switch over to film like conditioning, used by both openai and google at this point

README.md

@@ -64,7 +64,7 @@ trainer = Trainer(
     diffusion,
     'path/to/your/images',
     train_batch_size = 32,
-    train_lr = 2e-5,
+    train_lr = 1e-4,
     train_num_steps = 700000,         # total training steps
     gradient_accumulate_every = 2,    # gradient accumulation steps
     ema_decay = 0.995,                # exponential moving average decay

denoising_diffusion_pytorch/__init__.py

@@ -1,2 +1,4 @@
 from denoising_diffusion_pytorch.denoising_diffusion_pytorch import GaussianDiffusion, Unet, Trainer
+
 from denoising_diffusion_pytorch.learned_gaussian_diffusion import LearnedGaussianDiffusion
+from denoising_diffusion_pytorch.weighted_objective_gaussian_diffusion import WeightedObjectiveGaussianDiffusion

denoising_diffusion_pytorch/denoising_diffusion_pytorch.py

@@ -118,20 +118,27 @@ class PreNorm(nn.Module):
 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),
-            nn.SiLU()
-        )
-    def forward(self, x):
-        return self.block(x)
+        self.proj = nn.Conv2d(dim, dim_out, 3, padding = 1)
+        self.norm = nn.GroupNorm(groups, dim_out)
+        self.act = nn.SiLU()
+
+    def forward(self, x, scale_shift = None):
+        x = self.proj(x)
+        x = self.norm(x)
+
+        if exists(scale_shift):
+            scale, shift = scale_shift
+            x = x * (scale + 1) + shift
+
+        x = self.act(x)
+        return x
 
 class ResnetBlock(nn.Module):
     def __init__(self, dim, dim_out, *, time_emb_dim = None, groups = 8):
         super().__init__()
         self.mlp = nn.Sequential(
             nn.SiLU(),
-            nn.Linear(time_emb_dim, dim_out)
+            nn.Linear(time_emb_dim, dim_out * 2)
         ) if exists(time_emb_dim) else None
 
         self.block1 = Block(dim, dim_out, groups = groups)
@@ -139,11 +146,14 @@ class ResnetBlock(nn.Module):
         self.res_conv = nn.Conv2d(dim, dim_out, 1) if dim != dim_out else nn.Identity()
 
     def forward(self, x, time_emb = None):
-        h = self.block1(x)
 
+        scale_shift = None
         if exists(self.mlp) and exists(time_emb):
             time_emb = self.mlp(time_emb)
-            h = rearrange(time_emb, 'b c -> b c 1 1') + h
+            time_emb = rearrange(time_emb, 'b c -> b c 1 1')
+            scale_shift = time_emb.chunk(2, dim = 1)
+
+        h = self.block1(x, scale_shift = scale_shift)
 
         h = self.block2(h)
         return h + self.res_conv(x)
@@ -439,6 +449,8 @@ class GaussianDiffusion(nn.Module):
 
         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))
+
+        img = unnormalize_to_zero_to_one(img)
         return img
 
     @torch.no_grad()
@@ -497,11 +509,13 @@ class GaussianDiffusion(nn.Module):
         loss = self.loss_fn(model_out, target)
         return loss
 
-    def forward(self, x, *args, **kwargs):
-        b, c, h, w, device, img_size, = *x.shape, x.device, self.image_size
+    def forward(self, img, *args, **kwargs):
+        b, c, h, w, device, img_size, = *img.shape, img.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)
+
+        img = normalize_to_neg_one_to_one(img)
+        return self.p_losses(img, t, *args, **kwargs)
 
 # dataset classes
 
@@ -516,8 +530,7 @@ class Dataset(data.Dataset):
             transforms.Resize(image_size),
             transforms.RandomHorizontalFlip(),
             transforms.CenterCrop(image_size),
-            transforms.ToTensor(),
-            transforms.Lambda(normalize_to_neg_one_to_one)
+            transforms.ToTensor()
         ])
 
     def __len__(self):
@@ -539,7 +552,7 @@ class Trainer(object):
         ema_decay = 0.995,
         image_size = 128,
         train_batch_size = 32,
-        train_lr = 2e-5,
+        train_lr = 1e-4,
         train_num_steps = 100000,
         gradient_accumulate_every = 2,
         amp = False,
@@ -603,34 +616,36 @@ class Trainer(object):
         self.scaler.load_state_dict(data['scaler'])
 
     def train(self):
-        while self.step < self.train_num_steps:
-            for i in range(self.gradient_accumulate_every):
-                data = next(self.dl).cuda()
+        with tqdm(initial = self.step, total = self.train_num_steps) as pbar:
 
-                with autocast(enabled = self.amp):
-                    loss = self.model(data)
-                    self.scaler.scale(loss / self.gradient_accumulate_every).backward()
+            while self.step < self.train_num_steps:
+                for i in range(self.gradient_accumulate_every):
+                    data = next(self.dl).cuda()
 
-                print(f'{self.step}: {loss.item()}')
+                    with autocast(enabled = self.amp):
+                        loss = self.model(data)
+                        self.scaler.scale(loss / self.gradient_accumulate_every).backward()
 
-            self.scaler.step(self.opt)
-            self.scaler.update()
-            self.opt.zero_grad()
+                    pbar.set_description(f'loss: {loss.item():.4f}')
 
-            if self.step % self.update_ema_every == 0:
-                self.step_ema()
+                self.scaler.step(self.opt)
+                self.scaler.update()
+                self.opt.zero_grad()
 
-            if self.step != 0 and self.step % self.save_and_sample_every == 0:
-                self.ema_model.eval()
+                if self.step % self.update_ema_every == 0:
+                    self.step_ema()
 
-                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 = unnormalize_to_zero_to_one(all_images)
-                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_model.eval()
 
-            self.step += 1
+                    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)
+                    utils.save_image(all_images, str(self.results_folder / f'sample-{milestone}.png'), nrow = 6)
+                    self.save(milestone)
 
-        print('training completed')
+                self.step += 1
+                pbar.update(1)
+
+        print('training complete')

denoising_diffusion_pytorch/weighted_objective_gaussian_diffusion.py

@@ -0,0 +1,80 @@
+import torch
+from inspect import isfunction
+from torch import nn, einsum
+from einops import rearrange
+
+from denoising_diffusion_pytorch.denoising_diffusion_pytorch import GaussianDiffusion
+
+# helper functions
+
+def exists(x):
+    return x is not None
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+# some improvisation on my end
+# where i have the model learn to both predict noise and x0
+# and learn the weighted sum for each depending on time step
+
+class WeightedObjectiveGaussianDiffusion(GaussianDiffusion):
+    def __init__(
+        self,
+        denoise_fn,
+        *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'
+
+        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)
+
+        pred_noise, pred_x_start, weights = model_output.split(self.split_dims, dim = 1)
+        normalized_weights = weights.softmax(dim = 1)
+
+        x_start_from_noise = self.predict_start_from_noise(x, t = t, noise = pred_noise)
+        
+        x_starts = torch.stack((x_start_from_noise, pred_x_start), dim = 1)
+        weighted_x_start = einsum('b j h w, b j c h w -> b c h w', normalized_weights, x_starts)
+
+        if clip_denoised:
+            weighted_x_start.clamp_(-1., 1.)
+
+        model_mean, model_variance, model_log_variance = self.q_posterior(weighted_x_start, x, t)
+
+        return model_mean, model_variance, model_log_variance
+
+    def p_losses(self, x_start, t, noise = None, clip_denoised = False):
+        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)
+        pred_noise, pred_x_start, weights = model_output.split(self.split_dims, dim = 1)
+
+        # get loss for predicted noise and x_start
+        # with the loss weight given at initialization
+
+        noise_loss = self.loss_fn(noise, pred_noise) * self.pred_noise_loss_weight
+        x_start_loss = self.loss_fn(x_start, pred_x_start) * self.pred_x_start_loss_weight
+
+        # calculate x_start from predicted noise
+        # then do a weighted sum of the x_start prediction, weights also predicted by the model (softmax normalized)
+
+        x_start_from_pred_noise = self.predict_start_from_noise(x_t, t, pred_noise)
+        x_start_from_pred_noise = x_start_from_pred_noise.clamp(-2., 2.)
+        weighted_x_start = einsum('b j h w, b j c h w -> b c h w', weights.softmax(dim = 1), torch.stack((x_start_from_pred_noise, pred_x_start), dim = 1))
+
+        # main loss to x_start with the weighted one
+
+        weighted_x_start_loss = self.loss_fn(x_start, weighted_x_start)
+        return weighted_x_start_loss + x_start_loss + noise_loss

setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'denoising-diffusion-pytorch',
   packages = find_packages(),
-  version = '0.15.0',
+  version = '0.16.0',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',