optimizer should be saved and loaded

Solve issue #26

README.md

@@ -80,6 +80,22 @@ trainer.train()
 
 Samples and model checkpoints will be logged to `./results` periodically
 
+## Multi-GPU Training
+
+The `Trainer` class is now equipped with <a href="https://huggingface.co/docs/accelerate/accelerator">🤗 Accelerator</a>. You can easily do multi-gpu training in two steps using their `accelerate` CLI
+
+At the project root directory, where the training script is, run
+
+```python
+$ accelerate config
+```
+
+Then, in the same directory
+
+```python
+$ accelerate launch train.py
+```
+
 ## Citations
 
 ```bibtex
@@ -133,3 +149,13 @@ Samples and model checkpoints will be logged to `./results` periodically
     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

@@ -3,3 +3,4 @@ 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

@@ -1,3 +1,4 @@
+import math
 import torch
 from torch import sqrt
 from torch import nn, einsum
@@ -66,7 +67,7 @@ def beta_linear_log_snr(t):
     return -log(expm1(1e-4 + 10 * (t ** 2)))
 
 def alpha_cosine_log_snr(t, s = 0.008):
-    return -log((torch.cos((t + s) / (1 + s) * torch.pi * 0.5) ** -2) - 1, eps = 1e-5)
+    return -log((torch.cos((t + s) / (1 + s) * math.pi * 0.5) ** -2) - 1, eps = 1e-5)
 
 class learned_noise_schedule(nn.Module):
     """ described in section H and then I.2 of the supplementary material for variational ddpm paper """

denoising_diffusion_pytorch/denoising_diffusion_pytorch.py

@@ -6,21 +6,22 @@ import torch.nn.functional as F
 from inspect import isfunction
 from functools import partial
 
-from torch.utils import data
+from torch.utils.data import Dataset, DataLoader
 from multiprocessing import cpu_count
-from torch.cuda.amp import autocast, GradScaler
 
 from pathlib import Path
 from torch.optim import Adam
-from torchvision import transforms, utils
+from torchvision import transforms as T, utils
 from PIL import Image
 
-from tqdm import tqdm
 from einops import rearrange, reduce
 from einops.layers.torch import Rearrange
 
+from tqdm.auto import tqdm
 from ema_pytorch import EMA
 
+from accelerate import Accelerator
+
 # helpers functions
 
 def exists(x):
@@ -36,6 +37,9 @@ def cycle(dl):
         for data in dl:
             yield data
 
+def has_int_squareroot(num):
+    return (math.sqrt(num) ** 2) == num
+
 def num_to_groups(num, divisor):
     groups = num // divisor
     remainder = num % divisor
@@ -44,6 +48,13 @@ def num_to_groups(num, divisor):
         arr.append(remainder)
     return arr
 
+def convert_image_to(img_type, image):
+    if image.mode != img_type:
+        return image.convert(img_type)
+    return image
+
+# normalization functions
+
 def normalize_to_neg_one_to_one(img):
     return img * 2 - 1
 
@@ -60,11 +71,14 @@ class Residual(nn.Module):
     def forward(self, x, *args, **kwargs):
         return self.fn(x, *args, **kwargs) + x
 
-def Upsample(dim):
-    return nn.ConvTranspose2d(dim, dim, 4, 2, 1)
+def Upsample(dim, dim_out = None):
+    return nn.Sequential(
+        nn.Upsample(scale_factor = 2, mode = 'nearest'),
+        nn.Conv2d(dim, default(dim_out, dim), 3, padding = 1)
+    )
 
-def Downsample(dim):
-    return nn.Conv2d(dim, dim, 4, 2, 1)
+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):
@@ -277,10 +291,10 @@ class Unet(nn.Module):
             is_last = ind >= (num_resolutions - 1)
 
             self.downs.append(nn.ModuleList([
-                block_klass(dim_in, dim_out, time_emb_dim = time_dim),
-                block_klass(dim_out, dim_out, time_emb_dim = time_dim),
-                Residual(PreNorm(dim_out, LinearAttention(dim_out))),
-                Downsample(dim_out) if not is_last else nn.Identity()
+                block_klass(dim_in, dim_in, time_emb_dim = time_dim),
+                block_klass(dim_in, dim_in, time_emb_dim = time_dim),
+                Residual(PreNorm(dim_in, LinearAttention(dim_in))),
+                Downsample(dim_in, dim_out) if not is_last else nn.Conv2d(dim_in, dim_out, 3, padding = 1)
             ]))
 
         mid_dim = dims[-1]
@@ -292,10 +306,10 @@ class Unet(nn.Module):
             is_last = ind == (len(in_out) - 1)
 
             self.ups.append(nn.ModuleList([
-                block_klass(dim_out * 2, dim_in, time_emb_dim = time_dim),
-                block_klass(dim_in, dim_in, time_emb_dim = time_dim),
-                Residual(PreNorm(dim_in, LinearAttention(dim_in))),
-                Upsample(dim_in) if not is_last else nn.Identity()
+                block_klass(dim_out + dim_in, dim_out, time_emb_dim = time_dim),
+                block_klass(dim_out + dim_in, dim_out, time_emb_dim = time_dim),
+                Residual(PreNorm(dim_out, LinearAttention(dim_out))),
+                Upsample(dim_out, dim_in) if not is_last else  nn.Conv2d(dim_out, dim_in, 3, padding = 1)
             ]))
 
         default_out_dim = channels * (1 if not learned_variance else 2)
@@ -314,9 +328,12 @@ class Unet(nn.Module):
 
         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)
@@ -326,8 +343,11 @@ class Unet(nn.Module):
         for block1, block2, attn, upsample in self.ups:
             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)
@@ -355,7 +375,7 @@ 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) * torch.pi * 0.5) ** 2
+    alphas_cumprod = torch.cos(((x / timesteps) + s) / (1 + s) * math.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)
@@ -553,18 +573,28 @@ class GaussianDiffusion(nn.Module):
 
 # dataset classes
 
-class Dataset(data.Dataset):
-    def __init__(self, folder, image_size, exts = ['jpg', 'jpeg', 'png'], augment_horizontal_flip = False):
+class Dataset(Dataset):
+    def __init__(
+        self,
+        folder,
+        image_size,
+        exts = ['jpg', 'jpeg', 'png', 'tiff'],
+        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}')]
 
-        self.transform = transforms.Compose([
-            transforms.Resize(image_size),
-            transforms.RandomHorizontalFlip() if augment_horizontal_flip else nn.Identity(),
-            transforms.CenterCrop(image_size),
-            transforms.ToTensor()
+        maybe_convert_fn = partial(convert_image_to, convert_image_to) if exists(convert_image_to) else nn.Identity()
+
+        self.transform = T.Compose([
+            T.Lambda(maybe_convert_fn),
+            T.Resize(image_size),
+            T.RandomHorizontalFlip() if augment_horizontal_flip else nn.Identity(),
+            T.CenterCrop(image_size),
+            T.ToTensor()
         ])
 
     def __len__(self):
@@ -583,93 +613,137 @@ class Trainer(object):
         diffusion_model,
         folder,
         *,
-        ema_decay = 0.995,
-        train_batch_size = 32,
+        train_batch_size = 16,
+        gradient_accumulate_every = 1,
+        augment_horizontal_flip = True,
         train_lr = 1e-4,
         train_num_steps = 100000,
-        gradient_accumulate_every = 2,
-        amp = False,
-        step_start_ema = 2000,
-        update_ema_every = 10,
+        ema_update_every = 10,
+        ema_decay = 0.995,
         save_and_sample_every = 1000,
+        num_samples = 25,
         results_folder = './results',
-        augment_horizontal_flip = True
+        amp = False,
+        fp16 = False,
+        split_batches = True,
+        convert_image_to = None
     ):
         super().__init__()
-        self.image_size = diffusion_model.image_size
+
+        self.accelerator = Accelerator(
+            split_batches = split_batches,
+            mixed_precision = 'fp16' if fp16 else 'no'
+        )
+
+        self.accelerator.native_amp = amp
 
         self.model = diffusion_model
-        self.ema = EMA(diffusion_model, beta = ema_decay)
-        self.update_ema_every = update_ema_every
 
-        self.step_start_ema = step_start_ema
+        assert has_int_squareroot(num_samples), 'number of samples must have an integer square root'
+        self.num_samples = num_samples
         self.save_and_sample_every = save_and_sample_every
 
         self.batch_size = train_batch_size
-        self.image_size = diffusion_model.image_size
         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.dl = cycle(data.DataLoader(self.ds, batch_size = train_batch_size, shuffle = True, pin_memory = True, num_workers = cpu_count()))
+        self.train_num_steps = train_num_steps
+        self.image_size = diffusion_model.image_size
+
+        # dataset and dataloader
+
+        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())
+
+        self.dl = cycle(dl)
+
+        # optimizer
+
         self.opt = Adam(diffusion_model.parameters(), lr = train_lr)
 
+        # for logging results in a folder periodically
+
+        if self.accelerator.is_main_process:
+            self.ema = EMA(diffusion_model, beta = ema_decay, update_every = ema_update_every)
+
+            self.results_folder = Path(results_folder)
+            self.results_folder.mkdir(exist_ok = True)
+
+        # step counter state
+
         self.step = 0
 
-        self.amp = amp
-        self.scaler = GradScaler(enabled = amp)
+        # prepare model, dataloader, optimizer with accelerator
 
-        self.results_folder = Path(results_folder)
-        self.results_folder.mkdir(exist_ok = True)
+        self.model, self.dl, self.opt = self.accelerator.prepare(self.model, self.dl, self.opt)
 
     def save(self, milestone):
+        if not self.accelerator.is_main_process:
+            return
+
         data = {
             'step': self.step,
-            'model': self.model.state_dict(),
+            'model': self.accelerator.get_state_dict(self.model),
+            'opt': self.opt.state_dict(),
             'ema': self.ema.state_dict(),
-            'scaler': self.scaler.state_dict()
+            'scaler': self.accelerator.scaler.state_dict() if exists(self.accelerator.scaler) else None
         }
+
         torch.save(data, str(self.results_folder / f'model-{milestone}.pt'))
 
     def load(self, milestone):
         data = torch.load(str(self.results_folder / f'model-{milestone}.pt'))
 
+        model = self.accelerator.unwrap_model(self.model)
+        model.load_state_dict(data['model'])
+
         self.step = data['step']
-        self.model.load_state_dict(data['model'])
+        self.opt.load_state_dict(data['opt'])
         self.ema.load_state_dict(data['ema'])
-        self.scaler.load_state_dict(data['scaler'])
+
+        if exists(self.accelerator.scaler) and exists(data['scaler']):
+            self.accelerator.scaler.load_state_dict(data['scaler'])
 
     def train(self):
-        with tqdm(initial = self.step, total = self.train_num_steps) as pbar:
+        accelerator = self.accelerator
+        device = accelerator.device
+
+        with tqdm(initial = self.step, total = self.train_num_steps, disable = not accelerator.is_main_process) as pbar:
 
             while self.step < self.train_num_steps:
-                for i in range(self.gradient_accumulate_every):
-                    data = next(self.dl).cuda()
 
-                    with autocast(enabled = self.amp):
+                for _ in range(self.gradient_accumulate_every):
+                    data = next(self.dl).to(device)
+
+                    with self.accelerator.autocast():
                         loss = self.model(data)
-                        self.scaler.scale(loss / self.gradient_accumulate_every).backward()
+                        self.accelerator.backward(loss / self.gradient_accumulate_every)
 
-                    pbar.set_description(f'loss: {loss.item():.4f}')
+                pbar.set_description(f'loss: {loss.item():.4f}')
 
-                self.scaler.step(self.opt)
-                self.scaler.update()
+                accelerator.wait_for_everyone()
+
+                self.opt.step()
                 self.opt.zero_grad()
 
-                self.ema.update()
+                accelerator.wait_for_everyone()
 
-                if self.step != 0 and self.step % self.save_and_sample_every == 0:
-                    self.ema.ema_model.eval()
-                    with torch.no_grad():
-                        milestone = self.step // self.save_and_sample_every
-                        batches = num_to_groups(36, self.batch_size)
-                        all_images_list = list(map(lambda n: self.ema.ema_model.sample(batch_size=n), batches))
+                if accelerator.is_main_process:
+                    self.ema.to(device)
+                    self.ema.update()
 
-                    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)
+                    if self.step != 0 and self.step % self.save_and_sample_every == 0:
+                        self.ema.ema_model.eval()
+
+                        with torch.no_grad():
+                            milestone = self.step // self.save_and_sample_every
+                            batches = num_to_groups(self.num_samples, self.batch_size)
+                            all_images_list = list(map(lambda n: self.ema.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 = int(math.sqrt(self.num_samples)))
+                        self.save(milestone)
 
                 self.step += 1
                 pbar.update(1)
 
-        print('training complete')
+        accelerator.print('training complete')

denoising_diffusion_pytorch/elucidated_diffusion.py

@@ -0,0 +1,220 @@
+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
+
+    # preconditioned network output
+    # equation (7) in the paper
+
+    def preconditioned_network_forward(self, noised_images, sigma, clamp = False):
+        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)
+        )
+
+        out = self.c_skip(padded_sigma) * noised_images +  self.c_out(padded_sigma) * net_out
+
+        if clamp:
+            out = out.clamp(-1., 1.)
+
+        return out
+
+    # sampling
+
+    # 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
+
+    @torch.no_grad()
+    def sample(self, batch_size = 16, num_sample_steps = None, clamp = True):
+        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 = self.S_noise * torch.randn(shape, device = self.device) # stochastic sampling
+
+            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, clamp = clamp)
+            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, clamp = clamp)
+                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 loss_weight(self, sigma):
+        return (sigma ** 2 + self.sigma_data ** 2) * (sigma * self.sigma_data) ** -2
+
+    def noise_distribution(self, batch_size):
+        return (self.P_mean + self.P_std * torch.randn((batch_size,), device = self.device)).exp()
+
+    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()

denoising_diffusion_pytorch/learned_gaussian_diffusion.py

@@ -22,7 +22,7 @@ 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):

setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'denoising-diffusion-pytorch',
   packages = find_packages(),
-  version = '0.21.0',
+  version = '0.24.3',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',
@@ -15,6 +15,7 @@ setup(
     'generative models'
   ],
   install_requires=[
+    'accelerate',
     'einops',
     'ema-pytorch',
     'pillow',