assert model.learned_sinusoidal_cond
 
-        self.denoise_fn = denoise_fn
+        self.model = model
 
         # image dimensions
 
@@ -170,7 +170,7 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
 
     @property
     def device(self):
-        return next(self.denoise_fn.parameters()).device
+        return next(self.model.parameters()).device
 
     @property
     def loss_fn(self):
@@ -195,7 +195,7 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
         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)
+        pred_noise = self.model(x, batch_log_snr)
 
         if self.clip_sample_denoised:
             x_start = (x - sigma * pred_noise) / alpha
@@ -266,7 +266,7 @@ 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)
+        model_out = self.model(x, log_snr)
 
         losses = self.loss_fn(model_out, noise, reduction = 'none')
         losses = reduce(losses, 'b ... -> b', 'mean')

denoising_diffusion_pytorch/denoising_diffusion_pytorch.py

@@ -4,6 +4,7 @@ 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
@@ -22,6 +23,10 @@ from ema_pytorch import EMA
 
 from accelerate import Accelerator
 
+# constants
+
+ModelPrediction =  namedtuple('ModelPrediction', ['pred_noise', 'pred_x_start'])
+
 # helpers functions
 
 def exists(x):
@@ -53,6 +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):
@@ -81,16 +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):
@@ -203,6 +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)
@@ -210,9 +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)
@@ -222,10 +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)
@@ -383,25 +392,29 @@ def cosine_beta_schedule(timesteps, s = 0.008):
 class GaussianDiffusion(nn.Module):
     def __init__(
         self,
-        denoise_fn,
+        model,
         *,
         image_size,
         channels = 3,
         timesteps = 1000,
+        sampling_timesteps = None,
         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
+        p2_loss_weight_k = 1,
+        ddim_sampling_eta = 1.
     ):
         super().__init__()
-        assert not (type(self) == GaussianDiffusion and denoise_fn.channels != denoise_fn.out_dim)
+        assert not (type(self) == GaussianDiffusion and model.channels != model.out_dim)
 
         self.channels = channels
         self.image_size = image_size
-        self.denoise_fn = denoise_fn
+        self.model = model
         self.objective = objective
 
+        assert objective in {'pred_noise', 'pred_x0'}, 'objective must be either pred_noise (predict noise) or pred_x0 (predict image start)'
+
         if beta_schedule == 'linear':
             betas = linear_beta_schedule(timesteps)
         elif beta_schedule == 'cosine':
@@ -417,6 +430,14 @@ class GaussianDiffusion(nn.Module):
         self.num_timesteps = int(timesteps)
         self.loss_type = loss_type
 
+        # sampling related parameters
+
+        self.sampling_timesteps = default(sampling_timesteps, timesteps) # default num sampling timesteps to number of timesteps at training
+
+        assert self.sampling_timesteps <= timesteps
+        self.is_ddim_sampling = self.sampling_timesteps < timesteps
+        self.ddim_sampling_eta = ddim_sampling_eta
+
         # helper function to register buffer from float64 to float32
 
         register_buffer = lambda name, val: self.register_buffer(name, val.to(torch.float32))
@@ -457,6 +478,12 @@ class GaussianDiffusion(nn.Module):
             extract(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise
         )
 
+    def predict_noise_from_start(self, x_t, t, x0):
+        return (
+            (extract(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - x0) / \
+            extract(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape)
+        )
+
     def q_posterior(self, x_start, x_t, t):
         posterior_mean = (
             extract(self.posterior_mean_coef1, t, x_t.shape) * x_start +
@@ -466,15 +493,22 @@ 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 p_mean_variance(self, x, t, clip_denoised: bool):
-        model_output = self.denoise_fn(x, t)
+    def model_predictions(self, x, t):
+        model_output = self.model(x, t)
 
         if self.objective == 'pred_noise':
-            x_start = self.predict_start_from_noise(x, t = t, noise = model_output)
+            pred_noise = model_output
+            x_start = self.predict_start_from_noise(x, t, model_output)
+
         elif self.objective == 'pred_x0':
+            pred_noise = self.predict_noise_from_start(x, t, model_output)
             x_start = model_output
-        else:
-            raise ValueError(f'unknown objective {self.objective}')
+
+        return ModelPrediction(pred_noise, x_start)
+
+    def p_mean_variance(self, x, t, clip_denoised: bool):
+        preds = self.model_predictions(x, t)
+        x_start = preds.pred_x_start
 
         if clip_denoised:
             x_start.clamp_(-1., 1.)
@@ -483,32 +517,63 @@ class GaussianDiffusion(nn.Module):
         return model_mean, posterior_variance, posterior_log_variance
 
     @torch.no_grad()
-    def p_sample(self, x, t, clip_denoised=True):
+    def p_sample(self, x, t: int, 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 = 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
+        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)
+        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
 
     @torch.no_grad()
     def p_sample_loop(self, shape):
-        device = self.betas.device
+        batch, device = shape[0], self.betas.device
 
-        b = shape[0]
         img = torch.randn(shape, device=device)
 
-        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))
+        for t in tqdm(reversed(range(0, self.num_timesteps)), desc = 'sampling loop time step'):
+            img = self.p_sample(img, t)
+
+        img = unnormalize_to_zero_to_one(img)
+        return img
+
+    @torch.no_grad()
+    def ddim_sample(self, shape, clip_denoised = True):
+        batch, device, total_timesteps, sampling_timesteps, eta, objective = shape[0], self.betas.device, self.num_timesteps, self.sampling_timesteps, self.ddim_sampling_eta, self.objective
+
+        times = torch.linspace(0., total_timesteps, steps = sampling_timesteps + 2)[:-1]
+        times = list(reversed(times.int().tolist()))
+        time_pairs = list(zip(times[:-1], times[1:]))
+
+        img = torch.randn(shape, device = device)
+
+        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)
+
+            if clip_denoised:
+                x_start.clamp_(-1., 1.)
+
+            sigma = eta * ((1 - alpha / alpha_next) * (1 - alpha_next) / (1 - alpha)).sqrt()
+            c = ((1 - alpha_next) - sigma ** 2).sqrt()
+
+            noise = torch.randn_like(img) if time_next > 0 else 0.
+
+            img = x_start * alpha_next.sqrt() + \
+                  c * pred_noise + \
+                  sigma * noise
 
         img = unnormalize_to_zero_to_one(img)
         return img
 
     @torch.no_grad()
     def sample(self, batch_size = 16):
-        image_size = self.image_size
-        channels = self.channels
-        return self.p_sample_loop((batch_size, channels, image_size, image_size))
+        image_size, channels = self.image_size, self.channels
+        sample_fn = self.p_sample_loop if not self.is_ddim_sampling else self.ddim_sample
+        return sample_fn((batch_size, channels, image_size, image_size))
 
     @torch.no_grad()
     def interpolate(self, x1, x2, t = None, lam = 0.5):
@@ -547,8 +612,8 @@ class GaussianDiffusion(nn.Module):
         b, c, h, w = x_start.shape
         noise = default(noise, lambda: torch.randn_like(x_start))
 
-        x = self.q_sample(x_start=x_start, t=t, noise=noise)
-        model_out = self.denoise_fn(x, t)
+        x = self.q_sample(x_start = x_start, t = t, noise = noise)
+        model_out = self.model(x, t)
 
         if self.objective == 'pred_noise':
             target = noise
@@ -579,15 +644,18 @@ class Dataset(Dataset):
         folder,
         image_size,
         exts = ['jpg', 'jpeg', 'png', 'tiff'],
-        augment_horizontal_flip = False
+        augment_horizontal_flip = False,
+        convert_image_to = None
     ):
         super().__init__()
         self.folder = folder
         self.image_size = image_size
         self.paths = [p for ext in exts for p in Path(f'{folder}').glob(f'**/*.{ext}')]
 
+        maybe_convert_fn = partial(convert_image_to, convert_image_to) if exists(convert_image_to) else nn.Identity()
+
         self.transform = T.Compose([
-            T.Lambda(partial(convert_image_to, 'RGB')),
+            T.Lambda(maybe_convert_fn),
             T.Resize(image_size),
             T.RandomHorizontalFlip() if augment_horizontal_flip else nn.Identity(),
             T.CenterCrop(image_size),
@@ -617,12 +685,14 @@ 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',
         amp = False,
         fp16 = False,
-        split_batches = True
+        split_batches = True,
+        convert_image_to = None
     ):
         super().__init__()
 
@@ -647,14 +717,15 @@ class Trainer(object):
 
         # dataset and dataloader
 
-        self.ds = Dataset(folder, self.image_size, augment_horizontal_flip = augment_horizontal_flip)
+        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
 
@@ -670,15 +741,16 @@ 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_main_process:
+        if not self.accelerator.is_local_main_process:
             return
 
         data = {
             'step': self.step,
             'model': self.accelerator.get_state_dict(self.model),
+            'opt': self.opt.state_dict(),
             'ema': self.ema.state_dict(),
             'scaler': self.accelerator.scaler.state_dict() if exists(self.accelerator.scaler) else None
         }
@@ -692,6 +764,7 @@ class Trainer(object):
         model.load_state_dict(data['model'])
 
         self.step = data['step']
+        self.opt.load_state_dict(data['opt'])
         self.ema.load_state_dict(data['ema'])
 
         if exists(self.accelerator.scaler) and exists(data['scaler']):
@@ -705,14 +778,19 @@ 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/learned_gaussian_diffusion.py

@@ -1,4 +1,5 @@
 import torch
+from collections import namedtuple
 from math import pi, sqrt, log as ln
 from inspect import isfunction
 from torch import nn, einsum
@@ -10,6 +11,8 @@ from denoising_diffusion_pytorch.denoising_diffusion_pytorch import GaussianDiff
 
 NAT = 1. / ln(2)
 
+ModelPrediction = namedtuple('ModelPrediction', ['pred_noise', 'pred_x_start', 'pred_variance'])
+
 # helper functions
 
 def exists(x):
@@ -22,7 +25,7 @@ def default(val, d):
 
 # tensor helpers
 
-def log(t, eps = 1e-12):
+def log(t, eps = 1e-15):
     return torch.log(t.clamp(min = eps))
 
 def meanflat(x):
@@ -67,17 +70,31 @@ def discretized_gaussian_log_likelihood(x, *, means, log_scales, thres = 0.999):
 class LearnedGaussianDiffusion(GaussianDiffusion):
     def __init__(
         self,
-        denoise_fn,
+        model,
         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`'
+        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`'
         self.vb_loss_weight = vb_loss_weight
 
+    def model_predictions(self, x, t):
+        model_output = self.model(x, t)
+        model_output, pred_variance = model_output.chunk(2, dim = 1)
+
+        if self.objective == 'pred_noise':
+            pred_noise = model_output
+            x_start = self.predict_start_from_noise(x, t, model_output)
+
+        elif self.objective == 'pred_x0':
+            pred_noise = self.predict_noise_from_start(x, t, model_output)
+            x_start = model_output
+
+        return ModelPrediction(pred_noise, x_start, pred_variance)
+
     def p_mean_variance(self, *, x, t, clip_denoised, model_output = None):
-        model_output = default(model_output, lambda: self.denoise_fn(x, t))
+        model_output = default(model_output, lambda: self.model(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)
@@ -102,7 +119,7 @@ class LearnedGaussianDiffusion(GaussianDiffusion):
 
         # model output
 
-        model_output = self.denoise_fn(x_t, t)
+        model_output = self.model(x_t, t)
 
         # calculating kl loss for learned variance (interpolation)
 

denoising_diffusion_pytorch/weighted_objective_gaussian_diffusion.py

@@ -22,22 +22,23 @@ def default(val, d):
 class WeightedObjectiveGaussianDiffusion(GaussianDiffusion):
     def __init__(
         self,
-        denoise_fn,
+        model,
         *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'
+        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 self.is_ddim_sampling, 'ddim sampling cannot be used'
 
         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)
+        model_output = self.model(x, t)
 
         pred_noise, pred_x_start, weights = model_output.split(self.split_dims, dim = 1)
         normalized_weights = weights.softmax(dim = 1)
@@ -58,7 +59,7 @@ class WeightedObjectiveGaussianDiffusion(GaussianDiffusion):
         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)
+        model_output = self.model(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

setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'denoising-diffusion-pytorch',
   packages = find_packages(),
-  version = '0.24.1',
+  version = '0.26.5',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',