autoregressive-diffusion-pytorch 0.2.2__tar.gz → 0.2.4__tar.gz

{autoregressive_diffusion_pytorch-0.2.2 → autoregressive_diffusion_pytorch-0.2.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: autoregressive-diffusion-pytorch
-Version: 0.2.2
+Version: 0.2.4
 Summary: Autoregressive Diffusion - Pytorch
 Project-URL: Homepage, https://pypi.org/project/autoregressive-diffusion-pytorch/
 Project-URL: Repository, https://github.com/lucidrains/autoregressive-diffusion-pytorch
@@ -36,7 +36,9 @@ Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
 Requires-Python: >=3.8
 Requires-Dist: einops>=0.8.0
 Requires-Dist: einx>=0.3.0
+Requires-Dist: ema-pytorch
 Requires-Dist: torch>=2.0
+Requires-Dist: torchdiffeq
 Requires-Dist: tqdm
 Requires-Dist: x-transformers>=1.31.14
 Provides-Extra: examples
@@ -52,6 +54,8 @@ Implementation of the architecture behind <a href="https://arxiv.org/abs/2406.11
 
 Official repository has been released <a href="https://github.com/LTH14/mar">here</a>
 
+<a href="https://github.com/lucidrains/transfusion-pytorch">Alternative route</a>
+
 <img src="./images/results.96600.png" width="400px"></img>
 
 *oxford flowers at 96k steps*
@@ -217,3 +221,25 @@ trainer()
     url     = {https://api.semanticscholar.org/CorpusID:249240415}
 }
 ```
+
+```bibtex
+@article{Liu2022FlowSA,
+    title   = {Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow},
+    author  = {Xingchao Liu and Chengyue Gong and Qiang Liu},
+    journal = {ArXiv},
+    year    = {2022},
+    volume  = {abs/2209.03003},
+    url     = {https://api.semanticscholar.org/CorpusID:252111177}
+}
+```
+
+```bibtex
+@article{Esser2024ScalingRF,
+    title   = {Scaling Rectified Flow Transformers for High-Resolution Image Synthesis},
+    author  = {Patrick Esser and Sumith Kulal and A. Blattmann and Rahim Entezari and Jonas Muller and Harry Saini and Yam Levi and Dominik Lorenz and Axel Sauer and Frederic Boesel and Dustin Podell and Tim Dockhorn and Zion English and Kyle Lacey and Alex Goodwin and Yannik Marek and Robin Rombach},
+    journal = {ArXiv},
+    year    = {2024},
+    volume  = {abs/2403.03206},
+    url     = {https://api.semanticscholar.org/CorpusID:268247980}
+}
+```

{autoregressive_diffusion_pytorch-0.2.2 → autoregressive_diffusion_pytorch-0.2.4}/README.md

@@ -6,6 +6,8 @@ Implementation of the architecture behind <a href="https://arxiv.org/abs/2406.11
 
 Official repository has been released <a href="https://github.com/LTH14/mar">here</a>
 
+<a href="https://github.com/lucidrains/transfusion-pytorch">Alternative route</a>
+
 <img src="./images/results.96600.png" width="400px"></img>
 
 *oxford flowers at 96k steps*
@@ -171,3 +173,25 @@ trainer()
     url     = {https://api.semanticscholar.org/CorpusID:249240415}
 }
 ```
+
+```bibtex
+@article{Liu2022FlowSA,
+    title   = {Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow},
+    author  = {Xingchao Liu and Chengyue Gong and Qiang Liu},
+    journal = {ArXiv},
+    year    = {2022},
+    volume  = {abs/2209.03003},
+    url     = {https://api.semanticscholar.org/CorpusID:252111177}
+}
+```
+
+```bibtex
+@article{Esser2024ScalingRF,
+    title   = {Scaling Rectified Flow Transformers for High-Resolution Image Synthesis},
+    author  = {Patrick Esser and Sumith Kulal and A. Blattmann and Rahim Entezari and Jonas Muller and Harry Saini and Yam Levi and Dominik Lorenz and Axel Sauer and Frederic Boesel and Dustin Podell and Tim Dockhorn and Zion English and Kyle Lacey and Alex Goodwin and Yannik Marek and Robin Rombach},
+    journal = {ArXiv},
+    year    = {2024},
+    volume  = {abs/2403.03206},
+    url     = {https://api.semanticscholar.org/CorpusID:268247980}
+}
+```

{autoregressive_diffusion_pytorch-0.2.2 → autoregressive_diffusion_pytorch-0.2.4}/autoregressive_diffusion_pytorch/autoregressive_flow.py

@@ -260,7 +260,8 @@ class AutoregressiveFlow(Module):
 
     def forward(
         self,
-        seq
+        seq,
+        noised_seq = None
     ):
         b, seq_len, dim = seq.shape
 
@@ -271,6 +272,9 @@ class AutoregressiveFlow(Module):
 
         seq, target = seq[:, :-1], seq
 
+        if exists(noised_seq):
+            seq = noised_seq[:, :-1]
+
         # append start tokens
 
         seq = self.proj_in(seq)
@@ -303,6 +307,7 @@ class ImageAutoregressiveFlow(Module):
         image_size,
         patch_size,
         channels = 3,
+        train_max_noise = 0.,
         model: dict = dict(),
     ):
         super().__init__()
@@ -314,6 +319,10 @@ class ImageAutoregressiveFlow(Module):
         self.image_size = image_size
         self.patch_size = patch_size
 
+        assert 0. <= train_max_noise < 1.
+
+        self.train_max_noise = train_max_noise
+
         self.to_tokens = Rearrange('b c (h p1) (w p2) -> b (h w) (c p1 p2)', p1 = patch_size, p2 = patch_size)
 
         self.model = AutoregressiveFlow(
@@ -330,6 +339,20 @@ class ImageAutoregressiveFlow(Module):
         return unnormalize_to_zero_to_one(images)
 
     def forward(self, images):
+        train_under_noise, device = self.train_max_noise > 0., images.device
+
         images = normalize_to_neg_one_to_one(images)
         tokens = self.to_tokens(images)
-        return self.model(tokens)
+
+        if not train_under_noise:
+            return self.model(tokens)
+
+        # allow for the network to predict from slightly noised images of the past
+
+        times = torch.rand(images.shape[0], device = device) * self.train_max_noise
+        noise = torch.randn_like(images)
+        padded_times = right_pad_dims_to(images, times)
+        noised_images = images * (1. - padded_times) + noise * padded_times
+        noised_tokens = self.to_tokens(noised_images)
+
+        return self.model(tokens, noised_tokens)

{autoregressive_diffusion_pytorch-0.2.2 → autoregressive_diffusion_pytorch-0.2.4}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "autoregressive-diffusion-pytorch"
-version = "0.2.2"
+version = "0.2.4"
 description = "Autoregressive Diffusion - Pytorch"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }
@@ -25,8 +25,10 @@ classifiers=[
 dependencies = [
     'einx>=0.3.0',
     'einops>=0.8.0',
+    'ema-pytorch',
     'x-transformers>=1.31.14',
     'torch>=2.0',
+    'torchdiffeq',
     'tqdm'
 ]