AI/diffusers
1
0
Fork 0
mirror of https://github.com/huggingface/diffusers.git synced 2026-01-27 17:22:53 +03:00
Code Activity

config fixes (#3060)

Browse Source
This commit is contained in:
Will Berman
2023-04-11 09:54:50 -07:00
committed by Daniel Gu
parent 1336668f0e
commit 2bc0a15818
4 changed files with 15 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
examples/community/sd_text2img_k_diffusion.py
View File

@@ -105,7 +105,7 @@ class StableDiffusionPipeline(DiffusionPipeline):
)
model = ModelWrapper(unet, scheduler.alphas_cumprod)
if scheduler.prediction_type == "v_prediction":
if scheduler.config.prediction_type == "v_prediction":
self.k_diffusion_model = CompVisVDenoiser(model)
else:
self.k_diffusion_model = CompVisDenoiser(model)

6
src/diffusers/pipelines/audio_diffusion/pipeline_audio_diffusion.py
View File

@@ -60,9 +60,9 @@ class AudioDiffusionPipeline(DiffusionPipeline):
input_module = self.vqvae if self.vqvae is not None else self.unet
# For backwards compatibility
sample_size = (
(input_module.sample_size, input_module.sample_size)
if type(input_module.sample_size) == int
else input_module.sample_size
(input_module.config.sample_size, input_module.config.sample_size)
if type(input_module.config.sample_size) == int
else input_module.config.sample_size
)
return sample_size

2
src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_k_diffusion.py
View File

@@ -113,7 +113,7 @@ class StableDiffusionKDiffusionPipeline(DiffusionPipeline, TextualInversionLoade
self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
model = ModelWrapper(unet, scheduler.alphas_cumprod)
if scheduler.prediction_type == "v_prediction":
if scheduler.config.prediction_type == "v_prediction":
self.k_diffusion_model = CompVisVDenoiser(model)
else:
self.k_diffusion_model = CompVisDenoiser(model)

17
tests/pipelines/audio_diffusion/test_audio_diffusion.py
View File

@@ -115,8 +115,11 @@ class PipelineFastTests(unittest.TestCase):
output = pipe(generator=generator, steps=4, return_dict=False)
image_from_tuple = output[0][0]
assert audio.shape == (1, (self.dummy_unet.sample_size[1] - 1) * mel.hop_length)
assert image.height == self.dummy_unet.sample_size[0] and image.width == self.dummy_unet.sample_size[1]
assert audio.shape == (1, (self.dummy_unet.config.sample_size[1] - 1) * mel.hop_length)
assert (
image.height == self.dummy_unet.config.sample_size[0]
and image.width == self.dummy_unet.config.sample_size[1]
)
image_slice = np.frombuffer(image.tobytes(), dtype="uint8")[:10]
image_from_tuple_slice = np.frombuffer(image_from_tuple.tobytes(), dtype="uint8")[:10]
expected_slice = np.array([69, 255, 255, 255, 0, 0, 77, 181, 12, 127])
@@ -133,14 +136,14 @@ class PipelineFastTests(unittest.TestCase):
pipe.set_progress_bar_config(disable=None)
np.random.seed(0)
raw_audio = np.random.uniform(-1, 1, ((dummy_vqvae_and_unet[0].sample_size[1] - 1) * mel.hop_length,))
raw_audio = np.random.uniform(-1, 1, ((dummy_vqvae_and_unet[0].config.sample_size[1] - 1) * mel.hop_length,))
generator = torch.Generator(device=device).manual_seed(42)
output = pipe(raw_audio=raw_audio, generator=generator, start_step=5, steps=10)
image = output.images[0]
assert (
image.height == self.dummy_vqvae_and_unet[0].sample_size[0]
and image.width == self.dummy_vqvae_and_unet[0].sample_size[1]
image.height == self.dummy_vqvae_and_unet[0].config.sample_size[0]
and image.width == self.dummy_vqvae_and_unet[0].config.sample_size[1]
)
image_slice = np.frombuffer(image.tobytes(), dtype="uint8")[:10]
expected_slice = np.array([120, 117, 110, 109, 138, 167, 138, 148, 132, 121])
@@ -183,8 +186,8 @@ class PipelineIntegrationTests(unittest.TestCase):
audio = output.audios[0]
image = output.images[0]
assert audio.shape == (1, (pipe.unet.sample_size[1] - 1) * pipe.mel.hop_length)
assert image.height == pipe.unet.sample_size[0] and image.width == pipe.unet.sample_size[1]
assert audio.shape == (1, (pipe.unet.config.sample_size[1] - 1) * pipe.mel.hop_length)
assert image.height == pipe.unet.config.sample_size[0] and image.width == pipe.unet.config.sample_size[1]
image_slice = np.frombuffer(image.tobytes(), dtype="uint8")[:10]
expected_slice = np.array([151, 167, 154, 144, 122, 134, 121, 105, 70, 26])