fix decoder needing separate conditional dropping probabilities for image embeddings and text encodings, thanks to @xiankgx !

Fix spelling and grammatical errors

README.md

@@ -47,7 +47,7 @@ clip = CLIP(
     use_all_token_embeds = True,            # whether to use fine-grained contrastive learning (FILIP)
     decoupled_contrastive_learning = True,  # use decoupled contrastive learning (DCL) objective function, removing positive pairs from the denominator of the InfoNCE loss (CLOOB + DCL)
     extra_latent_projection = True,         # whether to use separate projections for text-to-image vs image-to-text comparisons (CLOOB)
-    use_visual_ssl = True,                  # whether to do self supervised learning on iages
+    use_visual_ssl = True,                  # whether to do self supervised learning on images
     visual_ssl_type = 'simclr',             # can be either 'simclr' or 'simsiam', depending on using DeCLIP or SLIP
     use_mlm = False,                        # use masked language learning (MLM) on text (DeCLIP)
     text_ssl_loss_weight = 0.05,            # weight for text MLM loss
@@ -110,7 +110,8 @@ decoder = Decoder(
     unet = unet,
     clip = clip,
     timesteps = 100,
-    cond_drop_prob = 0.2
+    image_cond_drop_prob = 0.1,
+    text_cond_drop_prob = 0.5
 ).cuda()
 
 # mock images (get a lot of this)
@@ -229,7 +230,8 @@ decoder = Decoder(
     unet = (unet1, unet2),            # insert both unets in order of low resolution to highest resolution (you can have as many stages as you want here)
     image_sizes = (256, 512),         # resolutions, 256 for first unet, 512 for second. these must be unique and in ascending order (matches with the unets passed in)
     timesteps = 1000,
-    cond_drop_prob = 0.2
+    image_cond_drop_prob = 0.1,
+    text_cond_drop_prob = 0.5
 ).cuda()
 
 # mock images (get a lot of this)
@@ -348,7 +350,8 @@ decoder = Decoder(
     image_sizes = (128, 256),
     clip = clip,
     timesteps = 100,
-    cond_drop_prob = 0.2,
+    image_cond_drop_prob = 0.1,
+    text_cond_drop_prob = 0.5,
     condition_on_text_encodings = False  # set this to True if you wish to condition on text during training and sampling
 ).cuda()
 
@@ -499,9 +502,7 @@ loss.backward()
 
 Although there is the possibility they are using an unreleased, more powerful CLIP, you can use one of the released ones, if you do not wish to train your own CLIP from scratch. This will also allow the community to more quickly validate the conclusions of the paper.
 
-First you'll need to install <a href="https://github.com/openai/CLIP#usage">the prerequisites</a>
-
-Then to use a pretrained OpenAI CLIP, simply import `OpenAIClipAdapter` and pass it into the `DiffusionPrior` or `Decoder` like so
+To use a pretrained OpenAI CLIP, simply import `OpenAIClipAdapter` and pass it into the `DiffusionPrior` or `Decoder` like so
 
 ```python
 import torch
@@ -560,7 +561,8 @@ decoder = Decoder(
     image_sizes = (128, 256),
     clip = clip,
     timesteps = 100,
-    cond_drop_prob = 0.2,
+    image_cond_drop_prob = 0.1,
+    text_cond_drop_prob = 0.5,
     condition_on_text_encodings = False  # set this to True if you wish to condition on text during training and sampling
 ).cuda()
 
@@ -618,7 +620,7 @@ clip = CLIP(
 # 3 unets for the decoder (a la cascading DDPM)
 
 # first two unets are doing latent diffusion
-# vqgan-vae must be trained before hand
+# vqgan-vae must be trained beforehand
 
 vae1 = VQGanVAE(
     dim = 32,
@@ -671,7 +673,8 @@ decoder = Decoder(
     unet = (unet1, unet2, unet3),      # insert unets in order of low resolution to highest resolution (you can have as many stages as you want here)
     image_sizes = (256, 512, 1024),    # resolutions, 256 for first unet, 512 for second, 1024 for third
     timesteps = 100,
-    cond_drop_prob = 0.2
+    image_cond_drop_prob = 0.1,
+    text_cond_drop_prob = 0.5
 ).cuda()
 
 # mock images (get a lot of this)

dalle2_pytorch/dalle2_pytorch.py

@@ -172,17 +172,13 @@ class OpenAIClipAdapter(BaseClipAdapter):
         self,
         name = 'ViT-B/32'
     ):
-        try:
-            import clip
-        except ImportError:
-            print('you must install openai clip in order to use this adapter - `pip install git+https://github.com/openai/CLIP.git` - more instructions at https://github.com/openai/CLIP#usage')
-
-        openai_clip, _ = clip.load(name)
+        import clip
+        openai_clip, preprocess = clip.load(name)
         super().__init__(openai_clip)
 
         text_attention_final = self.find_layer('ln_final')
         self.handle = text_attention_final.register_forward_hook(self._hook)
-        self.clip_normalize = T.Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711))
+        self.clip_normalize = preprocess.transforms[-1]
         self.cleared = False
 
     def find_layer(self,  layer):
@@ -1178,7 +1174,7 @@ class Unet(nn.Module):
         if cond_scale == 1:
             return logits
 
-        null_logits = self.forward(*args, cond_drop_prob = 1., **kwargs)
+        null_logits = self.forward(*args, text_cond_drop_prob = 1., image_cond_drop_prob = 1., **kwargs)
         return null_logits + (logits - null_logits) * cond_scale
 
     def forward(
@@ -1189,7 +1185,8 @@ class Unet(nn.Module):
         image_embed,
         lowres_cond_img = None,
         text_encodings = None,
-        cond_drop_prob = 0.,
+        image_cond_drop_prob = 0.,
+        text_cond_drop_prob = 0.,
         blur_sigma = None,
         blur_kernel_size = None
     ):
@@ -1208,8 +1205,10 @@ class Unet(nn.Module):
 
         # conditional dropout
 
-        keep_mask = prob_mask_like((batch_size,), 1 - cond_drop_prob, device = device)
-        keep_mask = rearrange(keep_mask, 'b -> b 1 1')
+        image_keep_mask = prob_mask_like((batch_size,), 1 - image_cond_drop_prob, device = device)
+        text_keep_mask = prob_mask_like((batch_size,), 1 - text_cond_drop_prob, device = device)
+
+        image_keep_mask, text_keep_mask = rearrange_many((image_keep_mask, text_keep_mask), 'b -> b 1 1')
 
         # mask out image embedding depending on condition dropout
         # for classifier free guidance
@@ -1220,7 +1219,7 @@ class Unet(nn.Module):
             image_tokens = self.image_to_cond(image_embed)
 
             image_tokens = torch.where(
-                keep_mask,
+                image_keep_mask,
                 image_tokens,
                 self.null_image_embed
             )
@@ -1232,7 +1231,7 @@ class Unet(nn.Module):
         if exists(text_encodings) and self.cond_on_text_encodings:
             text_tokens = self.text_to_cond(text_encodings)
             text_tokens = torch.where(
-                keep_mask,
+                text_keep_mask,
                 text_tokens,
                 self.null_text_embed[:, :text_tokens.shape[1]]
             )
@@ -1322,7 +1321,8 @@ class Decoder(BaseGaussianDiffusion):
         clip,
         vae = tuple(),
         timesteps = 1000,
-        cond_drop_prob = 0.2,
+        image_cond_drop_prob = 0.1,
+        text_cond_drop_prob = 0.5,
         loss_type = 'l1',
         beta_schedule = 'cosine',
         predict_x_start = False,
@@ -1406,7 +1406,8 @@ class Decoder(BaseGaussianDiffusion):
 
         # classifier free guidance
 
-        self.cond_drop_prob = cond_drop_prob
+        self.image_cond_drop_prob = image_cond_drop_prob
+        self.text_cond_drop_prob = text_cond_drop_prob
 
     def get_unet(self, unet_number):
         assert 0 < unet_number <= len(self.unets)
@@ -1488,7 +1489,8 @@ class Decoder(BaseGaussianDiffusion):
             image_embed = image_embed,
             text_encodings = text_encodings,
             lowres_cond_img = lowres_cond_img,
-            cond_drop_prob = self.cond_drop_prob
+            image_cond_drop_prob = self.image_cond_drop_prob,
+            text_cond_drop_prob = self.text_cond_drop_prob,
         )
 
         target = noise if not predict_x_start else x_start
@@ -1636,4 +1638,3 @@ class DALLE2(nn.Module):
             return images[0]
 
         return images
-

setup.py

@@ -10,7 +10,7 @@ setup(
       'dream = dalle2_pytorch.cli:dream'
     ],
   },
-  version = '0.0.72',
+  version = '0.0.74',
   license='MIT',
   description = 'DALL-E 2',
   author = 'Phil Wang',
@@ -23,6 +23,7 @@ setup(
   ],
   install_requires=[
     'click',
+    'clip-anytorch',
     'einops>=0.4',
     'einops-exts>=0.0.3',
     'kornia>=0.5.4',