force first unet in the cascade to be conditioned on image embeds

dalle2_pytorch/dalle2_pytorch.py

@@ -647,9 +647,12 @@ class DiffusionPrior(BaseGaussianDiffusion):
         )
 
         if exists(clip):
-            assert isinstance(clip, CLIP)
+            if isinstance(clip, CLIP):
+                clip = XClipAdapter(clip)
+
+            assert isinstance(clip, BaseClipAdapter)
             freeze_model_and_make_eval_(clip)
-            self.clip = XClipAdapter(clip)
+            self.clip = clip
         else:
             assert exists(image_embed_dim), 'latent dimension must be given, if training prior network without CLIP given'
             self.clip = None
@@ -740,7 +743,7 @@ class DiffusionPrior(BaseGaussianDiffusion):
         text_cond = dict(text_embed = text_embed)
 
         if self.condition_on_text_encodings:
-            text_cond = {**text_cond, 'text_encodings': text_encodings, 'mask': text_mask}
+            text_cond = {**text_cond, 'text_encodings': text_encodings, 'mask': text != 0}
 
         image_embeds = self.p_sample_loop((batch_size, image_embed_dim), text_cond = text_cond)
         text_embeds = text_cond['text_embed']
@@ -1063,13 +1066,14 @@ class Unet(nn.Module):
         self,
         *,
         lowres_cond,
-        channels
+        channels,
+        cond_on_image_embeds
     ):
-        if lowres_cond == self.lowres_cond and channels == self.channels:
+        if lowres_cond == self.lowres_cond and channels == self.channels and cond_on_image_embeds == self.cond_on_image_embeds:
             return self
 
-        updated_kwargs = {**self._locals, 'lowres_cond': lowres_cond, 'channels': channels}
+        updated_kwargs = {'lowres_cond': lowres_cond, 'channels': channels, 'cond_on_image_embeds': cond_on_image_embeds}
-        return self.__class__(**updated_kwargs)
+        return self.__class__(**{**self._locals, **updated_kwargs})
 
     def forward_with_cond_scale(
         self,
@@ -1248,6 +1252,8 @@ class Decoder(BaseGaussianDiffusion):
             clip = XClipAdapter(clip)
 
         freeze_model_and_make_eval_(clip)
+        assert isinstance(clip, BaseClipAdapter)
+
         self.clip = clip
         self.clip_image_size = clip.image_size
         self.channels = clip.image_channels
@@ -1274,6 +1280,7 @@ class Decoder(BaseGaussianDiffusion):
 
             one_unet = one_unet.cast_model_parameters(
                 lowres_cond = not is_first,
+                cond_on_image_embeds = is_first,
                 channels = unet_channels
             )
 

dalle2_pytorch/vqgan_vae.py

@@ -545,6 +545,7 @@ class VQGanVAE(nn.Module):
         l2_recon_loss = False,
         use_hinge_loss = True,
         vgg = None,
+        vq_codebook_dim = 256,
         vq_codebook_size = 512,
         vq_decay = 0.8,
         vq_commitment_weight = 1.,
@@ -579,6 +580,7 @@ class VQGanVAE(nn.Module):
 
         self.vq = VQ(
             dim = self.enc_dec.encoded_dim,
+            codebook_dim = vq_codebook_dim,
             codebook_size = vq_codebook_size,
             decay = vq_decay,
             commitment_weight = vq_commitment_weight,

setup.py

@@ -10,7 +10,7 @@ setup(
       'dream = dalle2_pytorch.cli:dream'
     ],
   },
-  version = '0.0.60',
+  version = '0.0.64',
   license='MIT',
   description = 'DALL-E 2',
   author = 'Phil Wang',