fix a dtype conversion issue for the diffusion timesteps in the diffusion prior, thanks to @JiaHeng-DLUT

README.md

@@ -634,10 +634,12 @@ Alternatively, you can also use <a href="https://github.com/mlfoundations/open_c
 $ pip install open-clip-torch
 ```
 
+Ex. using the <a href="https://laion.ai/blog/large-openclip/">SOTA Open Clip</a> model trained by <a href="https://github.com/rom1504">Romain</a>
+
 ```python
 from dalle2_pytorch import OpenClipAdapter
 
-clip = OpenClipAdapter()
+clip = OpenClipAdapter('ViT-H/14')
 ```
 
 Now you'll just have to worry about training the Prior and the Decoder!
@@ -1066,7 +1068,7 @@ dataloader = create_image_embedding_dataloader(
 )
 for img, emb in dataloader:
     print(img.shape)  # torch.Size([32, 3, 256, 256])
-    print(emb.shape)  # torch.Size([32, 512])
+    print(emb["img"].shape)  # torch.Size([32, 512])
     # Train decoder only as shown above
 
 # Or create a dataset without a loader so you can configure it manually

dalle2_pytorch/dalle2_pytorch.py

@@ -100,6 +100,9 @@ def eval_decorator(fn):
         return out
     return inner
 
+def is_float_dtype(dtype):
+    return any([dtype == float_dtype for float_dtype in (torch.float64, torch.float32, torch.float16, torch.bfloat16)])
+
 def is_list_str(x):
     if not isinstance(x, (list, tuple)):
         return False
@@ -314,7 +317,10 @@ class OpenAIClipAdapter(BaseClipAdapter):
         self.eos_id = 49407 # for handling 0 being also '!'
 
         text_attention_final = self.find_layer('ln_final')
+
+        self.dim_latent_ = text_attention_final.weight.shape[0]
         self.handle = text_attention_final.register_forward_hook(self._hook)
+
         self.clip_normalize = preprocess.transforms[-1]
         self.cleared = False
 
@@ -333,7 +339,7 @@ class OpenAIClipAdapter(BaseClipAdapter):
 
     @property
     def dim_latent(self):
-        return 512
+        return self.dim_latent_
 
     @property
     def image_size(self):
@@ -965,6 +971,8 @@ class DiffusionPriorNetwork(nn.Module):
             Rearrange('b (n d) -> b n d', n = num_text_embeds)
         )
 
+        self.continuous_embedded_time = not exists(num_timesteps)
+
         self.to_time_embeds = nn.Sequential(
             nn.Embedding(num_timesteps, dim * num_time_embeds) if exists(num_timesteps) else nn.Sequential(SinusoidalPosEmb(dim), MLP(dim, dim * num_time_embeds)), # also offer a continuous version of timestep embeddings, with a 2 layer MLP
             Rearrange('b (n d) -> b n d', n = num_time_embeds)
@@ -1092,6 +1100,9 @@ class DiffusionPriorNetwork(nn.Module):
         # whether text embedding is used for conditioning depends on whether text encodings are available for attention (for classifier free guidance, even though it seems from the paper it was not used in the prior ddpm, as the objective is different)
         # but let's just do it right
 
+        if self.continuous_embedded_time:
+            diffusion_timesteps = diffusion_timesteps.type(dtype)
+
         time_embed = self.to_time_embeds(diffusion_timesteps)
 
         learned_queries = repeat(self.learned_query, 'd -> b 1 d', b = batch)
@@ -1429,7 +1440,7 @@ class DiffusionPrior(nn.Module):
         **kwargs
     ):
         assert exists(text) ^ exists(text_embed), 'either text or text embedding must be supplied'
-        assert exists(image) ^ exists(image_embed), 'either text or text embedding must be supplied'
+        assert exists(image) ^ exists(image_embed), 'either image or image embedding must be supplied'
         assert not (self.condition_on_text_encodings and (not exists(text_encodings) and not exists(text))), 'text encodings must be present if you specified you wish to condition on it on initialization'
 
         if exists(image):
@@ -1535,6 +1546,8 @@ class SinusoidalPosEmb(nn.Module):
 
     def forward(self, x):
         dtype, device = x.dtype, x.device
+        assert is_float_dtype(dtype), 'input to sinusoidal pos emb must be a float type'
+
         half_dim = self.dim // 2
         emb = math.log(10000) / (half_dim - 1)
         emb = torch.exp(torch.arange(half_dim, device = device, dtype = dtype) * -emb)

dalle2_pytorch/version.py

@@ -1 +1 @@
-__version__ = '1.10.6'
+__version__ = '1.10.8'