File size: 12,326 Bytes
3a25a0a |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 |
import tempfile
from typing import Dict, List, Tuple
import torch
from diffusers import LCMScheduler
from diffusers.utils.testing_utils import torch_device
from .test_schedulers import SchedulerCommonTest
class LCMSchedulerTest(SchedulerCommonTest):
scheduler_classes = (LCMScheduler,)
forward_default_kwargs = (("num_inference_steps", 10),)
def get_scheduler_config(self, **kwargs):
config = {
"num_train_timesteps": 1000,
"beta_start": 0.00085,
"beta_end": 0.0120,
"beta_schedule": "scaled_linear",
"prediction_type": "epsilon",
}
config.update(**kwargs)
return config
@property
def default_valid_timestep(self):
kwargs = dict(self.forward_default_kwargs)
num_inference_steps = kwargs.pop("num_inference_steps", None)
scheduler_config = self.get_scheduler_config()
scheduler = self.scheduler_classes[0](**scheduler_config)
scheduler.set_timesteps(num_inference_steps)
timestep = scheduler.timesteps[-1]
return timestep
def test_timesteps(self):
for timesteps in [100, 500, 1000]:
# 0 is not guaranteed to be in the timestep schedule, but timesteps - 1 is
self.check_over_configs(time_step=timesteps - 1, num_train_timesteps=timesteps)
def test_betas(self):
for beta_start, beta_end in zip([0.0001, 0.001, 0.01, 0.1], [0.002, 0.02, 0.2, 2]):
self.check_over_configs(time_step=self.default_valid_timestep, beta_start=beta_start, beta_end=beta_end)
def test_schedules(self):
for schedule in ["linear", "scaled_linear", "squaredcos_cap_v2"]:
self.check_over_configs(time_step=self.default_valid_timestep, beta_schedule=schedule)
def test_prediction_type(self):
for prediction_type in ["epsilon", "v_prediction"]:
self.check_over_configs(time_step=self.default_valid_timestep, prediction_type=prediction_type)
def test_clip_sample(self):
for clip_sample in [True, False]:
self.check_over_configs(time_step=self.default_valid_timestep, clip_sample=clip_sample)
def test_thresholding(self):
self.check_over_configs(time_step=self.default_valid_timestep, thresholding=False)
for threshold in [0.5, 1.0, 2.0]:
for prediction_type in ["epsilon", "v_prediction"]:
self.check_over_configs(
time_step=self.default_valid_timestep,
thresholding=True,
prediction_type=prediction_type,
sample_max_value=threshold,
)
def test_time_indices(self):
# Get default timestep schedule.
kwargs = dict(self.forward_default_kwargs)
num_inference_steps = kwargs.pop("num_inference_steps", None)
scheduler_config = self.get_scheduler_config()
scheduler = self.scheduler_classes[0](**scheduler_config)
scheduler.set_timesteps(num_inference_steps)
timesteps = scheduler.timesteps
for t in timesteps:
self.check_over_forward(time_step=t)
def test_inference_steps(self):
# Hardcoded for now
for t, num_inference_steps in zip([99, 39, 39, 19], [10, 25, 26, 50]):
self.check_over_forward(time_step=t, num_inference_steps=num_inference_steps)
# Override test_add_noise_device because the hardcoded num_inference_steps of 100 doesn't work
# for LCMScheduler under default settings
def test_add_noise_device(self, num_inference_steps=10):
for scheduler_class in self.scheduler_classes:
scheduler_config = self.get_scheduler_config()
scheduler = scheduler_class(**scheduler_config)
scheduler.set_timesteps(num_inference_steps)
sample = self.dummy_sample.to(torch_device)
scaled_sample = scheduler.scale_model_input(sample, 0.0)
self.assertEqual(sample.shape, scaled_sample.shape)
noise = torch.randn_like(scaled_sample).to(torch_device)
t = scheduler.timesteps[5][None]
noised = scheduler.add_noise(scaled_sample, noise, t)
self.assertEqual(noised.shape, scaled_sample.shape)
# Override test_from_save_pretrained because it hardcodes a timestep of 1
def test_from_save_pretrained(self):
kwargs = dict(self.forward_default_kwargs)
num_inference_steps = kwargs.pop("num_inference_steps", None)
for scheduler_class in self.scheduler_classes:
timestep = self.default_valid_timestep
scheduler_config = self.get_scheduler_config()
scheduler = scheduler_class(**scheduler_config)
sample = self.dummy_sample
residual = 0.1 * sample
with tempfile.TemporaryDirectory() as tmpdirname:
scheduler.save_config(tmpdirname)
new_scheduler = scheduler_class.from_pretrained(tmpdirname)
scheduler.set_timesteps(num_inference_steps)
new_scheduler.set_timesteps(num_inference_steps)
kwargs["generator"] = torch.manual_seed(0)
output = scheduler.step(residual, timestep, sample, **kwargs).prev_sample
kwargs["generator"] = torch.manual_seed(0)
new_output = new_scheduler.step(residual, timestep, sample, **kwargs).prev_sample
assert torch.sum(torch.abs(output - new_output)) < 1e-5, "Scheduler outputs are not identical"
# Override test_step_shape because uses 0 and 1 as hardcoded timesteps
def test_step_shape(self):
kwargs = dict(self.forward_default_kwargs)
num_inference_steps = kwargs.pop("num_inference_steps", None)
for scheduler_class in self.scheduler_classes:
scheduler_config = self.get_scheduler_config()
scheduler = scheduler_class(**scheduler_config)
sample = self.dummy_sample
residual = 0.1 * sample
scheduler.set_timesteps(num_inference_steps)
timestep_0 = scheduler.timesteps[-2]
timestep_1 = scheduler.timesteps[-1]
output_0 = scheduler.step(residual, timestep_0, sample, **kwargs).prev_sample
output_1 = scheduler.step(residual, timestep_1, sample, **kwargs).prev_sample
self.assertEqual(output_0.shape, sample.shape)
self.assertEqual(output_0.shape, output_1.shape)
# Override test_set_scheduler_outputs_equivalence since it uses 0 as a hardcoded timestep
def test_scheduler_outputs_equivalence(self):
def set_nan_tensor_to_zero(t):
t[t != t] = 0
return t
def recursive_check(tuple_object, dict_object):
if isinstance(tuple_object, (List, Tuple)):
for tuple_iterable_value, dict_iterable_value in zip(tuple_object, dict_object.values()):
recursive_check(tuple_iterable_value, dict_iterable_value)
elif isinstance(tuple_object, Dict):
for tuple_iterable_value, dict_iterable_value in zip(tuple_object.values(), dict_object.values()):
recursive_check(tuple_iterable_value, dict_iterable_value)
elif tuple_object is None:
return
else:
self.assertTrue(
torch.allclose(
set_nan_tensor_to_zero(tuple_object), set_nan_tensor_to_zero(dict_object), atol=1e-5
),
msg=(
"Tuple and dict output are not equal. Difference:"
f" {torch.max(torch.abs(tuple_object - dict_object))}. Tuple has `nan`:"
f" {torch.isnan(tuple_object).any()} and `inf`: {torch.isinf(tuple_object)}. Dict has"
f" `nan`: {torch.isnan(dict_object).any()} and `inf`: {torch.isinf(dict_object)}."
),
)
kwargs = dict(self.forward_default_kwargs)
num_inference_steps = kwargs.pop("num_inference_steps", 50)
timestep = self.default_valid_timestep
for scheduler_class in self.scheduler_classes:
scheduler_config = self.get_scheduler_config()
scheduler = scheduler_class(**scheduler_config)
sample = self.dummy_sample
residual = 0.1 * sample
scheduler.set_timesteps(num_inference_steps)
kwargs["generator"] = torch.manual_seed(0)
outputs_dict = scheduler.step(residual, timestep, sample, **kwargs)
scheduler.set_timesteps(num_inference_steps)
kwargs["generator"] = torch.manual_seed(0)
outputs_tuple = scheduler.step(residual, timestep, sample, return_dict=False, **kwargs)
recursive_check(outputs_tuple, outputs_dict)
def full_loop(self, num_inference_steps=10, seed=0, **config):
scheduler_class = self.scheduler_classes[0]
scheduler_config = self.get_scheduler_config(**config)
scheduler = scheduler_class(**scheduler_config)
model = self.dummy_model()
sample = self.dummy_sample_deter
generator = torch.manual_seed(seed)
scheduler.set_timesteps(num_inference_steps)
for t in scheduler.timesteps:
residual = model(sample, t)
sample = scheduler.step(residual, t, sample, generator).prev_sample
return sample
def test_full_loop_onestep(self):
sample = self.full_loop(num_inference_steps=1)
result_sum = torch.sum(torch.abs(sample))
result_mean = torch.mean(torch.abs(sample))
# TODO: get expected sum and mean
assert abs(result_sum.item() - 18.7097) < 1e-3
assert abs(result_mean.item() - 0.0244) < 1e-3
def test_full_loop_multistep(self):
sample = self.full_loop(num_inference_steps=10)
result_sum = torch.sum(torch.abs(sample))
result_mean = torch.mean(torch.abs(sample))
# TODO: get expected sum and mean
assert abs(result_sum.item() - 197.7616) < 1e-3
assert abs(result_mean.item() - 0.2575) < 1e-3
def test_custom_timesteps(self):
scheduler_class = self.scheduler_classes[0]
scheduler_config = self.get_scheduler_config()
scheduler = scheduler_class(**scheduler_config)
timesteps = [100, 87, 50, 1, 0]
scheduler.set_timesteps(timesteps=timesteps)
scheduler_timesteps = scheduler.timesteps
for i, timestep in enumerate(scheduler_timesteps):
if i == len(timesteps) - 1:
expected_prev_t = -1
else:
expected_prev_t = timesteps[i + 1]
prev_t = scheduler.previous_timestep(timestep)
prev_t = prev_t.item()
self.assertEqual(prev_t, expected_prev_t)
def test_custom_timesteps_increasing_order(self):
scheduler_class = self.scheduler_classes[0]
scheduler_config = self.get_scheduler_config()
scheduler = scheduler_class(**scheduler_config)
timesteps = [100, 87, 50, 51, 0]
with self.assertRaises(ValueError, msg="`custom_timesteps` must be in descending order."):
scheduler.set_timesteps(timesteps=timesteps)
def test_custom_timesteps_passing_both_num_inference_steps_and_timesteps(self):
scheduler_class = self.scheduler_classes[0]
scheduler_config = self.get_scheduler_config()
scheduler = scheduler_class(**scheduler_config)
timesteps = [100, 87, 50, 1, 0]
num_inference_steps = len(timesteps)
with self.assertRaises(ValueError, msg="Can only pass one of `num_inference_steps` or `custom_timesteps`."):
scheduler.set_timesteps(num_inference_steps=num_inference_steps, timesteps=timesteps)
def test_custom_timesteps_too_large(self):
scheduler_class = self.scheduler_classes[0]
scheduler_config = self.get_scheduler_config()
scheduler = scheduler_class(**scheduler_config)
timesteps = [scheduler.config.num_train_timesteps]
with self.assertRaises(
ValueError,
msg="`timesteps` must start before `self.config.train_timesteps`: {scheduler.config.num_train_timesteps}}",
):
scheduler.set_timesteps(timesteps=timesteps)
|