turn off classifier free guidance if predicting x_start for diffusion prior

README.md

@@ -981,6 +981,8 @@ Once built, images will be saved to the same directory the command is invoked
 - [ ] make sure FILIP works with DALL-E2 from x-clip https://arxiv.org/abs/2111.07783
 - [ ] make sure resnet hyperparameters can be configurable across unet depth (groups and expansion factor)
 - [ ] offer save / load methods on the trainer classes to automatically take care of state dicts for scalers / optimizers / saving versions and checking for breaking changes
+- [ ] offer setting in diffusion prior to split time and image embeddings into multiple tokens, configurable, for more surface area during attention
+- [ ] bring in skip-layer excitatons (from lightweight gan paper) to see if it helps for either decoder of unet or vqgan-vae training
 
 ## Citations
 

dalle2_pytorch/dalle2_pytorch.py

@@ -706,7 +706,7 @@ class DiffusionPriorNetwork(nn.Module):
         **kwargs
     ):
         super().__init__()
-        self.time_embeddings = nn.Embedding(num_timesteps, dim) if exists(num_timesteps) else nn.Sequential(Rearrange('b -> b 1'), MLP(1, dim)) # also offer a continuous version of timestep embeddings, with a 2 layer MLP
+        self.time_embeddings = nn.Embedding(num_timesteps, dim) if exists(num_timesteps) else nn.Sequential(SinusoidalPosEmb(dim), MLP(dim, dim)) # also offer a continuous version of timestep embeddings, with a 2 layer MLP
         self.learned_query = nn.Parameter(torch.randn(dim))
         self.causal_transformer = CausalTransformer(dim = dim, **kwargs)
 
@@ -800,7 +800,7 @@ class DiffusionPrior(BaseGaussianDiffusion):
         image_size = None,
         image_channels = 3,
         timesteps = 1000,
-        cond_drop_prob = 0.2,
+        cond_drop_prob = 0.,
         loss_type = "l1",
         predict_x_start = True,
         beta_schedule = "cosine",
@@ -834,7 +834,7 @@ class DiffusionPrior(BaseGaussianDiffusion):
         self.image_embed_dim = default(image_embed_dim, lambda: clip.dim_latent)
         self.channels = default(image_channels, lambda: clip.image_channels)
 
-        self.cond_drop_prob = cond_drop_prob
+        self.cond_drop_prob = cond_drop_prob if not predict_x_start else 0.
         self.condition_on_text_encodings = condition_on_text_encodings
 
         # in paper, they do not predict the noise, but predict x0 directly for image embedding, claiming empirically better results. I'll just offer both.

setup.py

@@ -10,7 +10,7 @@ setup(
       'dream = dalle2_pytorch.cli:dream'
     ],
   },
-  version = '0.1.7',
+  version = '0.1.9',
   license='MIT',
   description = 'DALL-E 2',
   author = 'Phil Wang',