use pre-layernorm with linear attention, and also allow for turning off time embedding

.gitignore

@@ -1,3 +1,6 @@
+# Generation results
+results/
+
 # Byte-compiled / optimized / DLL files
 __pycache__/
 *.py[cod]

README.md

@@ -6,6 +6,8 @@ Implementation of <a href="https://arxiv.org/abs/2006.11239">Denoising Diffusion
 
 <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)
+
 ## Install
 
 ```bash
@@ -25,10 +27,9 @@ model = Unet(
 
 diffusion = GaussianDiffusion(
     model,
-    beta_start = 0.0001,
-    beta_end = 0.02,
-    num_diffusion_timesteps = 1000,   # number of steps
-    loss_type = 'l1'                  # L1 or L2 (wavegrad paper claims l1 is better?)
+    image_size = 128,
+    timesteps = 1000,   # number of steps
+    loss_type = 'l1'    # L1 or L2
 )
 
 training_images = torch.randn(8, 3, 128, 128)
@@ -36,7 +37,7 @@ loss = diffusion(training_images)
 loss.backward()
 # after a lot of training
 
-sampled_images = diffusion.sample(128, batch_size = 4)
+sampled_images = diffusion.sample(batch_size = 4)
 sampled_images.shape # (4, 3, 128, 128)
 ```
 
@@ -52,37 +53,47 @@ model = Unet(
 
 diffusion = GaussianDiffusion(
     model,
-    beta_start = 0.0001,
-    beta_end = 0.02,
-    num_diffusion_timesteps = 1000,   # number of steps
-    loss_type = 'l1'                  # L1 or L2
+    image_size = 128,
+    timesteps = 1000,   # number of steps
+    loss_type = 'l1'    # L1 or L2
 ).cuda()
 
 trainer = Trainer(
     diffusion,
     'path/to/your/images',
-    image_size = 128,
     train_batch_size = 32,
     train_lr = 2e-5,
-    train_num_steps = 100000,         # total training steps
+    train_num_steps = 700000,         # total training steps
     gradient_accumulate_every = 2,    # gradient accumulation steps
-    ema_decay = 0.995                 # exponential moving average decay
+    ema_decay = 0.995,                # exponential moving average decay
+    fp16 = True                       # turn on mixed precision training with apex
 )
 
 trainer.train()
 ```
 
-Todo: Command line tool for one-line training
+Samples and model checkpoints will be logged to `./results` periodically
 
 ## Citations
 
 ```bibtex
 @misc{ho2020denoising,
-    title={Denoising Diffusion Probabilistic Models},
-    author={Jonathan Ho and Ajay Jain and Pieter Abbeel},
-    year={2020},
-    eprint={2006.11239},
-    archivePrefix={arXiv},
-    primaryClass={cs.LG}
+    title   = {Denoising Diffusion Probabilistic Models},
+    author  = {Jonathan Ho and Ajay Jain and Pieter Abbeel},
+    year    = {2020},
+    eprint  = {2006.11239},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.LG}
+}
+```
+
+```bibtex
+@inproceedings{anonymous2021improved,
+    title   = {Improved Denoising Diffusion Probabilistic Models},
+    author  = {Anonymous},
+    booktitle = {Submitted to International Conference on Learning Representations},
+    year    = {2021},
+    url     = {https://openreview.net/forum?id=-NEXDKk8gZ},
+    note    = {under review}
 }
 ```

denoising_diffusion_pytorch/denoising_diffusion_pytorch.py

@@ -16,11 +16,11 @@ import numpy as np
 from tqdm import tqdm
 from einops import rearrange
 
-# constants
-
-SAVE_AND_SAMPLE_EVERY = 1000
-UPDATE_EMA_EVERY = 100
-EXTS = ['jpg', 'png']
+try:
+    from apex import amp
+    APEX_AVAILABLE = True
+except:
+    APEX_AVAILABLE = False
 
 # helpers functions
 
@@ -37,6 +37,21 @@ def cycle(dl):
         for data in dl:
             yield data
 
+def num_to_groups(num, divisor):
+    groups = num // divisor
+    remainder = num % divisor
+    arr = [divisor] * groups
+    if remainder > 0:
+        arr.append(remainder)
+    return arr
+
+def loss_backwards(fp16, loss, optimizer, **kwargs):
+    if fp16:
+        with amp.scale_loss(loss, optimizer) as scaled_loss:
+            scaled_loss.backward(**kwargs)
+    else:
+        loss.backward(**kwargs)
+
 # small helper modules
 
 class EMA():
@@ -96,18 +111,20 @@ class Downsample(nn.Module):
     def forward(self, x):
         return self.conv(x)
 
-class Rezero(nn.Module):
-    def __init__(self, dim):
+class PreNorm(nn.Module):
+    def __init__(self, dim, fn):
         super().__init__()
-        self.g = nn.Parameter(torch.zeros(1))
+        self.fn = fn
+        self.norm = nn.InstanceNorm2d(dim, affine = True)
 
     def forward(self, x):
-        return x * self.g
+        x = self.norm(x)
+        return self.fn(x)
 
 # building block modules
 
 class Block(nn.Module):
-    def __init__(self, dim, dim_out, groups = 32):
+    def __init__(self, dim, dim_out, groups = 8):
         super().__init__()
         self.block = nn.Sequential(
             nn.Conv2d(dim, dim_out, 3, padding=1),
@@ -118,12 +135,12 @@ class Block(nn.Module):
         return self.block(x)
 
 class ResnetBlock(nn.Module):
-    def __init__(self, dim, dim_out, *, time_emb_dim, groups = 32):
+    def __init__(self, dim, dim_out, *, time_emb_dim = None, groups = 8):
         super().__init__()
         self.mlp = nn.Sequential(
             Mish(),
             nn.Linear(time_emb_dim, dim_out)
-        )
+        ) if exists(time_emb_dim) else None
 
         self.block1 = Block(dim, dim_out)
         self.block2 = Block(dim_out, dim_out)
@@ -131,23 +148,26 @@ class ResnetBlock(nn.Module):
 
     def forward(self, x, time_emb):
         h = self.block1(x)
-        h += self.mlp(time_emb)[:, :, None, None]
+
+        if exists(self.mlp):
+            print('hmmm')
+            h += self.mlp(time_emb)[:, :, None, None]
+
         h = self.block2(h)
         return h + self.res_conv(x)
 
 class LinearAttention(nn.Module):
-    def __init__(self, dim, heads = 8, dim_head = 32):
+    def __init__(self, dim, heads = 4, dim_head = 32):
         super().__init__()
         self.heads = heads
         hidden_dim = dim_head * heads
-        self.to_qkv = nn.Conv2d(dim, hidden_dim, 1, bias = False)
+        self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias = False)
         self.to_out = nn.Conv2d(hidden_dim, dim, 1)
 
     def forward(self, x):
         b, c, h, w = x.shape
         qkv = self.to_qkv(x)
-        q, k, v = rearrange(qkv, 'b (qkv heads c) h w -> qkv b heads c (h w)', heads = self.heads)
-        q = q.softmax(dim=-2)
+        q, k, v = rearrange(qkv, 'b (qkv heads c) h w -> qkv b heads c (h w)', heads = self.heads, qkv=3)
         k = k.softmax(dim=-1)
         context = torch.einsum('bhdn,bhen->bhde', k, v)
         out = torch.einsum('bhde,bhdn->bhen', context, q)
@@ -157,17 +177,32 @@ class LinearAttention(nn.Module):
 # model
 
 class Unet(nn.Module):
-    def __init__(self, dim, out_dim = None, dim_mults=(1, 2, 4, 8), groups = 32):
+    def __init__(
+        self,
+        dim,
+        out_dim = None,
+        dim_mults=(1, 2, 4, 8),
+        groups = 8,
+        channels = 3,
+        with_time_emb = True
+    ):
         super().__init__()
-        dims = [3, *map(lambda m: dim * m, dim_mults)]
+        self.channels = channels
+
+        dims = [channels, *map(lambda m: dim * m, dim_mults)]
         in_out = list(zip(dims[:-1], dims[1:]))
 
-        self.time_pos_emb = SinusoidalPosEmb(dim)
-        self.mlp = nn.Sequential(
-            nn.Linear(dim, dim * 4),
-            Mish(),
-            nn.Linear(dim * 4, dim)
-        )
+        if with_time_emb:
+            time_dim = dim
+            self.time_mlp = nn.Sequential(
+                SinusoidalPosEmb(dim),
+                nn.Linear(dim, dim * 4),
+                Mish(),
+                nn.Linear(dim * 4, dim)
+            )
+        else:
+            time_dim = None
+            self.time_mlp = None
 
         self.downs = nn.ModuleList([])
         self.ups = nn.ModuleList([])
@@ -177,39 +212,41 @@ class Unet(nn.Module):
             is_last = ind >= (num_resolutions - 1)
 
             self.downs.append(nn.ModuleList([
-                ResnetBlock(dim_in, dim_out, time_emb_dim = dim),
-                Residual(Rezero(LinearAttention(dim_out))),
+                ResnetBlock(dim_in, dim_out, time_emb_dim = time_dim),
+                ResnetBlock(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()
             ]))
 
         mid_dim = dims[-1]
-        self.mid_block1 = ResnetBlock(mid_dim, mid_dim, time_emb_dim = dim)
-        self.mid_attn = Residual(Rezero(LinearAttention(mid_dim)))
-        self.mid_block2 = ResnetBlock(mid_dim, mid_dim, time_emb_dim = dim)
+        self.mid_block1 = ResnetBlock(mid_dim, mid_dim, time_emb_dim = time_dim)
+        self.mid_attn = Residual(PreNorm(mid_dim, LinearAttention(mid_dim)))
+        self.mid_block2 = ResnetBlock(mid_dim, mid_dim, time_emb_dim = time_dim)
 
         for ind, (dim_in, dim_out) in enumerate(reversed(in_out[1:])):
             is_last = ind >= (num_resolutions - 1)
 
             self.ups.append(nn.ModuleList([
-                ResnetBlock(dim_out * 2, dim_in, time_emb_dim = dim),
-                Residual(Rezero(LinearAttention(dim_in))),
+                ResnetBlock(dim_out * 2, dim_in, time_emb_dim = time_dim),
+                ResnetBlock(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()
             ]))
 
-        out_dim = default(out_dim, 3)
+        out_dim = default(out_dim, channels)
         self.final_conv = nn.Sequential(
             Block(dim, dim),
             nn.Conv2d(dim, out_dim, 1)
         )
 
     def forward(self, x, time):
-        t = self.time_pos_emb(time)
-        t = self.mlp(t)
+        t = self.time_mlp(time) if exists(self.time_mlp) else None
 
         h = []
 
-        for resnet, attn, downsample in self.downs:
+        for resnet, resnet2, attn, downsample in self.downs:
             x = resnet(x, t)
+            x = resnet2(x, t)
             x = attn(x)
             h.append(x)
             x = downsample(x)
@@ -218,9 +255,10 @@ class Unet(nn.Module):
         x = self.mid_attn(x)
         x = self.mid_block2(x, t)
 
-        for resnet, attn, upsample in self.ups:
+        for resnet, resnet2, attn, upsample in self.ups:
             x = torch.cat((x, h.pop()), dim=1)
             x = resnet(x, t)
+            x = resnet2(x, t)
             x = attn(x)
             x = upsample(x)
 
@@ -238,24 +276,47 @@ def noise_like(shape, device, repeat=False):
     noise = lambda: torch.randn(shape, device=device)
     return repeat_noise() if repeat else noise()
 
+def cosine_beta_schedule(timesteps, s = 0.008):
+    """
+    cosine schedule
+    as proposed in https://openreview.net/forum?id=-NEXDKk8gZ
+    """
+    steps = timesteps + 1
+    x = np.linspace(0, steps, steps)
+    alphas_cumprod = np.cos(((x / steps) + s) / (1 + s) * np.pi * 0.5) ** 2
+    alphas_cumprod = alphas_cumprod / alphas_cumprod[0]
+    betas = 1 - (alphas_cumprod[1:] / alphas_cumprod[:-1])
+    return np.clip(betas, a_min = 0, a_max = 0.999)
+
 class GaussianDiffusion(nn.Module):
-    def __init__(self, denoise_fn, beta_start=0.0001, beta_end=0.02, num_diffusion_timesteps=1000, loss_type='l1', betas = None):
+    def __init__(
+        self,
+        denoise_fn,
+        *,
+        image_size,
+        channels = 3,
+        timesteps = 1000,
+        loss_type = 'l1',
+        betas = None
+    ):
         super().__init__()
+        self.channels = channels
+        self.image_size = image_size
         self.denoise_fn = denoise_fn
 
         if exists(betas):
-            self.np_betas = betas.detach().cpu().numpy() if isinstance(betas, torch.Tensor) else betas
+            betas = betas.detach().cpu().numpy() if isinstance(betas, torch.Tensor) else betas
         else:
-            self.np_betas = betas = np.linspace(beta_start, beta_end, num_diffusion_timesteps).astype(np.float64)
-
-        timesteps, = betas.shape
-        self.num_timesteps = int(timesteps)
-        self.loss_type = loss_type
+            betas = cosine_beta_schedule(timesteps)
 
         alphas = 1. - betas
         alphas_cumprod = np.cumprod(alphas, axis=0)
         alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
 
+        timesteps, = betas.shape
+        self.num_timesteps = int(timesteps)
+        self.loss_type = loss_type
+
         to_torch = partial(torch.tensor, dtype=torch.float32)
 
         self.register_buffer('betas', to_torch(betas))
@@ -331,8 +392,10 @@ class GaussianDiffusion(nn.Module):
         return img
 
     @torch.no_grad()
-    def sample(self, image_size, batch_size = 16):
-        return self.p_sample_loop((16, 3, image_size, image_size))
+    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))
 
     @torch.no_grad()
     def interpolate(self, x1, x2, t = None, lam = 0.5):
@@ -375,24 +438,26 @@ class GaussianDiffusion(nn.Module):
         return loss
 
     def forward(self, x, *args, **kwargs):
-        b, *_, device = *x.shape, x.device
+        b, c, h, w, device, img_size, = *x.shape, x.device, self.image_size
+        assert h == img_size and w == img_size, f'height and width of image must be {img_size}'
         t = torch.randint(0, self.num_timesteps, (b,), device=device).long()
         return self.p_losses(x, t, *args, **kwargs)
 
 # dataset classes
 
 class Dataset(data.Dataset):
-    def __init__(self, folder, image_size):
+    def __init__(self, folder, image_size, exts = ['jpg', 'jpeg', 'png']):
         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.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(),
             transforms.CenterCrop(image_size),
-            transforms.ToTensor()
+            transforms.ToTensor(),
+            transforms.Lambda(lambda t: (t * 2) - 1)
         ])
 
     def __len__(self):
@@ -417,16 +482,25 @@ class Trainer(object):
         train_lr = 2e-5,
         train_num_steps = 100000,
         gradient_accumulate_every = 2,
+        fp16 = False,
+        step_start_ema = 2000,
+        update_ema_every = 10,
+        save_and_sample_every = 1000,
+        results_folder = './results'
     ):
         super().__init__()
         self.model = diffusion_model
-
-        self.image_size = image_size
-        self.gradient_accumulate_every = gradient_accumulate_every
-        self.train_num_steps = train_num_steps
-
         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
+
+        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, image_size)
         self.dl = cycle(data.DataLoader(self.ds, batch_size = train_batch_size, shuffle=True, pin_memory=True))
@@ -434,38 +508,64 @@ class Trainer(object):
 
         self.step = 0
 
+        assert not fp16 or fp16 and APEX_AVAILABLE, 'Apex must be installed in order for mixed precision training to be turned on'
+
+        self.fp16 = fp16
+        if fp16:
+            (self.model, self.ema_model), self.opt = amp.initialize([self.model, self.ema_model], self.opt, opt_level='O1')
+
+        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_model.state_dict()
         }
-        torch.save(data, f'./model-{milestone}.pt')
+        torch.save(data, str(self.results_folder / f'model-{milestone}.pt'))
 
     def load(self, milestone):
-        data = torch.load(f'./model-{milestone}.pt')
+        data = torch.load(str(self.results_folder / f'model-{milestone}.pt'))
+
         self.step = data['step']
-        self.model = data['model']
-        self.ema_model = data['ema']
+        self.model.load_state_dict(data['model'])
+        self.ema_model.load_state_dict(data['ema'])
 
     def train(self):
+        backwards = partial(loss_backwards, self.fp16)
+
         while self.step < self.train_num_steps:
             for i in range(self.gradient_accumulate_every):
                 data = next(self.dl).cuda()
                 loss = self.model(data)
                 print(f'{self.step}: {loss.item()}')
-                (loss / self.gradient_accumulate_every).backward()
+                backwards(loss / self.gradient_accumulate_every, self.opt)
 
             self.opt.step()
             self.opt.zero_grad()
 
-            if self.step % UPDATE_EMA_EVERY == 0:
-                self.ema.update_model_average(self.ema_model, self.model)
+            if self.step % self.update_ema_every == 0:
+                self.step_ema()
 
-            if self.step % SAVE_AND_SAMPLE_EVERY == 0:
-                milestone = self.step // SAVE_AND_SAMPLE_EVERY
-                all_images = self.ema_model.p_sample_loop((64, 3, self.image_size, self.image_size))
-                utils.save_image(all_images, f'./sample-{milestone}.png', nrow=8)
+            if self.step != 0 and self.step % self.save_and_sample_every == 0:
+                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)
+                all_images = (all_images + 1) * 0.5
+                utils.save_image(all_images, str(self.results_folder / f'sample-{milestone}.png'), nrow = 6)
                 self.save(milestone)
 
             self.step += 1

setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'denoising-diffusion-pytorch',
   packages = find_packages(),
-  version = '0.2.0',
+  version = '0.6.7',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',