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

Fix [core/GLIGEN]: TypeError when iterating over 0-d tensor with In-painting mode when EulerAncestralDiscreteScheduler is used (#5305)

Browse Source 
* fix(gligen_inpaint_pipeline): 🐛 Wrap the timestep() 0-d tensor in a list to convert to 1-d tensor. This avoids the TypeError caused by trying to directly iterate over a 0-dimensional tensor in the denoising stage

* test(gligen/gligen_text_image): unit test using the EulerAncestralDiscreteScheduler

---------

Co-authored-by: zhen-hao.chu <zhen-hao.chu@vitrox.com>
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
This commit is contained in:
chuzh
2023-10-09 15:54:01 +08:00
committed by GitHub
parent 0513a8cfd8
commit 6bd55b54bc
5 changed files with 47 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
examples/custom_diffusion/train_custom_diffusion.py
View File

@@ -207,7 +207,7 @@ class CustomDiffusionDataset(Dataset):
with open(concept["class_prompt"], "r") as f:
class_prompt = f.read().splitlines()
class_img_path = [(x, y) for (x, y) in zip(class_images_path, class_prompt)]
class_img_path = list(zip(class_images_path, class_prompt))
self.class_images_path.extend(class_img_path[:num_class_images])
random.shuffle(self.instance_images_path)

4
src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_gligen.py
View File

@@ -803,7 +803,9 @@ class StableDiffusionGLIGENPipeline(DiffusionPipeline):
if gligen_inpaint_image is not None:
gligen_inpaint_latent_with_noise = (
self.scheduler.add_noise(gligen_inpaint_latent, torch.randn_like(gligen_inpaint_latent), t)
self.scheduler.add_noise(
gligen_inpaint_latent, torch.randn_like(gligen_inpaint_latent), torch.tensor([t])
)
.expand(latents.shape[0], -1, -1, -1)
.clone()
)

4
src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_gligen_text_image.py
View File

@@ -965,7 +965,9 @@ class StableDiffusionGLIGENTextImagePipeline(DiffusionPipeline):
if gligen_inpaint_image is not None:
gligen_inpaint_latent_with_noise = (
self.scheduler.add_noise(gligen_inpaint_latent, torch.randn_like(gligen_inpaint_latent), t)
self.scheduler.add_noise(
gligen_inpaint_latent, torch.randn_like(gligen_inpaint_latent), torch.tensor([t])
)
.expand(latents.shape[0], -1, -1, -1)
.clone()
)

21
tests/pipelines/stable_diffusion/test_stable_diffusion_gligen.py
View File

@@ -22,6 +22,7 @@ from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer
from diffusers import (
AutoencoderKL,
DDIMScheduler,
EulerAncestralDiscreteScheduler,
StableDiffusionGLIGENPipeline,
UNet2DConditionModel,
)
@@ -120,7 +121,7 @@ class GligenPipelineFastTests(
}
return inputs
def test_gligen(self):
def test_stable_diffusion_gligen_default_case(self):
device = "cpu" # ensure determinism for the device-dependent torch.Generator
components = self.get_dummy_components()
sd_pipe = StableDiffusionGLIGENPipeline(**components)
@@ -136,6 +137,24 @@ class GligenPipelineFastTests(
assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
def test_stable_diffusion_gligen_k_euler_ancestral(self):
device = "cpu" # ensure determinism for the device-dependent torch.Generator
components = self.get_dummy_components()
sd_pipe = StableDiffusionGLIGENPipeline(**components)
sd_pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(sd_pipe.scheduler.config)
sd_pipe = sd_pipe.to(device)
sd_pipe.set_progress_bar_config(disable=None)
inputs = self.get_dummy_inputs(device)
output = sd_pipe(**inputs)
image = output.images
image_slice = image[0, -3:, -3:, -1]
assert image.shape == (1, 64, 64, 3)
expected_slice = np.array([0.425, 0.494, 0.429, 0.469, 0.525, 0.417, 0.533, 0.5, 0.47])
assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
def test_attention_slicing_forward_pass(self):
super().test_attention_slicing_forward_pass(expected_max_diff=3e-3)

21
tests/pipelines/stable_diffusion/test_stable_diffusion_gligen_text_image.py
View File

@@ -29,6 +29,7 @@ from transformers import (
from diffusers import (
AutoencoderKL,
DDIMScheduler,
EulerAncestralDiscreteScheduler,
StableDiffusionGLIGENTextImagePipeline,
UNet2DConditionModel,
)
@@ -150,7 +151,7 @@ class GligenTextImagePipelineFastTests(
}
return inputs
def test_gligen(self):
def test_stable_diffusion_gligen_text_image_default_case(self):
device = "cpu" # ensure determinism for the device-dependent torch.Generator
components = self.get_dummy_components()
sd_pipe = StableDiffusionGLIGENTextImagePipeline(**components)
@@ -166,6 +167,24 @@ class GligenTextImagePipelineFastTests(
assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
def test_stable_diffusion_gligen_k_euler_ancestral(self):
device = "cpu" # ensure determinism for the device-dependent torch.Generator
components = self.get_dummy_components()
sd_pipe = StableDiffusionGLIGENTextImagePipeline(**components)
sd_pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(sd_pipe.scheduler.config)
sd_pipe = sd_pipe.to(device)
sd_pipe.set_progress_bar_config(disable=None)
inputs = self.get_dummy_inputs(device)
image = sd_pipe(**inputs).images
image_slice = image[0, -3:, -3:, -1]
assert image.shape == (1, 64, 64, 3)
expected_slice = np.array([0.425, 0.494, 0.429, 0.469, 0.525, 0.417, 0.533, 0.5, 0.47])
assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
def test_attention_slicing_forward_pass(self):
super().test_attention_slicing_forward_pass(expected_max_diff=3e-3)