refiners/tests/foundationals/latent_diffusion/test_solvers.py

330 lines
12 KiB
Python
Raw Permalink Normal View History

2024-09-06 10:56:24 +00:00
import itertools
2023-08-04 13:28:41 +00:00
from typing import cast
2023-09-21 09:47:11 +00:00
from warnings import warn
import pytest
2024-09-25 10:44:35 +00:00
import torch
from torch import Tensor, device as Device
2023-12-04 14:08:34 +00:00
from refiners.fluxion import manual_seed
from refiners.foundationals.latent_diffusion.solvers import (
DDIM,
DDPM,
DPMSolver,
Euler,
FrankenSolver,
LCMSolver,
ModelPredictionType,
NoiseSchedule,
Solver,
SolverParams,
TimestepSpacing,
)
def test_ddpm_diffusers():
from diffusers import DDPMScheduler # type: ignore
diffusers_scheduler = DDPMScheduler(beta_schedule="scaled_linear", beta_start=0.00085, beta_end=0.012)
diffusers_scheduler.set_timesteps(1000)
solver = DDPM(num_inference_steps=1000)
2024-09-25 10:44:35 +00:00
assert torch.equal(diffusers_scheduler.timesteps, solver.timesteps)
2023-08-04 13:28:41 +00:00
2024-09-06 10:56:24 +00:00
@pytest.mark.parametrize(
"n_steps, last_step_first_order, sde_variance, use_karras_sigmas",
list(itertools.product([5, 30], [False, True], [0.0, 1.0], [False, True])),
)
def test_dpm_solver_diffusers(n_steps: int, last_step_first_order: bool, sde_variance: float, use_karras_sigmas: bool):
from diffusers import DPMSolverMultistepScheduler as DiffuserScheduler # type: ignore
manual_seed(0)
diffusers_scheduler = DiffuserScheduler(
beta_schedule="scaled_linear",
beta_start=0.00085,
beta_end=0.012,
lower_order_final=False,
euler_at_final=last_step_first_order,
final_sigmas_type="sigma_min", # default before Diffusers 0.26.0
2024-09-06 10:56:24 +00:00
algorithm_type="sde-dpmsolver++" if sde_variance == 1.0 else "dpmsolver++",
use_karras_sigmas=use_karras_sigmas,
)
diffusers_scheduler.set_timesteps(n_steps)
solver = DPMSolver(
num_inference_steps=n_steps,
last_step_first_order=last_step_first_order,
2024-09-06 10:56:24 +00:00
params=SolverParams(
sde_variance=sde_variance,
sigma_schedule=NoiseSchedule.KARRAS if use_karras_sigmas else None,
),
)
2024-09-25 10:44:35 +00:00
assert torch.equal(solver.timesteps, diffusers_scheduler.timesteps)
2024-09-25 10:44:35 +00:00
sample = torch.randn(1, 3, 32, 32)
predicted_noise = torch.randn(1, 3, 32, 32)
manual_seed(37)
diffusers_outputs: list[Tensor] = [
cast(Tensor, diffusers_scheduler.step(predicted_noise, timestep, sample).prev_sample) # type: ignore
for timestep in diffusers_scheduler.timesteps
]
manual_seed(37)
refiners_outputs = [solver(x=sample, predicted_noise=predicted_noise, step=step) for step in range(n_steps)]
if use_karras_sigmas:
atol = 1e-4
elif sde_variance == 1.0:
atol = 1e-6
else:
atol = 1e-8
for step, (diffusers_output, refiners_output) in enumerate(zip(diffusers_outputs, refiners_outputs)):
2024-09-25 10:44:35 +00:00
assert torch.allclose(diffusers_output, refiners_output, rtol=0.01, atol=atol), f"outputs differ at step {step}"
2023-12-12 16:18:29 +00:00
def test_ddim_diffusers():
2023-08-04 13:28:41 +00:00
from diffusers import DDIMScheduler # type: ignore
2023-12-12 16:18:29 +00:00
manual_seed(0)
2023-08-04 13:28:41 +00:00
diffusers_scheduler = DDIMScheduler(
beta_end=0.012,
beta_schedule="scaled_linear",
beta_start=0.00085,
num_train_timesteps=1000,
steps_offset=1,
clip_sample=False,
)
diffusers_scheduler.set_timesteps(30)
solver = DDIM(num_inference_steps=30)
2024-09-25 10:44:35 +00:00
assert torch.equal(solver.timesteps, diffusers_scheduler.timesteps)
2023-08-04 13:28:41 +00:00
2024-09-25 10:44:35 +00:00
sample = torch.randn(1, 4, 32, 32)
predicted_noise = torch.randn(1, 4, 32, 32)
2023-08-04 13:28:41 +00:00
for step, timestep in enumerate(diffusers_scheduler.timesteps):
diffusers_output = cast(Tensor, diffusers_scheduler.step(predicted_noise, timestep, sample).prev_sample) # type: ignore
refiners_output = solver(x=sample, predicted_noise=predicted_noise, step=step)
2023-08-04 13:28:41 +00:00
2024-09-25 10:44:35 +00:00
assert torch.allclose(diffusers_output, refiners_output, rtol=0.01), f"outputs differ at step {step}"
2023-09-21 09:47:11 +00:00
@pytest.mark.parametrize("model_prediction_type", [ModelPredictionType.NOISE, ModelPredictionType.SAMPLE])
def test_euler_diffusers(model_prediction_type: ModelPredictionType):
2024-01-10 10:45:19 +00:00
from diffusers import EulerDiscreteScheduler # type: ignore
2024-01-10 10:32:40 +00:00
manual_seed(0)
diffusers_prediction_type = "epsilon" if model_prediction_type == ModelPredictionType.NOISE else "sample"
2024-01-10 10:32:40 +00:00
diffusers_scheduler = EulerDiscreteScheduler(
beta_end=0.012,
beta_schedule="scaled_linear",
beta_start=0.00085,
num_train_timesteps=1000,
steps_offset=1,
timestep_spacing="linspace",
use_karras_sigmas=False,
prediction_type=diffusers_prediction_type,
2024-01-10 10:32:40 +00:00
)
diffusers_scheduler.set_timesteps(30)
solver = Euler(num_inference_steps=30, params=SolverParams(model_prediction_type=model_prediction_type))
2024-09-25 10:44:35 +00:00
assert torch.equal(solver.timesteps, diffusers_scheduler.timesteps)
2024-01-10 10:32:40 +00:00
2024-09-25 10:44:35 +00:00
sample = torch.randn(1, 4, 32, 32)
predicted_noise = torch.randn(1, 4, 32, 32)
2024-01-10 10:32:40 +00:00
ref_init_noise_sigma = diffusers_scheduler.init_noise_sigma # type: ignore
assert isinstance(ref_init_noise_sigma, Tensor)
2024-09-25 10:44:35 +00:00
assert torch.isclose(ref_init_noise_sigma, solver.init_noise_sigma), "init_noise_sigma differ"
2024-01-10 10:32:40 +00:00
for step, timestep in enumerate(diffusers_scheduler.timesteps):
diffusers_output = cast(Tensor, diffusers_scheduler.step(predicted_noise, timestep, sample).prev_sample) # type: ignore
refiners_output = solver(x=sample, predicted_noise=predicted_noise, step=step)
2024-01-10 10:32:40 +00:00
2024-09-25 10:44:35 +00:00
assert torch.allclose(diffusers_output, refiners_output, rtol=0.02), f"outputs differ at step {step}"
2024-01-10 10:32:40 +00:00
def test_franken_diffusers():
from diffusers import EulerDiscreteScheduler # type: ignore
manual_seed(0)
params = {
"beta_end": 0.012,
"beta_schedule": "scaled_linear",
"beta_start": 0.00085,
"num_train_timesteps": 1000,
"steps_offset": 1,
"timestep_spacing": "linspace",
"use_karras_sigmas": False,
}
diffusers_scheduler = EulerDiscreteScheduler(**params) # type: ignore
diffusers_scheduler.set_timesteps(30)
diffusers_scheduler_2 = EulerDiscreteScheduler(**params) # type: ignore
solver = FrankenSolver(lambda: diffusers_scheduler_2, num_inference_steps=30)
2024-09-25 10:44:35 +00:00
assert torch.equal(solver.timesteps, diffusers_scheduler.timesteps)
2024-09-25 10:44:35 +00:00
sample = torch.randn(1, 4, 32, 32)
predicted_noise = torch.randn(1, 4, 32, 32)
ref_init_noise_sigma = diffusers_scheduler.init_noise_sigma # type: ignore
assert isinstance(ref_init_noise_sigma, Tensor)
2024-09-25 10:44:35 +00:00
init_noise_sigma = solver.scale_model_input(torch.tensor(1), step=-1)
assert torch.equal(ref_init_noise_sigma, init_noise_sigma), "init_noise_sigma differ"
for step, timestep in enumerate(diffusers_scheduler.timesteps):
diffusers_output = cast(Tensor, diffusers_scheduler.step(predicted_noise, timestep, sample).prev_sample) # type: ignore
refiners_output = solver(x=sample, predicted_noise=predicted_noise, step=step)
2024-09-25 10:44:35 +00:00
assert torch.equal(diffusers_output, refiners_output), f"outputs differ at step {step}"
def test_lcm_diffusers():
from diffusers import LCMScheduler # type: ignore
manual_seed(0)
# LCMScheduler is stochastic, make sure we use identical generators
2024-09-25 10:44:35 +00:00
diffusers_generator = torch.Generator().manual_seed(42)
refiners_generator = torch.Generator().manual_seed(42)
diffusers_scheduler = LCMScheduler()
diffusers_scheduler.set_timesteps(4)
solver = LCMSolver(num_inference_steps=4)
2024-09-25 10:44:35 +00:00
assert torch.equal(solver.timesteps, diffusers_scheduler.timesteps)
2024-09-25 10:44:35 +00:00
sample = torch.randn(1, 4, 32, 32)
predicted_noise = torch.randn(1, 4, 32, 32)
for step, timestep in enumerate(diffusers_scheduler.timesteps):
alpha_prod_t = diffusers_scheduler.alphas_cumprod[timestep]
diffusers_noise_ratio = (1 - alpha_prod_t).sqrt()
diffusers_scale_factor = alpha_prod_t.sqrt()
refiners_scale_factor = solver.cumulative_scale_factors[timestep]
refiners_noise_ratio = solver.noise_std[timestep]
assert refiners_scale_factor == diffusers_scale_factor
assert refiners_noise_ratio == diffusers_noise_ratio
d_out = diffusers_scheduler.step(predicted_noise, timestep, sample, generator=diffusers_generator) # type: ignore
diffusers_output = cast(Tensor, d_out.prev_sample) # type: ignore
refiners_output = solver(
x=sample,
predicted_noise=predicted_noise,
step=step,
generator=refiners_generator,
)
2024-09-25 10:44:35 +00:00
assert torch.allclose(refiners_output, diffusers_output, rtol=0.01), f"outputs differ at step {step}"
def test_solver_remove_noise():
from diffusers import DDIMScheduler # type: ignore
2023-12-12 16:18:29 +00:00
manual_seed(0)
diffusers_scheduler = DDIMScheduler(
beta_end=0.012,
beta_schedule="scaled_linear",
beta_start=0.00085,
num_train_timesteps=1000,
steps_offset=1,
clip_sample=False,
)
diffusers_scheduler.set_timesteps(30)
solver = DDIM(num_inference_steps=30)
2024-09-25 10:44:35 +00:00
sample = torch.randn(1, 4, 32, 32)
noise = torch.randn(1, 4, 32, 32)
for step, timestep in enumerate(diffusers_scheduler.timesteps):
2024-01-10 10:41:47 +00:00
diffusers_output = cast(Tensor, diffusers_scheduler.step(noise, timestep, sample).pred_original_sample) # type: ignore
refiners_output = solver.remove_noise(x=sample, noise=noise, step=step)
2024-09-25 10:44:35 +00:00
assert torch.allclose(diffusers_output, refiners_output, rtol=0.01), f"outputs differ at step {step}"
def test_solver_device(test_device: Device):
2023-09-21 09:47:11 +00:00
if test_device.type == "cpu":
warn("not running on CPU, skipping")
pytest.skip()
scheduler = DDIM(num_inference_steps=30, device=test_device)
2024-09-25 10:44:35 +00:00
x = torch.randn(1, 4, 32, 32, device=test_device)
noise = torch.randn(1, 4, 32, 32, device=test_device)
noised = scheduler.add_noise(x, noise, scheduler.first_inference_step)
2023-09-21 09:47:11 +00:00
assert noised.device == test_device
2024-01-30 16:47:06 +00:00
def test_solver_add_noise(test_device: Device):
scheduler = DDIM(num_inference_steps=30, device=test_device)
2024-09-25 10:44:35 +00:00
latent = torch.randn(1, 4, 32, 32, device=test_device)
noise = torch.randn(1, 4, 32, 32, device=test_device)
noised = scheduler.add_noise(
x=latent,
noise=noise,
step=0,
)
noised_double = scheduler.add_noise(
x=latent.repeat(2, 1, 1, 1),
noise=noise.repeat(2, 1, 1, 1),
step=[0, 0],
)
2024-09-25 10:44:35 +00:00
assert torch.allclose(noised, noised_double[0])
assert torch.allclose(noised, noised_double[1])
2024-01-30 16:47:06 +00:00
@pytest.mark.parametrize("noise_schedule", [NoiseSchedule.UNIFORM, NoiseSchedule.QUADRATIC, NoiseSchedule.KARRAS])
def test_solver_noise_schedules(noise_schedule: NoiseSchedule, test_device: Device):
scheduler = DDIM(
num_inference_steps=30,
params=SolverParams(noise_schedule=noise_schedule),
device=test_device,
)
2024-01-30 16:47:06 +00:00
assert len(scheduler.scale_factors) == 1000
assert scheduler.scale_factors[0] == 1 - scheduler.params.initial_diffusion_rate
assert scheduler.scale_factors[-1] == 1 - scheduler.params.final_diffusion_rate
def test_solver_timestep_spacing():
# Tests we get the results from [[arXiv:2305.08891] Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/abs/2305.08891) table 2.
linspace_int = Solver.generate_timesteps(
spacing=TimestepSpacing.LINSPACE_ROUNDED,
num_inference_steps=10,
num_train_timesteps=1000,
offset=1,
)
2024-09-25 10:44:35 +00:00
assert torch.equal(linspace_int, torch.tensor([1000, 889, 778, 667, 556, 445, 334, 223, 112, 1]))
leading = Solver.generate_timesteps(
spacing=TimestepSpacing.LEADING,
num_inference_steps=10,
num_train_timesteps=1000,
offset=1,
)
2024-09-25 10:44:35 +00:00
assert torch.equal(leading, torch.tensor([901, 801, 701, 601, 501, 401, 301, 201, 101, 1]))
trailing = Solver.generate_timesteps(
spacing=TimestepSpacing.TRAILING,
num_inference_steps=10,
num_train_timesteps=1000,
offset=1,
)
2024-09-25 10:44:35 +00:00
assert torch.equal(trailing, torch.tensor([1000, 900, 800, 700, 600, 500, 400, 300, 200, 100]))
def test_dpm_bfloat16(test_device: Device):
if test_device.type == "cpu":
warn("not running on CPU, skipping")
pytest.skip()
n_steps = 5
manual_seed(0)
solver_f32 = DPMSolver(num_inference_steps=n_steps, dtype=torch.float32)
solver_bf16 = DPMSolver(num_inference_steps=n_steps, dtype=torch.bfloat16)
assert torch.equal(solver_bf16.timesteps, solver_f32.timesteps)