bring in the dynamic thresholding technique from the Imagen paper, which purportedly improves classifier free guidance for the cascading ddpm

README.md

@@ -12,7 +12,7 @@ This model is SOTA for text-to-image for now.
 
 Please join <a href="https://discord.gg/xBPBXfcFHd"><img alt="Join us on Discord" src="https://img.shields.io/discord/823813159592001537?color=5865F2&logo=discord&logoColor=white"></a> if you are interested in helping out with the replication with the <a href="https://laion.ai/">LAION</a> community | <a href="https://www.youtube.com/watch?v=AIOE1l1W0Tw">Yannic Interview</a>
 
-There was enough interest for a <a href="https://github.com/lucidrains/dalle2-jax">Jax version</a>. I will also eventually extend this to <a href="https://github.com/lucidrains/dalle2-video">text to video</a>, once the repository is in a good place.
+As of 5/23/22, it is no longer SOTA. SOTA will be <a href="https://github.com/lucidrains/imagen-pytorch">here</a>. Jax versions as well as text-to-video project will be shifted towards the Imagen architecture, as it is way simpler.
 
 ## Status
 
@@ -26,7 +26,7 @@ There was enough interest for a <a href="https://github.com/lucidrains/dalle2-ja
 
 ## Pre-Trained Models
 - LAION is training prior models. Checkpoints are available on <a href="https://huggingface.co/zenglishuci/conditioned-prior">🤗huggingface</a> and the training statistics are available on <a href="https://wandb.ai/nousr_laion/conditioned-prior/reports/LAION-DALLE2-PyTorch-Prior--VmlldzoyMDI2OTIx">🐝WANDB</a>.
-- Decoder 🚧
+- Decoder - <a href="https://wandb.ai/veldrovive/dalle2_train_decoder/runs/jkrtg0so?workspace=user-veldrovive">In-progress test run</a> 🚧
 - DALL-E 2 🚧
 
 ## Install
@@ -1195,4 +1195,12 @@ This library would not have gotten to this working state without the help of
 }
 ```
 
+```bibtex
+@misc{Saharia2022,
+    title   = {Imagen: unprecedented photorealism × deep level of language understanding},
+    author  = {Chitwan Saharia*, William Chan*, Saurabh Saxena†, Lala Li†, Jay Whang†, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho†, David Fleet†, Mohammad Norouzi*},
+    year    = {2022}
+}
+```
+
 *Creating noise from data is easy; creating data from noise is generative modeling.* - <a href="https://arxiv.org/abs/2011.13456">Yang Song's paper</a>

dalle2_pytorch/dalle2_pytorch.py

@@ -890,6 +890,8 @@ class DiffusionPrior(BaseGaussianDiffusion):
         )
 
         if exists(clip):
+            assert image_channels == clip.image_channels, f'channels of image ({image_channels}) should be equal to the channels that CLIP accepts ({clip.image_channels})'
+
             if isinstance(clip, CLIP):
                 clip = XClipAdapter(clip, **clip_adapter_overrides)
             elif isinstance(clip, CoCa):
@@ -1702,6 +1704,8 @@ class Decoder(BaseGaussianDiffusion):
         vb_loss_weight = 0.001,
         unconditional = False,
         auto_normalize_img = True,                  # whether to take care of normalizing the image from [0, 1] to [-1, 1] and back automatically - you can turn this off if you want to pass in the [-1, 1] ranged image yourself from the dataloader
+        use_dynamic_thres = False,                  # from the Imagen paper
+        dynamic_thres_percentile = 0.9
     ):
         super().__init__(
             beta_schedule = beta_schedule,
@@ -1710,12 +1714,19 @@ class Decoder(BaseGaussianDiffusion):
         )
 
         self.unconditional = unconditional
-        assert not (condition_on_text_encodings and unconditional), 'unconditional decoder image generation cannot be set to True if conditioning on text is present'
 
-        assert self.unconditional or (exists(clip) or exists(image_size) or exists(image_sizes)), 'either CLIP is supplied, or you must give the image_size and channels (usually 3 for RGB)'
+        # text conditioning
+
+        assert not (condition_on_text_encodings and unconditional), 'unconditional decoder image generation cannot be set to True if conditioning on text is present'
+        self.condition_on_text_encodings = condition_on_text_encodings
+
+        # clip
 
         self.clip = None
         if exists(clip):
+            assert not unconditional, 'clip must not be given if doing unconditional image training'
+            assert channels == clip.image_channels, f'channels of image ({channels}) should be equal to the channels that CLIP accepts ({clip.image_channels})'
+
             if isinstance(clip, CLIP):
                 clip = XClipAdapter(clip, **clip_adapter_overrides)
             elif isinstance(clip, CoCa):
@@ -1725,13 +1736,20 @@ class Decoder(BaseGaussianDiffusion):
             assert isinstance(clip, BaseClipAdapter)
 
             self.clip = clip
-            self.clip_image_size = clip.image_size
-            self.channels = clip.image_channels
-        else:
-            self.clip_image_size = default(image_size, lambda: image_sizes[-1])
-            self.channels = channels
 
-        self.condition_on_text_encodings = condition_on_text_encodings
+        # determine image size, with image_size and image_sizes taking precedence
+
+        if exists(image_size) or exists(image_sizes):
+            assert exists(image_size) ^ exists(image_sizes), 'only one of image_size or image_sizes must be given'
+            image_size = default(image_size, lambda: image_sizes[-1])
+        elif exists(clip):
+            image_size = clip.image_size
+        else:
+            raise Error('either image_size, image_sizes, or clip must be given to decoder')
+
+        # channels
+
+        self.channels = channels
 
         # automatically take care of ensuring that first unet is unconditional
         # while the rest of the unets are conditioned on the low resolution image produced by previous unet
@@ -1773,7 +1791,7 @@ class Decoder(BaseGaussianDiffusion):
 
         # unet image sizes
 
-        image_sizes = default(image_sizes, (self.clip_image_size,))
+        image_sizes = default(image_sizes, (image_size,))
         image_sizes = tuple(sorted(set(image_sizes)))
 
         assert len(self.unets) == len(image_sizes), f'you did not supply the correct number of u-nets ({len(self.unets)}) for resolutions {image_sizes}'
@@ -1810,7 +1828,13 @@ class Decoder(BaseGaussianDiffusion):
         self.clip_denoised = clip_denoised
         self.clip_x_start = clip_x_start
 
+        # dynamic thresholding settings, if clipping denoised during sampling
+
+        self.use_dynamic_thres = use_dynamic_thres
+        self.dynamic_thres_percentile = dynamic_thres_percentile
+
         # normalize and unnormalize image functions
+
         self.normalize_img = normalize_neg_one_to_one if auto_normalize_img else identity
         self.unnormalize_img = unnormalize_zero_to_one if auto_normalize_img else identity
 
@@ -1851,7 +1875,21 @@ class Decoder(BaseGaussianDiffusion):
             x_recon = self.predict_start_from_noise(x, t = t, noise = pred)
 
         if clip_denoised:
-            x_recon.clamp_(-1., 1.)
+            # s is the threshold amount
+            # static thresholding would just be s = 1
+            s = 1.
+            if self.use_dynamic_thres:
+                s = torch.quantile(
+                    rearrange(x_recon, 'b ... -> b (...)').abs(),
+                    self.dynamic_thres_percentile,
+                    dim = -1
+                )
+
+                s.clamp_(min = 1.)
+                s = s.view(-1, *((1,) * (x_recon.ndim - 1)))
+
+            # clip by threshold, depending on whether static or dynamic
+            x_recon = x_recon.clamp(-s, s) / s
 
         model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
 

setup.py

@@ -10,7 +10,7 @@ setup(
       'dream = dalle2_pytorch.cli:dream'
     ],
   },
-  version = '0.4.10',
+  version = '0.5.0',
   license='MIT',
   description = 'DALL-E 2',
   author = 'Phil Wang',