Finegrain/refiners
1
0
Fork 0
mirror of https://github.com/finegrain-ai/refiners.git synced 2024-11-22 06:08:46 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

freeu: add one more test for identity scales

Browse source 
It should act as a NOP when [1.0, 1.0] is used for backbone and skip
scales.
This commit is contained in:
Cédric Deltheil 2023-12-01 12:32:52 +01:00 committed by Cédric Deltheil
parent 761678d9a5
commit b306c7db1b
1 changed files with 26 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

26
tests/foundationals/latent_diffusion/test_freeu.py
View file

@ -1,9 +1,11 @@
from typing import Iterator from typing import Iterator
import pytest import pytest
import torch
from refiners.foundationals.latent_diffusion import SD1UNet, SDXLUNet from refiners.foundationals.latent_diffusion import SD1UNet, SDXLUNet
from refiners.foundationals.latent_diffusion.freeu import SDFreeUAdapter, FreeUResidualConcatenator from refiners.foundationals.latent_diffusion.freeu import SDFreeUAdapter, FreeUResidualConcatenator
from refiners.fluxion import manual_seed
@pytest.fixture(scope="module", params=[True, False]) @pytest.fixture(scope="module", params=[True, False])
@ -39,3 +41,27 @@ def test_freeu_adapter_too_many_scales(unet: SD1UNet | SDXLUNet) -> None:
def test_freeu_adapter_inconsistent_scales(unet: SD1UNet | SDXLUNet) -> None: def test_freeu_adapter_inconsistent_scales(unet: SD1UNet | SDXLUNet) -> None:
with pytest.raises(AssertionError): with pytest.raises(AssertionError):
SDFreeUAdapter(unet, backbone_scales=[1.2, 1.2], skip_scales=[0.9, 0.9, 0.9]) SDFreeUAdapter(unet, backbone_scales=[1.2, 1.2], skip_scales=[0.9, 0.9, 0.9])
def test_freeu_identity_scales() -> None:
manual_seed(0)
text_embedding = torch.randn(1, 77, 768)
timestep = torch.randint(0, 999, size=(1, 1))
x = torch.randn(1, 4, 32, 32)
unet = SD1UNet(in_channels=4)
unet.set_clip_text_embedding(clip_text_embedding=text_embedding) # not flushed between forward-s
with torch.no_grad():
unet.set_timestep(timestep=timestep)
y_1 = unet(x.clone())
freeu = SDFreeUAdapter(unet, backbone_scales=[1.0, 1.0], skip_scales=[1.0, 1.0])
freeu.inject()
with torch.no_grad():
unet.set_timestep(timestep=timestep)
y_2 = unet(x.clone())
# The FFT -> inverse FFT sequence (skip features) introduces small numerical differences
assert torch.allclose(y_1, y_2, atol=1e-5)