remove outdated apex in favor of native pytorch AMP

README.md

@@ -2,7 +2,9 @@
 
 ## Denoising Diffusion Probabilistic Model, in Pytorch
 
-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>. 
+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> and then modified to use <a href="https://arxiv.org/abs/2201.03545">ConvNext</a> blocks instead of Resnets.
 
 <img src="./sample.png" width="500px"><img>
 
@@ -66,7 +68,7 @@ trainer = Trainer(
     train_num_steps = 700000,         # 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
+    amp = True                        # turn on mixed precision
 )
 
 trainer.train()
@@ -97,3 +99,14 @@ Samples and model checkpoints will be logged to `./results` periodically
     note    = {under review}
 }
 ```
+
+```bibtex
+@misc{liu2022convnet,
+    title   = {A ConvNet for the 2020s},
+    author  = {Zhuang Liu and Hanzi Mao and Chao-Yuan Wu and Christoph Feichtenhofer and Trevor Darrell and Saining Xie},
+    year    = {2022},
+    eprint  = {2201.03545},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.CV}
+}
+```

denoising_diffusion_pytorch/denoising_diffusion_pytorch.py

@@ -7,21 +7,16 @@ from inspect import isfunction
 from functools import partial
 
 from torch.utils import data
+from torch.cuda.amp import autocast, GradScaler
+
 from pathlib import Path
 from torch.optim import Adam
 from torchvision import transforms, utils
 from PIL import Image
 
-import numpy as np
 from tqdm import tqdm
 from einops import rearrange
 
-try:
-    from apex import amp
-    APEX_AVAILABLE = True
-except:
-    APEX_AVAILABLE = False
-
 # helpers functions
 
 def exists(x):
@@ -45,13 +40,6 @@ def num_to_groups(num, divisor):
         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():
@@ -91,31 +79,29 @@ class SinusoidalPosEmb(nn.Module):
         emb = torch.cat((emb.sin(), emb.cos()), dim=-1)
         return emb
 
-class Mish(nn.Module):
-    def forward(self, x):
-        return x * torch.tanh(F.softplus(x))
+def Upsample(dim):
+    return nn.ConvTranspose2d(dim, dim, 4, 2, 1)
 
-class Upsample(nn.Module):
-    def __init__(self, dim):
+def Downsample(dim):
+    return nn.Conv2d(dim, dim, 4, 2, 1)
+
+class LayerNorm(nn.Module):
+    def __init__(self, dim, eps = 1e-5):
         super().__init__()
-        self.conv = nn.ConvTranspose2d(dim, dim, 4, 2, 1)
+        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):
-        return self.conv(x)
-
-class Downsample(nn.Module):
-    def __init__(self, dim):
-        super().__init__()
-        self.conv = nn.Conv2d(dim, dim, 3, 2, 1)
-
-    def forward(self, x):
-        return self.conv(x)
+        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
 
 class PreNorm(nn.Module):
     def __init__(self, dim, fn):
         super().__init__()
         self.fn = fn
-        self.norm = nn.InstanceNorm2d(dim, affine = True)
+        self.norm = LayerNorm(dim)
 
     def forward(self, x):
         x = self.norm(x)
@@ -123,42 +109,42 @@ class PreNorm(nn.Module):
 
 # building block modules
 
-class Block(nn.Module):
-    def __init__(self, dim, dim_out, groups = 8):
-        super().__init__()
-        self.block = nn.Sequential(
-            nn.Conv2d(dim, dim_out, 3, padding=1),
-            nn.GroupNorm(groups, dim_out),
-            Mish()
-        )
-    def forward(self, x):
-        return self.block(x)
+class ConvNextBlock(nn.Module):
+    """ https://arxiv.org/abs/2201.03545 """
 
-class ResnetBlock(nn.Module):
-    def __init__(self, dim, dim_out, *, time_emb_dim = None, groups = 8):
+    def __init__(self, dim, dim_out, *, time_emb_dim = None, mult = 2, norm = True):
         super().__init__()
         self.mlp = nn.Sequential(
-            Mish(),
-            nn.Linear(time_emb_dim, dim_out)
+            nn.GELU(),
+            nn.Linear(time_emb_dim, dim)
         ) if exists(time_emb_dim) else None
 
-        self.block1 = Block(dim, dim_out)
-        self.block2 = Block(dim_out, dim_out)
+        self.ds_conv = nn.Conv2d(dim, dim, 7, padding = 3, groups = dim)
+
+        self.net = nn.Sequential(
+            LayerNorm(dim) if norm else nn.Identity(),
+            nn.Conv2d(dim, dim_out * mult, 3, padding = 1),
+            nn.GELU(),
+            nn.Conv2d(dim_out * mult, dim_out, 3, padding = 1)
+        )
+
         self.res_conv = nn.Conv2d(dim, dim_out, 1) if dim != dim_out else nn.Identity()
 
-    def forward(self, x, time_emb):
-        h = self.block1(x)
+    def forward(self, x, time_emb = None):
+        h = self.ds_conv(x)
 
         if exists(self.mlp):
-            print('hmmm')
-            h += self.mlp(time_emb)[:, :, None, None]
+            assert exists(time_emb), 'time emb must be passed in'
+            condition = self.mlp(time_emb)
+            h = h + rearrange(condition, 'b c -> b c 1 1')
 
-        h = self.block2(h)
+        h = self.net(h)
         return h + self.res_conv(x)
 
 class LinearAttention(nn.Module):
     def __init__(self, dim, heads = 4, dim_head = 32):
         super().__init__()
+        self.scale = dim_head ** -0.5
         self.heads = heads
         hidden_dim = dim_head * heads
         self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias = False)
@@ -166,12 +152,38 @@ class LinearAttention(nn.Module):
 
     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, qkv=3)
-        k = k.softmax(dim=-1)
-        context = torch.einsum('bhdn,bhen->bhde', k, v)
-        out = torch.einsum('bhde,bhdn->bhen', context, q)
-        out = rearrange(out, 'b heads c (h w) -> b (heads c) h w', heads=self.heads, h=h, w=w)
+        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
+
+        k = k.softmax(dim = -1)
+        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)
+        out = rearrange(out, 'b h c (x y) -> b (h c) x y', h = self.heads, x = h, y = w)
+        return self.to_out(out)
+
+class Attention(nn.Module):
+    def __init__(self, dim, heads = 4, dim_head = 32):
+        super().__init__()
+        self.scale = dim_head ** -0.5
+        self.heads = heads
+        hidden_dim = dim_head * heads
+        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).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()
+        attn = sim.softmax(dim = -1)
+
+        out = einsum('b h i j, b h d j -> b h i d', attn, v)
+        out = rearrange(out, 'b h (x y) d -> b (h d) x y', x = h, y = w)
         return self.to_out(out)
 
 # model
@@ -182,7 +194,6 @@ class Unet(nn.Module):
         dim,
         out_dim = None,
         dim_mults=(1, 2, 4, 8),
-        groups = 8,
         channels = 3,
         with_time_emb = True
     ):
@@ -197,7 +208,7 @@ class Unet(nn.Module):
             self.time_mlp = nn.Sequential(
                 SinusoidalPosEmb(dim),
                 nn.Linear(dim, dim * 4),
-                Mish(),
+                nn.GELU(),
                 nn.Linear(dim * 4, dim)
             )
         else:
@@ -212,30 +223,30 @@ class Unet(nn.Module):
             is_last = ind >= (num_resolutions - 1)
 
             self.downs.append(nn.ModuleList([
-                ResnetBlock(dim_in, dim_out, time_emb_dim = time_dim),
-                ResnetBlock(dim_out, dim_out, time_emb_dim = time_dim),
+                ConvNextBlock(dim_in, dim_out, time_emb_dim = time_dim, norm = ind != 0),
+                ConvNextBlock(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 = 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)
+        self.mid_block1 = ConvNextBlock(mid_dim, mid_dim, time_emb_dim = time_dim)
+        self.mid_attn = Residual(PreNorm(mid_dim, Attention(mid_dim)))
+        self.mid_block2 = ConvNextBlock(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 = time_dim),
-                ResnetBlock(dim_in, dim_in, time_emb_dim = time_dim),
+                ConvNextBlock(dim_out * 2, dim_in, time_emb_dim = time_dim),
+                ConvNextBlock(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, channels)
         self.final_conv = nn.Sequential(
-            Block(dim, dim),
+            ConvNextBlock(dim, dim),
             nn.Conv2d(dim, out_dim, 1)
         )
 
@@ -244,9 +255,9 @@ class Unet(nn.Module):
 
         h = []
 
-        for resnet, resnet2, attn, downsample in self.downs:
-            x = resnet(x, t)
-            x = resnet2(x, t)
+        for convnext, convnext2, attn, downsample in self.downs:
+            x = convnext(x, t)
+            x = convnext2(x, t)
             x = attn(x)
             h.append(x)
             x = downsample(x)
@@ -255,10 +266,10 @@ class Unet(nn.Module):
         x = self.mid_attn(x)
         x = self.mid_block2(x, t)
 
-        for resnet, resnet2, attn, upsample in self.ups:
+        for convnext, convnext2, attn, upsample in self.ups:
             x = torch.cat((x, h.pop()), dim=1)
-            x = resnet(x, t)
-            x = resnet2(x, t)
+            x = convnext(x, t)
+            x = convnext2(x, t)
             x = attn(x)
             x = upsample(x)
 
@@ -282,11 +293,11 @@ def cosine_beta_schedule(timesteps, s = 0.008):
     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
+    x = torch.linspace(0, steps, steps)
+    alphas_cumprod = torch.cos(((x / steps) + s) / (1 + s) * torch.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)
+    return torch.clip(betas, 0, 0.999)
 
 class GaussianDiffusion(nn.Module):
     def __init__(
@@ -296,50 +307,48 @@ class GaussianDiffusion(nn.Module):
         image_size,
         channels = 3,
         timesteps = 1000,
-        loss_type = 'l1',
-        betas = None
+        loss_type = 'l1'
     ):
         super().__init__()
         self.channels = channels
         self.image_size = image_size
         self.denoise_fn = denoise_fn
 
-        if exists(betas):
-            betas = betas.detach().cpu().numpy() if isinstance(betas, torch.Tensor) else betas
-        else:
-            betas = cosine_beta_schedule(timesteps)
+        betas = cosine_beta_schedule(timesteps)
 
         alphas = 1. - betas
-        alphas_cumprod = np.cumprod(alphas, axis=0)
-        alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
+        alphas_cumprod = torch.cumprod(alphas, axis=0)
+        alphas_cumprod_prev = F.pad(alphas_cumprod[:-1], (0, 1), value = 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))
-        self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
-        self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev))
+        self.register_buffer('betas', betas)
+        self.register_buffer('alphas_cumprod', alphas_cumprod)
+        self.register_buffer('alphas_cumprod_prev', alphas_cumprod_prev)
 
         # calculations for diffusion q(x_t | x_{t-1}) and others
-        self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod)))
-        self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod)))
-        self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod)))
-        self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod)))
-        self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1)))
+
+        self.register_buffer('sqrt_alphas_cumprod', torch.sqrt(alphas_cumprod))
+        self.register_buffer('sqrt_one_minus_alphas_cumprod', torch.sqrt(1. - alphas_cumprod))
+        self.register_buffer('log_one_minus_alphas_cumprod', torch.log(1. - alphas_cumprod))
+        self.register_buffer('sqrt_recip_alphas_cumprod', torch.sqrt(1. / alphas_cumprod))
+        self.register_buffer('sqrt_recipm1_alphas_cumprod', torch.sqrt(1. / alphas_cumprod - 1))
 
         # calculations for posterior q(x_{t-1} | x_t, x_0)
+
         posterior_variance = betas * (1. - alphas_cumprod_prev) / (1. - alphas_cumprod)
+
         # above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t)
-        self.register_buffer('posterior_variance', to_torch(posterior_variance))
+
+        self.register_buffer('posterior_variance', posterior_variance)
+
         # below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain
-        self.register_buffer('posterior_log_variance_clipped', to_torch(np.log(np.maximum(posterior_variance, 1e-20))))
-        self.register_buffer('posterior_mean_coef1', to_torch(
-            betas * np.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod)))
-        self.register_buffer('posterior_mean_coef2', to_torch(
-            (1. - alphas_cumprod_prev) * np.sqrt(alphas) / (1. - alphas_cumprod)))
+
+        self.register_buffer('posterior_log_variance_clipped', torch.log(posterior_variance.clamp(min =1e-20)))
+        self.register_buffer('posterior_mean_coef1', betas * torch.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod))
+        self.register_buffer('posterior_mean_coef2', (1. - alphas_cumprod_prev) * torch.sqrt(alphas) / (1. - alphas_cumprod))
 
     def q_mean_variance(self, x_start, t):
         mean = extract(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start
@@ -482,7 +491,7 @@ class Trainer(object):
         train_lr = 2e-5,
         train_num_steps = 100000,
         gradient_accumulate_every = 2,
-        fp16 = False,
+        amp = False,
         step_start_ema = 2000,
         update_ema_every = 10,
         save_and_sample_every = 1000,
@@ -508,11 +517,8 @@ 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.amp = amp
+        self.scaler = GradScaler(enabled = amp)
 
         self.results_folder = Path(results_folder)
         self.results_folder.mkdir(exist_ok = True)
@@ -532,7 +538,8 @@ class Trainer(object):
         data = {
             'step': self.step,
             'model': self.model.state_dict(),
-            'ema': self.ema_model.state_dict()
+            'ema': self.ema_model.state_dict(),
+            'scaler': self.scaler.state_dict()
         }
         torch.save(data, str(self.results_folder / f'model-{milestone}.pt'))
 
@@ -542,18 +549,21 @@ class Trainer(object):
         self.step = data['step']
         self.model.load_state_dict(data['model'])
         self.ema_model.load_state_dict(data['ema'])
+        self.scaler.load_state_dict(data['scaler'])
 
     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()}')
-                backwards(loss / self.gradient_accumulate_every, self.opt)
 
-            self.opt.step()
+                with autocast(enabled = self.amp):
+                    loss = self.model(data)
+                    self.scaler.scale(loss / self.gradient_accumulate_every).backward()
+
+                print(f'{self.step}: {loss.item()}')
+
+            self.scaler.step(self.opt)
+            self.scaler.update()
             self.opt.zero_grad()
 
             if self.step % self.update_ema_every == 0:

setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'denoising-diffusion-pytorch',
   packages = find_packages(),
-  version = '0.6.7',
+  version = '0.9.0',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',
@@ -15,7 +15,6 @@ setup(
   ],
   install_requires=[
     'einops',
-    'numpy',
     'pillow',
     'torch',
     'torchvision',