add ability for DALL-E2 to return PIL images with return_pil_images = True on forward, for those who have no clue about deep learning

dalle2_pytorch/dalle2_pytorch.py

@@ -7,6 +7,7 @@ from contextlib import contextmanager
 import torch
 import torch.nn.functional as F
 from torch import nn, einsum
+import torchvision.transforms as T
 
 from einops import rearrange, repeat
 from einops.layers.torch import Rearrange
@@ -736,11 +737,10 @@ class DiffusionPrior(BaseGaussianDiffusion):
 
         text_embed, text_encodings = self.clip.embed_text(text)
 
-        text_cond = dict(
-            text_embed = text_embed,
-            text_encodings = text_encodings,
-            mask = text != 0
-        )
+        text_cond = dict(text_embed = text_embed)
+
+        if self.condition_on_text_encodings:
+            text_cond = {**text_cond, 'text_encodings': text_encodings, 'mask': text_mask}
 
         image_embeds = self.p_sample_loop((batch_size, image_embed_dim), text_cond = text_cond)
         text_embeds = text_cond['text_embed']
@@ -780,11 +780,10 @@ class DiffusionPrior(BaseGaussianDiffusion):
             text_embed, text_encodings = self.clip.embed_text(text)
             text_mask = text != 0
 
-        text_cond = dict(
-            text_embed = text_embed,
-            text_encodings = text_encodings,
-            mask = text_mask
-        )
+        text_cond = dict(text_embed = text_embed)
+
+        if self.condition_on_text_encodings:
+            text_cond = {**text_cond, 'text_encodings': text_encodings, 'mask': text_mask}
 
         # timestep conditioning from ddpm
 
@@ -1518,12 +1517,15 @@ class DALLE2(nn.Module):
         self.prior_num_samples = prior_num_samples
         self.decoder_need_text_cond = self.decoder.condition_on_text_encodings
 
+        self.to_pil = T.ToPILImage()
+
     @torch.no_grad()
     @eval_decorator
     def forward(
         self,
         text,
-        cond_scale = 1.
+        cond_scale = 1.,
+        return_pil_images = False
     ):
         device = next(self.parameters()).device
         one_text = isinstance(text, str) or (not is_list_str(text) and text.shape[0] == 1)
@@ -1537,7 +1539,13 @@ class DALLE2(nn.Module):
         text_cond = text if self.decoder_need_text_cond else None
         images = self.decoder.sample(image_embed, text = text_cond, cond_scale = cond_scale)
 
+        if return_pil_images:
+            # do some magic - if the user passed in a string text, or a list of strings
+            # assume they do not know anything about tensors and return PIL Image(s)
+            images = list(map(self.to_pil, images.unbind(dim = 0)))
+
         if one_text:
             return images[0]
 
         return images
+

setup.py

@@ -10,7 +10,7 @@ setup(
       'dream = dalle2_pytorch.cli:dream'
     ],
   },
-  version = '0.0.57',
+  version = '0.0.59',
   license='MIT',
   description = 'DALL-E 2',
   author = 'Phil Wang',