cleanup

README.md

@@ -199,7 +199,7 @@ dalle2 = DALLE2(
     decoder = decoder
 )
 
-# send the text as a string if you want to use the simple tokenizer from DALL-E1
+# send the text as a string if you want to use the simple tokenizer from DALLE v1
 # or you can do it as token ids, if you have your own tokenizer
 
 texts = ['glistening morning dew on a flower petal']
@@ -212,10 +212,7 @@ Let's see the whole script below
 
 ```python
 import torch
-from dalle2_pytorch.dalle2_pytorch import DALLE2
+from dalle2_pytorch import DALLE2, DiffusionPriorNetwork, DiffusionPrior, Unet, Decoder, CLIP
-from dalle2_pytorch import DiffusionPriorNetwork, DiffusionPrior, Unet, Decoder, CLIP
-
-import torch
 
 clip = CLIP(
     dim_text = 512,
@@ -304,6 +301,8 @@ images = dalle2(['cute puppy chasing after a squirrel'])
 
 Everything in this readme should run without error
 
+For the layperson, no worries, training will all be automated into a CLI tool, at least for small scale training.
+
 ## Training CLI (wip)
 
 <a href="https://github.com/lucidrains/stylegan2-pytorch">template</a>
@@ -365,3 +364,5 @@ Everything in this readme should run without error
     primaryClass = {cs.LG}
 }
 ```
+
+*Creating noise from data is easy; creating data from noise is generative modeling.* - Yang Song's <a href="https://arxiv.org/abs/2011.13456">paper</a>

dalle2_pytorch/dalle2_pytorch.py

@@ -374,12 +374,13 @@ class DiffusionPrior(nn.Module):
         image_encoding = self.clip.visual_transformer(image)
         image_cls = image_encoding[:, 0]
         image_embed = self.clip.to_visual_latent(image_cls)
-        return image_embed
+        return l2norm(image_embed)
 
     def get_text_cond(self, text):
         text_encodings = self.clip.text_transformer(text)
         text_cls, text_encodings = text_encodings[:, 0], text_encodings[:, 1:]
         text_embed = self.clip.to_text_latent(text_cls)
+        text_embed = l2norm(text_embed)
         return dict(text_encodings = text_encodings, text_embed = text_embed, mask = text != 0)
 
     def q_mean_variance(self, x_start, t):
@@ -750,7 +751,7 @@ class Decoder(nn.Module):
         image_encoding = self.clip.visual_transformer(image)
         image_cls = image_encoding[:, 0]
         image_embed = self.clip.to_visual_latent(image_cls)
-        return image_embed
+        return l2norm(image_embed)
 
     def q_mean_variance(self, x_start, t):
         mean = extract(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start
@@ -876,7 +877,6 @@ class DALLE2(nn.Module):
             text = [text] if not isinstance(text, (list, tuple)) else text
             text = tokenizer.tokenize(text).to(device)
 
-        print(text.shape, type(text))
         image_embed = self.prior.sample(text, num_samples_per_batch = self.prior_num_samples)
         images = self.decoder.sample(image_embed)
         return images

setup.py

@@ -10,7 +10,7 @@ setup(
       'dream = dalle2_pytorch.cli:dream'
     ],
   },
-  version = '0.0.4',
+  version = '0.0.6',
   license='MIT',
   description = 'DALL-E 2',
   author = 'Phil Wang',