revert restriction for classifier free guidance for diffusion prior, given @crowsonkb advice

Added metric to track image_embed vs predicted_image_embed

dalle2_pytorch/dalle2_pytorch.py

@@ -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 if not predict_x_start else 0.
+        self.cond_drop_prob = cond_drop_prob
         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 = '0.1.9',
+  version = '0.1.10',
   license='MIT',
   description = 'DALL-E 2',
   author = 'Phil Wang',

train_diffusion_prior.py

@@ -93,6 +93,8 @@ def report_cosine_sims(diffusion_prior, image_reader, text_reader, train_set_siz
             text_embed, predicted_image_embeddings).cpu().numpy()
         unrelated_similarity = cos(
             text_embed, predicted_unrelated_embeddings).cpu().numpy()
+        predicted_img_similarity = cos(
+            test_image_embeddings, predicted_image_embeddings).cpu().numpy()
 
         wandb.log(
             {"CosineSimilarity(text_embed,image_embed)": np.mean(original_similarity)})
@@ -100,6 +102,8 @@ def report_cosine_sims(diffusion_prior, image_reader, text_reader, train_set_siz
             predicted_similarity)})
         wandb.log({"CosineSimilarity(text_embed,predicted_unrelated_embed)": np.mean(
             unrelated_similarity)})
+        wandb.log({"CosineSimilarity(image_embed,predicted_image_embed)": np.mean(
+            predicted_img_similarity)})
 
     return np.mean(predicted_similarity - original_similarity)