Commit1cc5d0a

committed

upgrade to best downsample

1 parent59fa101 commit1cc5d0aCopy full SHA for 1cc5d0a

File tree

+19

-4

lines changed

+19

-4

lines changed

Lines changed: 10 additions & 0 deletions

Original file line number	Diff line number	Diff line change
`@@ -1285,4 +1285,14 @@ For detailed information on training the diffusion prior, please refer to the [d`
`1285`	`1285`	`}`
`1286`	`1286`	```
`1287`	`1287`
	`1288`	+```bibtex
	`1289`	`+@article{Sunkara2022NoMS,`
	`1290`	`+ title = {No More Strided Convolutions or Pooling: A New CNN Building Block for Low-Resolution Images and Small Objects},`
	`1291`	`+ author = {Raja Sunkara and Tie Luo},`
	`1292`	`+ journal = {ArXiv},`
	`1293`	`+ year = {2022},`
	`1294`	`+ volume = {abs/2208.03641}`
	`1295`	`+}`
	`1296`	+```
	`1297`	`+`
`1288`	`1298`	`Creating noise from data is easy; creating data from noise is generative modeling. - <ahref="https://arxiv.org/abs/2011.13456">Yang Song's paper</a>`

Lines changed: 7 additions & 2 deletions

Original file line number	Diff line number	Diff line change
`@@ -1479,9 +1479,14 @@ def init_conv_(self, conv):`
`1479`	`1479`	`defforward(self,x):`
`1480`	`1480`	`returnself.net(x)`
`1481`	`1481`
`1482`		`-defDownsample(dim,*,dim_out=None):`
	`1482`	`+defDownsample(dim,dim_out=None):`
	`1483`	`+# https://arxiv.org/abs/2208.03641 shows this is the most optimal way to downsample`
	`1484`	`+# named SP-conv in the paper, but basically a pixel unshuffle`
`1483`	`1485`	`dim_out=default(dim_out,dim)`
`1484`		`-returnnn.Conv2d(dim,dim_out,4,2,1)`
	`1486`	`+returnnn.Sequential(`
	`1487`	`+Rearrange('b c (h s1) (w s2) -> b (c s1 s2) h w',s1=2,s2=2),`
	`1488`	`+nn.Conv2d(dim*4,dim_out,1)`
	`1489`	`+ )`
`1485`	`1490`
`1486`	`1491`	`classWeightStandardizedConv2d(nn.Conv2d):`
`1487`	`1492`	`"""`

Lines changed: 1 addition & 1 deletion

Original file line number	Diff line number	Diff line change
`@@ -519,7 +519,7 @@ def __init__(`
`519`	`519`	`clip=decoder.clip`
`520`	`520`	`clip.to(precision_type)`
`521`	`521`
`522`		`-decoder,train_dataloader,optimizers=list(self.accelerator.prepare(decoder,dataloaders['train'],optimizers))`
	`522`	`+decoder,optimizers=list(self.accelerator.prepare(decoder,optimizers))`
`523`	`523`
`524`	`524`	`self.decoder=decoder`
`525`	`525`

Lines changed: 1 addition & 1 deletion

Comments

(0)