fix everything and make sure it runs end to end, document everything in readme for public

README.md

@@ -34,7 +34,7 @@ Once built, images will be saved to the same directory the command is invoked
 
 To train DALLE-2 is a 3 step process, with the training of CLIP being the most important
 
-To train CLIP, you can either use <a href="https://github.com/lucidrains/x-clip">x-clip</a> package, or join the LAION discord, where a lot of replication efforts are already underway.
+To train CLIP, you can either use `x-clip` package, or join the LAION discord, where a lot of replication efforts are already underway.
 
 This repository will demonstrate integration with `x-clip` for starters
 
@@ -136,14 +136,12 @@ loss.backward()
 # then it will learn to generate images based on the CLIP image embeddings
 ```
 
-Finally, the main contribution of the paper. The repository offers the diffusion prior network. It takes the CLIP text embeddings and tries to generate the CLIP image embeddings. Again, you will need the trained CLIP from the first step
+Finally, the main contribution of the paper. The repository offers the diffusion prior network. It takes the CLIP text embeddings and tries to generate the CLIP image embeddings. Again, you will need the trained CLIP fron the first step
 
 ```python
 import torch
 from dalle2_pytorch import DiffusionPriorNetwork, DiffusionPrior, CLIP
 
-# get trained CLIP from step one
-
 clip = CLIP(
     dim_text = 512,
     dim_image = 512,
@@ -201,7 +199,7 @@ dalle2 = DALLE2(
     decoder = decoder
 )
 
-# send the text as a string if you want to use the simple tokenizer from DALLE v1
+# send the text as a string if you want to use the simple tokenizer from DALL-E1
 # or you can do it as token ids, if you have your own tokenizer
 
 texts = ['glistening morning dew on a flower petal']
@@ -214,7 +212,10 @@ Let's see the whole script below
 
 ```python
 import torch
-from dalle2_pytorch import DALLE2, DiffusionPriorNetwork, DiffusionPrior, Unet, Decoder, CLIP
+from dalle2_pytorch.dalle2_pytorch import DALLE2
+from dalle2_pytorch import DiffusionPriorNetwork, DiffusionPrior, Unet, Decoder, CLIP
+
+import torch
 
 clip = CLIP(
     dim_text = 512,
@@ -296,18 +297,13 @@ dalle2 = DALLE2(
     decoder = decoder
 )
 
-images = dalle2(
-    ['cute puppy chasing after a squirrel'],
-    cond_scale = 2. # classifier free guidance strength (> 1 would strengthen the condition)
-)
+images = dalle2(['cute puppy chasing after a squirrel'])
 
 # save your image
 ```
 
 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>
@@ -322,7 +318,6 @@ For the layperson, no worries, training will all be automated into a CLI tool, a
 - [ ] figure out all the current bag of tricks needed to make DDPMs great (starting with the blur trick mentioned in paper)
 - [ ] train on a toy task, offer in colab
 - [ ] add attention to unet - apply some personal tricks with efficient attention
-- [ ] figure out the big idea behind latent diffusion and what can be ported over
 
 ## Citations
 
@@ -370,5 +365,3 @@ For the layperson, no worries, training will all be automated into a CLI tool, a
     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

@@ -246,16 +246,15 @@ class DiffusionPriorNetwork(nn.Module):
     def forward_with_cond_scale(
         self,
         x,
-        *args,
+        *,
         cond_scale = 1.,
         **kwargs
     ):
-        logits = self.forward(x, *args, **kwargs)
-
         if cond_scale == 1:
-            return logits
+            return self.forward(x, **kwargs)
 
-        null_logits = self.forward(x, *args, cond_drop_prob = 1., **kwargs)
+        logits = self.forward(x, **kwargs)
+        null_logits = self.forward(x, cond_drop_prob = 1., **kwargs)
         return null_logits + (logits - null_logits) * cond_scale
 
     def forward(
@@ -375,13 +374,12 @@ 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 l2norm(image_embed)
+        return 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):
@@ -636,16 +634,15 @@ class Unet(nn.Module):
     def forward_with_cond_scale(
         self,
         x,
-        *args,
+        *,
         cond_scale = 1.,
         **kwargs
     ):
-        logits = self.forward(x, *args, **kwargs)
-
         if cond_scale == 1:
-            return logits
+            return self.forward(x, **kwargs)
 
-        null_logits = self.forward(x, *args, cond_drop_prob = 1., **kwargs)
+        logits = self.forward(x, **kwargs)
+        null_logits = self.forward(x, cond_drop_prob = 1., **kwargs)
         return null_logits + (logits - null_logits) * cond_scale
 
     def forward(
@@ -753,7 +750,7 @@ 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 l2norm(image_embed)
+        return image_embed
 
     def q_mean_variance(self, x_start, t):
         mean = extract(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start
@@ -776,8 +773,8 @@ class Decoder(nn.Module):
         posterior_log_variance_clipped = extract(self.posterior_log_variance_clipped, t, x_t.shape)
         return posterior_mean, posterior_variance, posterior_log_variance_clipped
 
-    def p_mean_variance(self, x, t, image_embed, clip_denoised = True, cond_scale = 1.):
-        x_recon = self.predict_start_from_noise(x, t = t, noise = self.net.forward_with_cond_scale(x, t, image_embed = image_embed, cond_scale = cond_scale))
+    def p_mean_variance(self, x, t, image_embed, clip_denoised: bool):
+        x_recon = self.predict_start_from_noise(x, t = t, noise = self.net(x, t, image_embed = image_embed))
 
         if clip_denoised:
             x_recon.clamp_(-1., 1.)
@@ -786,31 +783,31 @@ class Decoder(nn.Module):
         return model_mean, posterior_variance, posterior_log_variance
 
     @torch.no_grad()
-    def p_sample(self, x, t, image_embed, cond_scale = 1., clip_denoised = True, repeat_noise = False):
+    def p_sample(self, x, t, image_embed, clip_denoised = True, repeat_noise = False):
         b, *_, device = *x.shape, x.device
-        model_mean, _, model_log_variance = self.p_mean_variance(x = x, t = t, image_embed = image_embed, cond_scale = cond_scale, clip_denoised = clip_denoised)
+        model_mean, _, model_log_variance = self.p_mean_variance(x = x, t = t, image_embed = image_embed, clip_denoised = clip_denoised)
         noise = noise_like(x.shape, device, repeat_noise)
         # no noise when t == 0
         nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
         return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
 
     @torch.no_grad()
-    def p_sample_loop(self, shape, image_embed, cond_scale = 1):
+    def p_sample_loop(self, shape, image_embed):
         device = self.betas.device
 
         b = shape[0]
         img = torch.randn(shape, device=device)
 
         for i in tqdm(reversed(range(0, self.num_timesteps)), desc='sampling loop time step', total=self.num_timesteps):
-            img = self.p_sample(img, torch.full((b,), i, device = device, dtype = torch.long), image_embed = image_embed, cond_scale = cond_scale)
+            img = self.p_sample(img, torch.full((b,), i, device = device, dtype = torch.long), image_embed = image_embed)
         return img
 
     @torch.no_grad()
-    def sample(self, image_embed, cond_scale = 1.):
+    def sample(self, image_embed):
         batch_size = image_embed.shape[0]
         image_size = self.image_size
         channels = self.channels
-        return self.p_sample_loop((batch_size, channels, image_size, image_size), image_embed = image_embed, cond_scale = cond_scale)
+        return self.p_sample_loop((batch_size, channels, image_size, image_size), image_embed = image_embed)
 
     def q_sample(self, x_start, t, noise=None):
         noise = default(noise, lambda: torch.randn_like(x_start))
@@ -871,8 +868,7 @@ class DALLE2(nn.Module):
     @torch.no_grad()
     def forward(
         self,
-        text,
-        cond_scale = 1.
+        text
     ):
         device = next(self.parameters()).device
 
@@ -880,6 +876,7 @@ 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, cond_scale = cond_scale)
+        images = self.decoder.sample(image_embed)
         return images

setup.py

@@ -10,7 +10,7 @@ setup(
       'dream = dalle2_pytorch.cli:dream'
     ],
   },
-  version = '0.0.7',
+  version = '0.0.4',
   license='MIT',
   description = 'DALL-E 2',
   author = 'Phil Wang',