update with the best type of downsample

README.md

@@ -768,3 +768,13 @@ Anything! It is MIT licensed. In other words, you can freely copy / paste for yo
     primaryClass = {cs.CV}
 }
 ```
+
+```bibtex
+@article{Sunkara2022NoMS,
+    title   = {No More Strided Convolutions or Pooling: A New CNN Building Block for Low-Resolution Images and Small Objects},
+    author  = {Raja Sunkara and Tie Luo},
+    journal = {ArXiv},
+    year    = {2022},
+    volume  = {abs/2208.03641}
+}
+```

imagen_pytorch/imagen_pytorch.py

@@ -583,8 +583,13 @@ def forward(self, x):
         return self.net(x)
 
 def Downsample(dim, dim_out = None):
+    # https://arxiv.org/abs/2208.03641 shows this is the most optimal way to downsample
+    # named SP-conv in the paper, but basically a pixel unshuffle
     dim_out = default(dim_out, dim)
-    return nn.Conv2d(dim, dim_out, 4, 2, 1)
+    return nn.Sequential(
+        Rearrange('b c (h s1) (w s2) -> b (c s1 s2) h w', s1 = 2, s2 = 2),
+        nn.Conv2d(dim * 4, dim_out, 1)
+    )
 
 class SinusoidalPosEmb(nn.Module):
     def __init__(self, dim):

imagen_pytorch/version.py

@@ -1 +1 @@
-__version__ = '1.10.1'
+__version__ = '1.11.0'