Skip to content

[Tests] reduce the model sizes in the SD fast tests #7580

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
yiyixuxu merged 9 commits into from
Apr 10, 2024
Merged

[Tests] reduce the model sizes in the SD fast tests #7580

Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
0d3e905
give it a shot.
sayakpaul Apr 5, 2024
d005fcb
print.
sayakpaul Apr 5, 2024
8762d6e
correct assertion.
sayakpaul Apr 5, 2024
eba1ea2
gather results from the rest of the tests.
sayakpaul Apr 5, 2024
a2a857c
change the assertion values where needed.
sayakpaul Apr 5, 2024
dce8042
remove print statements.
sayakpaul Apr 5, 2024
a432c81
Merge branch 'main' into tests/reduce-block-sizes-in-pipelines
sayakpaul Apr 5, 2024
86caf22
Merge branch 'main' into tests/reduce-block-sizes-in-pipelines
sayakpaul Apr 7, 2024
cb0f778
Merge branch 'main' into tests/reduce-block-sizes-in-pipelines
sayakpaul Apr 10, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
36 changes: 16 additions & 20 deletions tests/pipelines/stable_diffusion/test_stable_diffusion.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -124,6 +124,8 @@ class StableDiffusionPipelineFastTests(
callback_cfg_params = TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS

def get_dummy_components(self, time_cond_proj_dim=None):
cross_attention_dim = 8

torch.manual_seed(0)
unet = UNet2DConditionModel(
block_out_channels=(4, 8),
Expand All @@ -134,7 +136,7 @@ def get_dummy_components(self, time_cond_proj_dim=None):
out_channels=4,
down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
cross_attention_dim=32,
cross_attention_dim=cross_attention_dim,
norm_num_groups=2,
)
scheduler = DDIMScheduler(
Expand All @@ -158,11 +160,11 @@ def get_dummy_components(self, time_cond_proj_dim=None):
text_encoder_config = CLIPTextConfig(
bos_token_id=0,
eos_token_id=2,
hidden_size=32,
intermediate_size=64,
hidden_size=cross_attention_dim,
intermediate_size=16,
layer_norm_eps=1e-05,
num_attention_heads=8,
num_hidden_layers=3,
num_attention_heads=2,
num_hidden_layers=2,
pad_token_id=1,
vocab_size=1000,
)
Expand Down Expand Up @@ -210,7 +212,7 @@ def test_stable_diffusion_ddim(self):
image_slice = image[0, -3:, -3:, -1]

assert image.shape == (1, 64, 64, 3)
expected_slice = np.array([0.3203, 0.4555, 0.4711, 0.3505, 0.3973, 0.4650, 0.5137, 0.3392, 0.4045])
expected_slice = np.array([0.1763, 0.4776, 0.4986, 0.2566, 0.3802, 0.4596, 0.5363, 0.3277, 0.3949])

assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

Expand All @@ -230,7 +232,7 @@ def test_stable_diffusion_lcm(self):
image_slice = image[0, -3:, -3:, -1]

assert image.shape == (1, 64, 64, 3)
expected_slice = np.array([0.3454, 0.5349, 0.5185, 0.2808, 0.4509, 0.4612, 0.4655, 0.3601, 0.4315])
expected_slice = np.array([0.2368, 0.4900, 0.5019, 0.2723, 0.4473, 0.4578, 0.4551, 0.3532, 0.4133])

assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

Expand All @@ -252,7 +254,7 @@ def test_stable_diffusion_lcm_custom_timesteps(self):
image_slice = image[0, -3:, -3:, -1]

assert image.shape == (1, 64, 64, 3)
expected_slice = np.array([0.3454, 0.5349, 0.5185, 0.2808, 0.4509, 0.4612, 0.4655, 0.3601, 0.4315])
expected_slice = np.array([0.2368, 0.4900, 0.5019, 0.2723, 0.4473, 0.4578, 0.4551, 0.3532, 0.4133])

assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

Expand Down Expand Up @@ -371,12 +373,6 @@ def test_stable_diffusion_prompt_embeds_with_plain_negative_prompt_list(self):

assert np.abs(image_slice_1.flatten() - image_slice_2.flatten()).max() < 1e-4

def test_ip_adapter_single(self):
expected_pipe_slice = None
if torch_device == "cpu":
expected_pipe_slice = np.array([0.3203, 0.4555, 0.4711, 0.3505, 0.3973, 0.4650, 0.5137, 0.3392, 0.4045])
return super().test_ip_adapter_single(expected_pipe_slice=expected_pipe_slice)
Comment on lines -374 to -378
Copy link
Member Author

@sayakpaul sayakpaul Apr 5, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

No point in keeping this test this way because we are reducing the model sizes dramatically so this won't contribute too much. LMK if that doesn't make sense.


def test_stable_diffusion_ddim_factor_8(self):
device = "cpu" # ensure determinism for the device-dependent torch.Generator

Expand All @@ -392,7 +388,7 @@ def test_stable_diffusion_ddim_factor_8(self):
image_slice = image[0, -3:, -3:, -1]

assert image.shape == (1, 136, 136, 3)
expected_slice = np.array([0.4346, 0.5621, 0.5016, 0.3926, 0.4533, 0.4134, 0.5625, 0.5632, 0.5265])
expected_slice = np.array([0.4720, 0.5426, 0.5160, 0.3961, 0.4696, 0.4296, 0.5738, 0.5888, 0.5481])

assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

Expand All @@ -410,7 +406,7 @@ def test_stable_diffusion_pndm(self):
image_slice = image[0, -3:, -3:, -1]

assert image.shape == (1, 64, 64, 3)
expected_slice = np.array([0.3411, 0.5032, 0.4704, 0.3135, 0.4323, 0.4740, 0.5150, 0.3498, 0.4022])
expected_slice = np.array([0.1941, 0.4748, 0.4880, 0.2222, 0.4221, 0.4545, 0.5604, 0.3488, 0.3902])

assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

Expand Down Expand Up @@ -450,7 +446,7 @@ def test_stable_diffusion_k_lms(self):
image_slice = image[0, -3:, -3:, -1]

assert image.shape == (1, 64, 64, 3)
expected_slice = np.array([0.3149, 0.5246, 0.4796, 0.3218, 0.4469, 0.4729, 0.5151, 0.3597, 0.3954])
expected_slice = np.array([0.2681, 0.4785, 0.4857, 0.2426, 0.4473, 0.4481, 0.5610, 0.3676, 0.3855])

assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

Expand All @@ -469,7 +465,7 @@ def test_stable_diffusion_k_euler_ancestral(self):
image_slice = image[0, -3:, -3:, -1]

assert image.shape == (1, 64, 64, 3)
expected_slice = np.array([0.3151, 0.5243, 0.4794, 0.3217, 0.4468, 0.4728, 0.5152, 0.3598, 0.3954])
expected_slice = np.array([0.2682, 0.4782, 0.4855, 0.2424, 0.4472, 0.4479, 0.5612, 0.3676, 0.3854])

assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

Expand All @@ -488,7 +484,7 @@ def test_stable_diffusion_k_euler(self):
image_slice = image[0, -3:, -3:, -1]

assert image.shape == (1, 64, 64, 3)
expected_slice = np.array([0.3149, 0.5246, 0.4796, 0.3218, 0.4469, 0.4729, 0.5151, 0.3597, 0.3954])
expected_slice = np.array([0.2681, 0.4785, 0.4857, 0.2426, 0.4473, 0.4481, 0.5610, 0.3676, 0.3855])

assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

Expand Down Expand Up @@ -560,7 +556,7 @@ def test_stable_diffusion_negative_prompt(self):
image_slice = image[0, -3:, -3:, -1]

assert image.shape == (1, 64, 64, 3)
expected_slice = np.array([0.3458, 0.5120, 0.4800, 0.3116, 0.4348, 0.4802, 0.5237, 0.3467, 0.3991])
expected_slice = np.array([0.1907, 0.4709, 0.4858, 0.2224, 0.4223, 0.4539, 0.5606, 0.3489, 0.3900])

assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

Expand Down