update with new and improved cosine noise scheduler

README.md

@@ -4,6 +4,10 @@
 
 Implementation of <a href="https://arxiv.org/abs/2006.11239">Denoising Diffusion Probabilistic Model</a> in Pytorch. It is a new approach to generative modeling that may <a href="https://ajolicoeur.wordpress.com/the-new-contender-to-gans-score-matching-with-langevin-sampling/">have the potential</a> to rival GANs. It uses denoising score matching to estimate the gradient of the data distribution, followed by Langevin sampling to sample from the true distribution. This implementation was transcribed from the official Tensorflow version <a href="https://github.com/hojonathanho/diffusion">here</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)
+
 ## Install
 
 ```bash
@@ -23,10 +27,8 @@ 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?)
+    timesteps = 1000,   # number of steps
+    loss_type = 'l1'                  # L1 or L2
 )
 
 training_images = torch.randn(8, 3, 128, 128)
@@ -50,10 +52,8 @@ 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
+    timesteps = 1000,   # number of steps
+    loss_type = 'l1'    # L1 or L2
 ).cuda()
 
 trainer = Trainer(
@@ -62,15 +62,15 @@ trainer = Trainer(
     image_size = 128,
     train_batch_size = 32,
     train_lr = 2e-5,
-    train_num_steps = 100000,
-    gradient_accumulate_every = 2
+    train_num_steps = 100000,         # total training steps
+    gradient_accumulate_every = 2,    # gradient accumulation steps
+    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
-
 ## Citations
 
 ```bibtex
@@ -83,3 +83,15 @@ Todo: Command line tool for one-line training
     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

@@ -1,4 +1,5 @@
 import math
+import copy
 import torch
 from torch import nn, einsum
 import torch.nn.functional as F
@@ -15,10 +16,17 @@ import numpy as np
 from tqdm import tqdm
 from einops import rearrange
 
+try:
+    from apex import amp
+    APEX_AVAILABLE = True
+except:
+    APEX_AVAILABLE = False
+
 # constants
 
 SAVE_AND_SAMPLE_EVERY = 1000
-EXTS = ['jpg', 'png']
+UPDATE_EMA_EVERY = 10
+EXTS = ['jpg', 'jpeg', 'png']
 
 # helpers functions
 
@@ -35,8 +43,38 @@ 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():
+    def __init__(self, beta):
+        super().__init__()
+        self.beta = beta
+
+    def update_model_average(self, ma_model, current_model):
+        for current_params, ma_params in zip(current_model.parameters(), ma_model.parameters()):
+            old_weight, up_weight = ma_params.data, current_params.data
+            ma_params.data = self.update_average(old_weight, up_weight)
+
+    def update_average(self, old, new):
+        if old is None:
+            return new
+        return old * self.beta + (1 - self.beta) * new
+
 class Residual(nn.Module):
     def __init__(self, fn):
         super().__init__()
@@ -80,17 +118,18 @@ class Downsample(nn.Module):
         return self.conv(x)
 
 class Rezero(nn.Module):
-    def __init__(self, dim):
+    def __init__(self, fn):
         super().__init__()
+        self.fn = fn
         self.g = nn.Parameter(torch.zeros(1))
 
     def forward(self, x):
-        return x * self.g
+        return self.fn(x) * self.g
 
 # 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),
@@ -101,7 +140,7 @@ 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, groups = 8):
         super().__init__()
         self.mlp = nn.Sequential(
             Mish(),
@@ -119,18 +158,17 @@ class ResnetBlock(nn.Module):
         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)
@@ -140,7 +178,7 @@ 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):
         super().__init__()
         dims = [3, *map(lambda m: dim * m, dim_mults)]
         in_out = list(zip(dims[:-1], dims[1:]))
@@ -161,6 +199,7 @@ class Unet(nn.Module):
 
             self.downs.append(nn.ModuleList([
                 ResnetBlock(dim_in, dim_out, time_emb_dim = dim),
+                ResnetBlock(dim_out, dim_out, time_emb_dim = dim),
                 Residual(Rezero(LinearAttention(dim_out))),
                 Downsample(dim_out) if not is_last else nn.Identity()
             ]))
@@ -175,6 +214,7 @@ class Unet(nn.Module):
 
             self.ups.append(nn.ModuleList([
                 ResnetBlock(dim_out * 2, dim_in, time_emb_dim = dim),
+                ResnetBlock(dim_in, dim_in, time_emb_dim = dim),
                 Residual(Rezero(LinearAttention(dim_in))),
                 Upsample(dim_in) if not is_last else nn.Identity()
             ]))
@@ -191,8 +231,9 @@ class Unet(nn.Module):
 
         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)
@@ -201,9 +242,10 @@ 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)
 
@@ -221,20 +263,36 @@ 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'):
+    def __init__(self, denoise_fn, timesteps=1000, loss_type='l1', betas = None):
         super().__init__()
         self.denoise_fn = denoise_fn
 
-        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
+        if exists(betas):
+            betas = betas.detach().cpu().numpy() if isinstance(betas, torch.Tensor) else betas
+        else:
+            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))
@@ -311,7 +369,7 @@ class GaussianDiffusion(nn.Module):
 
     @torch.no_grad()
     def sample(self, image_size, batch_size = 16):
-        return self.p_sample_loop((16, 3, image_size, image_size))
+        return self.p_sample_loop((batch_size, 3, image_size, image_size))
 
     @torch.no_grad()
     def interpolate(self, x1, x2, t = None, lam = 0.5):
@@ -390,14 +448,22 @@ class Trainer(object):
         diffusion_model,
         folder,
         *,
+        ema_decay = 0.995,
         image_size = 128,
         train_batch_size = 32,
         train_lr = 2e-5,
         train_num_steps = 100000,
-        gradient_accumulate_every = 2
+        gradient_accumulate_every = 2,
+        fp16 = False,
+        step_start_ema = 2000
     ):
         super().__init__()
         self.model = diffusion_model
+        self.ema = EMA(ema_decay)
+        self.ema_model = copy.deepcopy(self.model)
+        self.step_start_ema = step_start_ema
+
+        self.batch_size = train_batch_size
         self.image_size = image_size
         self.gradient_accumulate_every = gradient_accumulate_every
         self.train_num_steps = train_num_steps
@@ -406,25 +472,64 @@ class Trainer(object):
         self.dl = cycle(data.DataLoader(self.ds, batch_size = train_batch_size, shuffle=True, pin_memory=True))
         self.opt = Adam(diffusion_model.parameters(), lr=train_lr)
 
-    def train(self):
-        ind = 0
+        self.step = 0
 
-        while ind < self.train_num_steps:
+        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.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')
+
+    def load(self, milestone):
+        data = torch.load(f'./model-{milestone}.pt')
+
+        self.step = data['step']
+        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'{ind}: {loss.item()}')
-                (loss / self.gradient_accumulate_every).backward()
+                print(f'{self.step}: {loss.item()}')
+                backwards(loss / self.gradient_accumulate_every, self.opt)
 
             self.opt.step()
             self.opt.zero_grad()
 
-            if ind % SAVE_AND_SAMPLE_EVERY == 0:
-                milestone = ind // SAVE_AND_SAMPLE_EVERY
-                all_images = self.model.p_sample_loop((64, 3, self.image_size, self.image_size))
-                utils.save_image(all_images, f'./sample-{milestone}.png', nrow=8)
-                torch.save(self.model.state_dict(), f'./model-{milestone}.pt')
+            if self.step % UPDATE_EMA_EVERY == 0:
+                self.step_ema()
 
-            ind += 1
+            if self.step != 0 and self.step % SAVE_AND_SAMPLE_EVERY == 0:
+                milestone = self.step // SAVE_AND_SAMPLE_EVERY
+                batches = num_to_groups(36, self.batch_size)
+                all_images_list = list(map(lambda n: self.ema_model.sample(self.image_size, batch_size=n), batches))
+                all_images = torch.cat(all_images_list, dim=0)
+                utils.save_image(all_images, f'./sample-{milestone}.png', nrow=6)
+                self.save(milestone)
+
+            self.step += 1
 
         print('training completed')

setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'denoising-diffusion-pytorch',
   packages = find_packages(),
-  version = '0.1.4',
+  version = '0.5.0',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',