test_scheduler.py 52.6 KB
Newer Older
Patrick von Platen's avatar
Patrick von Platen committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
# coding=utf-8
# Copyright 2022 HuggingFace Inc.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
hlky's avatar
hlky committed
15
import inspect
Patrick von Platen's avatar
Patrick von Platen committed
16
import tempfile
Patrick von Platen's avatar
Patrick von Platen committed
17
import unittest
18
from typing import Dict, List, Tuple
Patrick von Platen's avatar
Patrick von Platen committed
19

Patrick von Platen's avatar
Patrick von Platen committed
20
21
import numpy as np
import torch
Will Berman's avatar
Will Berman committed
22
import torch.nn.functional as F
Patrick von Platen's avatar
Patrick von Platen committed
23

24
25
26
from diffusers import (
    DDIMScheduler,
    DDPMScheduler,
hlky's avatar
hlky committed
27
28
    EulerAncestralDiscreteScheduler,
    EulerDiscreteScheduler,
29
30
31
32
    IPNDMScheduler,
    LMSDiscreteScheduler,
    PNDMScheduler,
    ScoreSdeVeScheduler,
Will Berman's avatar
Will Berman committed
33
    VQDiffusionScheduler,
34
)
35
from diffusers.utils import torch_device
Patrick von Platen's avatar
Patrick von Platen committed
36
37
38
39
40
41


torch.backends.cuda.matmul.allow_tf32 = False


class SchedulerCommonTest(unittest.TestCase):
Patrick von Platen's avatar
Patrick von Platen committed
42
43
    scheduler_classes = ()
    forward_default_kwargs = ()
Patrick von Platen's avatar
Patrick von Platen committed
44
45

    @property
46
    def dummy_sample(self):
Patrick von Platen's avatar
Patrick von Platen committed
47
48
49
50
51
        batch_size = 4
        num_channels = 3
        height = 8
        width = 8

52
        sample = torch.rand((batch_size, num_channels, height, width))
Patrick von Platen's avatar
Patrick von Platen committed
53

54
        return sample
Patrick von Platen's avatar
Patrick von Platen committed
55
56

    @property
57
    def dummy_sample_deter(self):
Patrick von Platen's avatar
Patrick von Platen committed
58
59
60
61
62
63
        batch_size = 4
        num_channels = 3
        height = 8
        width = 8

        num_elems = batch_size * num_channels * height * width
64
        sample = torch.arange(num_elems)
65
66
        sample = sample.reshape(num_channels, height, width, batch_size)
        sample = sample / num_elems
67
        sample = sample.permute(3, 0, 1, 2)
Patrick von Platen's avatar
Patrick von Platen committed
68

69
        return sample
Patrick von Platen's avatar
Patrick von Platen committed
70
71
72
73
74

    def get_scheduler_config(self):
        raise NotImplementedError

    def dummy_model(self):
75
76
        def model(sample, t, *args):
            return sample * t / (t + 1)
Patrick von Platen's avatar
Patrick von Platen committed
77
78
79

        return model

Patrick von Platen's avatar
Patrick von Platen committed
80
81
82
    def check_over_configs(self, time_step=0, **config):
        kwargs = dict(self.forward_default_kwargs)

83
84
        num_inference_steps = kwargs.pop("num_inference_steps", None)

hlky's avatar
hlky committed
85
        # TODO(Suraj) - delete the following two lines once DDPM, DDIM, and PNDM have timesteps casted to float by default
Patrick von Platen's avatar
Patrick von Platen committed
86
        for scheduler_class in self.scheduler_classes:
hlky's avatar
hlky committed
87
88
89
            if scheduler_class in (EulerAncestralDiscreteScheduler, EulerDiscreteScheduler, LMSDiscreteScheduler):
                time_step = float(time_step)

Patrick von Platen's avatar
Patrick von Platen committed
90
91
92
            scheduler_config = self.get_scheduler_config(**config)
            scheduler = scheduler_class(**scheduler_config)

Will Berman's avatar
Will Berman committed
93
94
95
96
97
98
99
100
101
            if scheduler_class == VQDiffusionScheduler:
                num_vec_classes = scheduler_config["num_vec_classes"]
                sample = self.dummy_sample(num_vec_classes)
                model = self.dummy_model(num_vec_classes)
                residual = model(sample, time_step)
            else:
                sample = self.dummy_sample
                residual = 0.1 * sample

Patrick von Platen's avatar
Patrick von Platen committed
102
103
104
105
            with tempfile.TemporaryDirectory() as tmpdirname:
                scheduler.save_config(tmpdirname)
                new_scheduler = scheduler_class.from_config(tmpdirname)

106
107
108
109
110
111
            if num_inference_steps is not None and hasattr(scheduler, "set_timesteps"):
                scheduler.set_timesteps(num_inference_steps)
                new_scheduler.set_timesteps(num_inference_steps)
            elif num_inference_steps is not None and not hasattr(scheduler, "set_timesteps"):
                kwargs["num_inference_steps"] = num_inference_steps

hlky's avatar
hlky committed
112
113
114
            # Set the seed before step() as some schedulers are stochastic like EulerAncestralDiscreteScheduler, EulerDiscreteScheduler
            if "generator" in set(inspect.signature(scheduler.step).parameters.keys()):
                kwargs["generator"] = torch.Generator().manual_seed(0)
115
            output = scheduler.step(residual, time_step, sample, **kwargs).prev_sample
hlky's avatar
hlky committed
116
117
118

            if "generator" in set(inspect.signature(scheduler.step).parameters.keys()):
                kwargs["generator"] = torch.Generator().manual_seed(0)
119
            new_output = new_scheduler.step(residual, time_step, sample, **kwargs).prev_sample
Patrick von Platen's avatar
Patrick von Platen committed
120

121
            assert torch.sum(torch.abs(output - new_output)) < 1e-5, "Scheduler outputs are not identical"
Patrick von Platen's avatar
Patrick von Platen committed
122
123
124
125
126

    def check_over_forward(self, time_step=0, **forward_kwargs):
        kwargs = dict(self.forward_default_kwargs)
        kwargs.update(forward_kwargs)

127
128
        num_inference_steps = kwargs.pop("num_inference_steps", None)

Patrick von Platen's avatar
Patrick von Platen committed
129
        for scheduler_class in self.scheduler_classes:
hlky's avatar
hlky committed
130
131
132
            if scheduler_class in (EulerAncestralDiscreteScheduler, EulerDiscreteScheduler, LMSDiscreteScheduler):
                time_step = float(time_step)

Patrick von Platen's avatar
Patrick von Platen committed
133
134
135
            scheduler_config = self.get_scheduler_config()
            scheduler = scheduler_class(**scheduler_config)

Will Berman's avatar
Will Berman committed
136
137
138
139
140
141
142
143
144
            if scheduler_class == VQDiffusionScheduler:
                num_vec_classes = scheduler_config["num_vec_classes"]
                sample = self.dummy_sample(num_vec_classes)
                model = self.dummy_model(num_vec_classes)
                residual = model(sample, time_step)
            else:
                sample = self.dummy_sample
                residual = 0.1 * sample

Patrick von Platen's avatar
Patrick von Platen committed
145
146
147
148
            with tempfile.TemporaryDirectory() as tmpdirname:
                scheduler.save_config(tmpdirname)
                new_scheduler = scheduler_class.from_config(tmpdirname)

149
150
151
152
153
154
            if num_inference_steps is not None and hasattr(scheduler, "set_timesteps"):
                scheduler.set_timesteps(num_inference_steps)
                new_scheduler.set_timesteps(num_inference_steps)
            elif num_inference_steps is not None and not hasattr(scheduler, "set_timesteps"):
                kwargs["num_inference_steps"] = num_inference_steps

hlky's avatar
hlky committed
155
156
            if "generator" in set(inspect.signature(scheduler.step).parameters.keys()):
                kwargs["generator"] = torch.Generator().manual_seed(0)
157
            output = scheduler.step(residual, time_step, sample, **kwargs).prev_sample
hlky's avatar
hlky committed
158
159
160

            if "generator" in set(inspect.signature(scheduler.step).parameters.keys()):
                kwargs["generator"] = torch.Generator().manual_seed(0)
161
            new_output = new_scheduler.step(residual, time_step, sample, **kwargs).prev_sample
Patrick von Platen's avatar
Patrick von Platen committed
162

163
            assert torch.sum(torch.abs(output - new_output)) < 1e-5, "Scheduler outputs are not identical"
Patrick von Platen's avatar
Patrick von Platen committed
164

Patrick von Platen's avatar
Patrick von Platen committed
165
    def test_from_pretrained_save_pretrained(self):
Patrick von Platen's avatar
Patrick von Platen committed
166
167
        kwargs = dict(self.forward_default_kwargs)

168
169
        num_inference_steps = kwargs.pop("num_inference_steps", None)

Patrick von Platen's avatar
Patrick von Platen committed
170
        for scheduler_class in self.scheduler_classes:
Will Berman's avatar
Will Berman committed
171
172
173
            timestep = 1
            if scheduler_class in (EulerAncestralDiscreteScheduler, EulerDiscreteScheduler, LMSDiscreteScheduler):
                timestep = float(timestep)
Patrick von Platen's avatar
Patrick von Platen committed
174
175
176
177

            scheduler_config = self.get_scheduler_config()
            scheduler = scheduler_class(**scheduler_config)

Will Berman's avatar
Will Berman committed
178
179
180
181
182
183
184
185
            if scheduler_class == VQDiffusionScheduler:
                num_vec_classes = scheduler_config["num_vec_classes"]
                sample = self.dummy_sample(num_vec_classes)
                model = self.dummy_model(num_vec_classes)
                residual = model(sample, timestep)
            else:
                sample = self.dummy_sample
                residual = 0.1 * sample
hlky's avatar
hlky committed
186

Patrick von Platen's avatar
Patrick von Platen committed
187
188
189
190
            with tempfile.TemporaryDirectory() as tmpdirname:
                scheduler.save_config(tmpdirname)
                new_scheduler = scheduler_class.from_config(tmpdirname)

191
192
193
194
195
196
            if num_inference_steps is not None and hasattr(scheduler, "set_timesteps"):
                scheduler.set_timesteps(num_inference_steps)
                new_scheduler.set_timesteps(num_inference_steps)
            elif num_inference_steps is not None and not hasattr(scheduler, "set_timesteps"):
                kwargs["num_inference_steps"] = num_inference_steps

hlky's avatar
hlky committed
197
198
199
200
201
202
203
            if "generator" in set(inspect.signature(scheduler.step).parameters.keys()):
                kwargs["generator"] = torch.Generator().manual_seed(0)
            output = scheduler.step(residual, timestep, sample, **kwargs).prev_sample

            if "generator" in set(inspect.signature(scheduler.step).parameters.keys()):
                kwargs["generator"] = torch.Generator().manual_seed(0)
            new_output = new_scheduler.step(residual, timestep, sample, **kwargs).prev_sample
Patrick von Platen's avatar
Patrick von Platen committed
204

205
            assert torch.sum(torch.abs(output - new_output)) < 1e-5, "Scheduler outputs are not identical"
Patrick von Platen's avatar
Patrick von Platen committed
206
207
208
209

    def test_step_shape(self):
        kwargs = dict(self.forward_default_kwargs)

210
211
        num_inference_steps = kwargs.pop("num_inference_steps", None)

hlky's avatar
hlky committed
212
213
214
        timestep_0 = 0
        timestep_1 = 1

Patrick von Platen's avatar
Patrick von Platen committed
215
        for scheduler_class in self.scheduler_classes:
hlky's avatar
hlky committed
216
217
218
219
            if scheduler_class in (EulerAncestralDiscreteScheduler, EulerDiscreteScheduler, LMSDiscreteScheduler):
                timestep_0 = float(timestep_0)
                timestep_1 = float(timestep_1)

Patrick von Platen's avatar
Patrick von Platen committed
220
221
222
            scheduler_config = self.get_scheduler_config()
            scheduler = scheduler_class(**scheduler_config)

Will Berman's avatar
Will Berman committed
223
224
225
226
227
228
229
230
            if scheduler_class == VQDiffusionScheduler:
                num_vec_classes = scheduler_config["num_vec_classes"]
                sample = self.dummy_sample(num_vec_classes)
                model = self.dummy_model(num_vec_classes)
                residual = model(sample, timestep_0)
            else:
                sample = self.dummy_sample
                residual = 0.1 * sample
Patrick von Platen's avatar
Patrick von Platen committed
231

232
233
234
235
236
            if num_inference_steps is not None and hasattr(scheduler, "set_timesteps"):
                scheduler.set_timesteps(num_inference_steps)
            elif num_inference_steps is not None and not hasattr(scheduler, "set_timesteps"):
                kwargs["num_inference_steps"] = num_inference_steps

hlky's avatar
hlky committed
237
238
            output_0 = scheduler.step(residual, timestep_0, sample, **kwargs).prev_sample
            output_1 = scheduler.step(residual, timestep_1, sample, **kwargs).prev_sample
Patrick von Platen's avatar
Patrick von Platen committed
239

240
            self.assertEqual(output_0.shape, sample.shape)
Patrick von Platen's avatar
Patrick von Platen committed
241
242
            self.assertEqual(output_0.shape, output_1.shape)

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
    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)
271
        num_inference_steps = kwargs.pop("num_inference_steps", 50)
272

273
274
275
276
        timestep = 0
        if len(self.scheduler_classes) > 0 and self.scheduler_classes[0] == IPNDMScheduler:
            timestep = 1

277
        for scheduler_class in self.scheduler_classes:
hlky's avatar
hlky committed
278
279
280
            if scheduler_class in (EulerAncestralDiscreteScheduler, EulerDiscreteScheduler, LMSDiscreteScheduler):
                timestep = float(timestep)

281
282
283
            scheduler_config = self.get_scheduler_config()
            scheduler = scheduler_class(**scheduler_config)

Will Berman's avatar
Will Berman committed
284
285
286
287
288
289
290
291
            if scheduler_class == VQDiffusionScheduler:
                num_vec_classes = scheduler_config["num_vec_classes"]
                sample = self.dummy_sample(num_vec_classes)
                model = self.dummy_model(num_vec_classes)
                residual = model(sample, timestep)
            else:
                sample = self.dummy_sample
                residual = 0.1 * sample
292
293
294
295
296
297

            if num_inference_steps is not None and hasattr(scheduler, "set_timesteps"):
                scheduler.set_timesteps(num_inference_steps)
            elif num_inference_steps is not None and not hasattr(scheduler, "set_timesteps"):
                kwargs["num_inference_steps"] = num_inference_steps

hlky's avatar
hlky committed
298
299
300
            # Set the seed before state as some schedulers are stochastic like EulerAncestralDiscreteScheduler, EulerDiscreteScheduler
            if "generator" in set(inspect.signature(scheduler.step).parameters.keys()):
                kwargs["generator"] = torch.Generator().manual_seed(0)
301
            outputs_dict = scheduler.step(residual, timestep, sample, **kwargs)
302
303
304
305
306
307

            if num_inference_steps is not None and hasattr(scheduler, "set_timesteps"):
                scheduler.set_timesteps(num_inference_steps)
            elif num_inference_steps is not None and not hasattr(scheduler, "set_timesteps"):
                kwargs["num_inference_steps"] = num_inference_steps

hlky's avatar
hlky committed
308
309
310
            # Set the seed before state as some schedulers are stochastic like EulerAncestralDiscreteScheduler, EulerDiscreteScheduler
            if "generator" in set(inspect.signature(scheduler.step).parameters.keys()):
                kwargs["generator"] = torch.Generator().manual_seed(0)
311
            outputs_tuple = scheduler.step(residual, timestep, sample, return_dict=False, **kwargs)
312
313
314

            recursive_check(outputs_tuple, outputs_dict)

315
316
317
318
    def test_scheduler_public_api(self):
        for scheduler_class in self.scheduler_classes:
            scheduler_config = self.get_scheduler_config()
            scheduler = scheduler_class(**scheduler_config)
Will Berman's avatar
Will Berman committed
319
320
321
322
323
324
325
326
327
328
329

            if scheduler_class != VQDiffusionScheduler:
                self.assertTrue(
                    hasattr(scheduler, "init_noise_sigma"),
                    f"{scheduler_class} does not implement a required attribute `init_noise_sigma`",
                )
                self.assertTrue(
                    hasattr(scheduler, "scale_model_input"),
                    f"{scheduler_class} does not implement a required class method `scale_model_input(sample,"
                    " timestep)`",
                )
330
331
332
333
334
            self.assertTrue(
                hasattr(scheduler, "step"),
                f"{scheduler_class} does not implement a required class method `step(...)`",
            )

Will Berman's avatar
Will Berman committed
335
336
337
338
            if scheduler_class != VQDiffusionScheduler:
                sample = self.dummy_sample
                scaled_sample = scheduler.scale_model_input(sample, 0.0)
                self.assertEqual(sample.shape, scaled_sample.shape)
339

340
341
342
343
344
345
346
    def test_add_noise_device(self):
        for scheduler_class in self.scheduler_classes:
            if scheduler_class == IPNDMScheduler:
                # Skip until #990 is addressed
                continue
            scheduler_config = self.get_scheduler_config()
            scheduler = scheduler_class(**scheduler_config)
347
            scheduler.set_timesteps(100)
348
349
350
351
352
353

            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)
354
            t = scheduler.timesteps[5][None]
355
356
357
            noised = scheduler.add_noise(scaled_sample, noise, t)
            self.assertEqual(noised.shape, scaled_sample.shape)

Patrick von Platen's avatar
Patrick von Platen committed
358
359

class DDPMSchedulerTest(SchedulerCommonTest):
Patrick von Platen's avatar
Patrick von Platen committed
360
    scheduler_classes = (DDPMScheduler,)
Patrick von Platen's avatar
Patrick von Platen committed
361
362
363

    def get_scheduler_config(self, **kwargs):
        config = {
Nathan Lambert's avatar
Nathan Lambert committed
364
            "num_train_timesteps": 1000,
Patrick von Platen's avatar
Patrick von Platen committed
365
366
367
368
            "beta_start": 0.0001,
            "beta_end": 0.02,
            "beta_schedule": "linear",
            "variance_type": "fixed_small",
Patrick von Platen's avatar
Patrick von Platen committed
369
            "clip_sample": True,
Patrick von Platen's avatar
Patrick von Platen committed
370
371
372
373
        }

        config.update(**kwargs)
        return config
Patrick von Platen's avatar
update  
Patrick von Platen committed
374

Patrick von Platen's avatar
Patrick von Platen committed
375
376
    def test_timesteps(self):
        for timesteps in [1, 5, 100, 1000]:
Nathan Lambert's avatar
Nathan Lambert committed
377
            self.check_over_configs(num_train_timesteps=timesteps)
Patrick von Platen's avatar
Patrick von Platen committed
378
379
380
381
382
383
384
385
386
387
388
389
390

    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(beta_start=beta_start, beta_end=beta_end)

    def test_schedules(self):
        for schedule in ["linear", "squaredcos_cap_v2"]:
            self.check_over_configs(beta_schedule=schedule)

    def test_variance_type(self):
        for variance in ["fixed_small", "fixed_large", "other"]:
            self.check_over_configs(variance_type=variance)

391
    def test_clip_sample(self):
Patrick von Platen's avatar
Patrick von Platen committed
392
393
        for clip_sample in [True, False]:
            self.check_over_configs(clip_sample=clip_sample)
Patrick von Platen's avatar
Patrick von Platen committed
394
395
396
397
398
399
400
401
402
403

    def test_time_indices(self):
        for t in [0, 500, 999]:
            self.check_over_forward(time_step=t)

    def test_variance(self):
        scheduler_class = self.scheduler_classes[0]
        scheduler_config = self.get_scheduler_config()
        scheduler = scheduler_class(**scheduler_config)

404
405
406
407
        assert torch.sum(torch.abs(scheduler._get_variance(0) - 0.0)) < 1e-5
        assert torch.sum(torch.abs(scheduler._get_variance(487) - 0.00979)) < 1e-5
        assert torch.sum(torch.abs(scheduler._get_variance(999) - 0.02)) < 1e-5

Patrick von Platen's avatar
Patrick von Platen committed
408
409
    def test_full_loop_no_noise(self):
        scheduler_class = self.scheduler_classes[0]
Patrick von Platen's avatar
Patrick von Platen committed
410
        scheduler_config = self.get_scheduler_config()
Patrick von Platen's avatar
Patrick von Platen committed
411
412
413
414
415
        scheduler = scheduler_class(**scheduler_config)

        num_trained_timesteps = len(scheduler)

        model = self.dummy_model()
416
        sample = self.dummy_sample_deter
417
        generator = torch.manual_seed(0)
Patrick von Platen's avatar
Patrick von Platen committed
418
419
420

        for t in reversed(range(num_trained_timesteps)):
            # 1. predict noise residual
421
            residual = model(sample, t)
Patrick von Platen's avatar
Patrick von Platen committed
422

423
            # 2. predict previous mean of sample x_t-1
424
            pred_prev_sample = scheduler.step(residual, t, sample, generator=generator).prev_sample
Patrick von Platen's avatar
Patrick von Platen committed
425

426
427
428
429
430
431
            # if t > 0:
            #     noise = self.dummy_sample_deter
            #     variance = scheduler.get_variance(t) ** (0.5) * noise
            #
            # sample = pred_prev_sample + variance
            sample = pred_prev_sample
Patrick von Platen's avatar
Patrick von Platen committed
432

433
434
        result_sum = torch.sum(torch.abs(sample))
        result_mean = torch.mean(torch.abs(sample))
Patrick von Platen's avatar
Patrick von Platen committed
435

436
        assert abs(result_sum.item() - 258.9070) < 1e-2
437
        assert abs(result_mean.item() - 0.3374) < 1e-3
Patrick von Platen's avatar
Patrick von Platen committed
438

Patrick von Platen's avatar
update  
Patrick von Platen committed
439

Patrick von Platen's avatar
Patrick von Platen committed
440
441
class DDIMSchedulerTest(SchedulerCommonTest):
    scheduler_classes = (DDIMScheduler,)
442
    forward_default_kwargs = (("eta", 0.0), ("num_inference_steps", 50))
Patrick von Platen's avatar
update  
Patrick von Platen committed
443

Patrick von Platen's avatar
Patrick von Platen committed
444
445
    def get_scheduler_config(self, **kwargs):
        config = {
Nathan Lambert's avatar
Nathan Lambert committed
446
            "num_train_timesteps": 1000,
Patrick von Platen's avatar
Patrick von Platen committed
447
448
449
            "beta_start": 0.0001,
            "beta_end": 0.02,
            "beta_schedule": "linear",
Patrick von Platen's avatar
Patrick von Platen committed
450
            "clip_sample": True,
Patrick von Platen's avatar
Patrick von Platen committed
451
        }
Patrick von Platen's avatar
Patrick von Platen committed
452

Patrick von Platen's avatar
Patrick von Platen committed
453
454
455
        config.update(**kwargs)
        return config

456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
    def full_loop(self, **config):
        scheduler_class = self.scheduler_classes[0]
        scheduler_config = self.get_scheduler_config(**config)
        scheduler = scheduler_class(**scheduler_config)

        num_inference_steps, eta = 10, 0.0

        model = self.dummy_model()
        sample = self.dummy_sample_deter

        scheduler.set_timesteps(num_inference_steps)

        for t in scheduler.timesteps:
            residual = model(sample, t)
            sample = scheduler.step(residual, t, sample, eta).prev_sample

        return sample

Patrick von Platen's avatar
Patrick von Platen committed
474
    def test_timesteps(self):
475
        for timesteps in [100, 500, 1000]:
Nathan Lambert's avatar
Nathan Lambert committed
476
            self.check_over_configs(num_train_timesteps=timesteps)
Patrick von Platen's avatar
Patrick von Platen committed
477

478
479
480
481
482
483
484
485
    def test_steps_offset(self):
        for steps_offset in [0, 1]:
            self.check_over_configs(steps_offset=steps_offset)

        scheduler_class = self.scheduler_classes[0]
        scheduler_config = self.get_scheduler_config(steps_offset=1)
        scheduler = scheduler_class(**scheduler_config)
        scheduler.set_timesteps(5)
486
        assert torch.equal(scheduler.timesteps, torch.LongTensor([801, 601, 401, 201, 1]))
487

Patrick von Platen's avatar
Patrick von Platen committed
488
489
490
491
492
493
494
495
    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(beta_start=beta_start, beta_end=beta_end)

    def test_schedules(self):
        for schedule in ["linear", "squaredcos_cap_v2"]:
            self.check_over_configs(beta_schedule=schedule)

496
    def test_clip_sample(self):
Patrick von Platen's avatar
Patrick von Platen committed
497
498
        for clip_sample in [True, False]:
            self.check_over_configs(clip_sample=clip_sample)
Patrick von Platen's avatar
Patrick von Platen committed
499
500
501
502
503
504
505

    def test_time_indices(self):
        for t in [1, 10, 49]:
            self.check_over_forward(time_step=t)

    def test_inference_steps(self):
        for t, num_inference_steps in zip([1, 10, 50], [10, 50, 500]):
506
            self.check_over_forward(time_step=t, num_inference_steps=num_inference_steps)
Patrick von Platen's avatar
Patrick von Platen committed
507
508
509
510
511
512
513

    def test_eta(self):
        for t, eta in zip([1, 10, 49], [0.0, 0.5, 1.0]):
            self.check_over_forward(time_step=t, eta=eta)

    def test_variance(self):
        scheduler_class = self.scheduler_classes[0]
Patrick von Platen's avatar
Patrick von Platen committed
514
        scheduler_config = self.get_scheduler_config()
Patrick von Platen's avatar
Patrick von Platen committed
515
516
        scheduler = scheduler_class(**scheduler_config)

517
518
519
520
521
522
        assert torch.sum(torch.abs(scheduler._get_variance(0, 0) - 0.0)) < 1e-5
        assert torch.sum(torch.abs(scheduler._get_variance(420, 400) - 0.14771)) < 1e-5
        assert torch.sum(torch.abs(scheduler._get_variance(980, 960) - 0.32460)) < 1e-5
        assert torch.sum(torch.abs(scheduler._get_variance(0, 0) - 0.0)) < 1e-5
        assert torch.sum(torch.abs(scheduler._get_variance(487, 486) - 0.00979)) < 1e-5
        assert torch.sum(torch.abs(scheduler._get_variance(999, 998) - 0.02)) < 1e-5
Patrick von Platen's avatar
Patrick von Platen committed
523
524

    def test_full_loop_no_noise(self):
525
        sample = self.full_loop()
Patrick von Platen's avatar
Patrick von Platen committed
526

527
528
        result_sum = torch.sum(torch.abs(sample))
        result_mean = torch.mean(torch.abs(sample))
Patrick von Platen's avatar
Patrick von Platen committed
529

530
531
        assert abs(result_sum.item() - 172.0067) < 1e-2
        assert abs(result_mean.item() - 0.223967) < 1e-3
Patrick von Platen's avatar
Patrick von Platen committed
532

533
534
535
536
537
    def test_full_loop_with_set_alpha_to_one(self):
        # We specify different beta, so that the first alpha is 0.99
        sample = self.full_loop(set_alpha_to_one=True, beta_start=0.01)
        result_sum = torch.sum(torch.abs(sample))
        result_mean = torch.mean(torch.abs(sample))
Patrick von Platen's avatar
Patrick von Platen committed
538

539
540
        assert abs(result_sum.item() - 149.8295) < 1e-2
        assert abs(result_mean.item() - 0.1951) < 1e-3
Patrick von Platen's avatar
Patrick von Platen committed
541

542
543
544
    def test_full_loop_with_no_set_alpha_to_one(self):
        # We specify different beta, so that the first alpha is 0.99
        sample = self.full_loop(set_alpha_to_one=False, beta_start=0.01)
545
546
        result_sum = torch.sum(torch.abs(sample))
        result_mean = torch.mean(torch.abs(sample))
Patrick von Platen's avatar
Patrick von Platen committed
547

548
549
        assert abs(result_sum.item() - 149.0784) < 1e-2
        assert abs(result_mean.item() - 0.1941) < 1e-3
Patrick von Platen's avatar
Patrick von Platen committed
550
551
552
553
554
555
556
557


class PNDMSchedulerTest(SchedulerCommonTest):
    scheduler_classes = (PNDMScheduler,)
    forward_default_kwargs = (("num_inference_steps", 50),)

    def get_scheduler_config(self, **kwargs):
        config = {
Nathan Lambert's avatar
Nathan Lambert committed
558
            "num_train_timesteps": 1000,
Patrick von Platen's avatar
Patrick von Platen committed
559
560
561
562
563
564
565
566
            "beta_start": 0.0001,
            "beta_end": 0.02,
            "beta_schedule": "linear",
        }

        config.update(**kwargs)
        return config

567
    def check_over_configs(self, time_step=0, **config):
Patrick von Platen's avatar
Patrick von Platen committed
568
        kwargs = dict(self.forward_default_kwargs)
Patrick von Platen's avatar
Patrick von Platen committed
569
        num_inference_steps = kwargs.pop("num_inference_steps", None)
570
571
        sample = self.dummy_sample
        residual = 0.1 * sample
Patrick von Platen's avatar
Patrick von Platen committed
572
573
574
575
576
        dummy_past_residuals = [residual + 0.2, residual + 0.15, residual + 0.1, residual + 0.05]

        for scheduler_class in self.scheduler_classes:
            scheduler_config = self.get_scheduler_config(**config)
            scheduler = scheduler_class(**scheduler_config)
Patrick von Platen's avatar
Patrick von Platen committed
577
            scheduler.set_timesteps(num_inference_steps)
Patrick von Platen's avatar
Patrick von Platen committed
578
579
580
581
582
583
            # copy over dummy past residuals
            scheduler.ets = dummy_past_residuals[:]

            with tempfile.TemporaryDirectory() as tmpdirname:
                scheduler.save_config(tmpdirname)
                new_scheduler = scheduler_class.from_config(tmpdirname)
Patrick von Platen's avatar
Patrick von Platen committed
584
                new_scheduler.set_timesteps(num_inference_steps)
Patrick von Platen's avatar
Patrick von Platen committed
585
586
587
                # copy over dummy past residuals
                new_scheduler.ets = dummy_past_residuals[:]

588
589
            output = scheduler.step_prk(residual, time_step, sample, **kwargs).prev_sample
            new_output = new_scheduler.step_prk(residual, time_step, sample, **kwargs).prev_sample
Patrick von Platen's avatar
Patrick von Platen committed
590

591
            assert torch.sum(torch.abs(output - new_output)) < 1e-5, "Scheduler outputs are not identical"
Patrick von Platen's avatar
Patrick von Platen committed
592

593
594
            output = scheduler.step_plms(residual, time_step, sample, **kwargs).prev_sample
            new_output = new_scheduler.step_plms(residual, time_step, sample, **kwargs).prev_sample
595

596
            assert torch.sum(torch.abs(output - new_output)) < 1e-5, "Scheduler outputs are not identical"
597
598
599
600
601

    def test_from_pretrained_save_pretrained(self):
        pass

    def check_over_forward(self, time_step=0, **forward_kwargs):
Patrick von Platen's avatar
Patrick von Platen committed
602
        kwargs = dict(self.forward_default_kwargs)
Patrick von Platen's avatar
Patrick von Platen committed
603
        num_inference_steps = kwargs.pop("num_inference_steps", None)
604
605
        sample = self.dummy_sample
        residual = 0.1 * sample
Patrick von Platen's avatar
Patrick von Platen committed
606
607
608
609
610
        dummy_past_residuals = [residual + 0.2, residual + 0.15, residual + 0.1, residual + 0.05]

        for scheduler_class in self.scheduler_classes:
            scheduler_config = self.get_scheduler_config()
            scheduler = scheduler_class(**scheduler_config)
Patrick von Platen's avatar
Patrick von Platen committed
611
            scheduler.set_timesteps(num_inference_steps)
612

Nathan Lambert's avatar
Nathan Lambert committed
613
            # copy over dummy past residuals (must be after setting timesteps)
Patrick von Platen's avatar
Patrick von Platen committed
614
615
616
617
618
619
            scheduler.ets = dummy_past_residuals[:]

            with tempfile.TemporaryDirectory() as tmpdirname:
                scheduler.save_config(tmpdirname)
                new_scheduler = scheduler_class.from_config(tmpdirname)
                # copy over dummy past residuals
Patrick von Platen's avatar
Patrick von Platen committed
620
                new_scheduler.set_timesteps(num_inference_steps)
Patrick von Platen's avatar
Patrick von Platen committed
621

Nathan Lambert's avatar
Nathan Lambert committed
622
623
624
                # copy over dummy past residual (must be after setting timesteps)
                new_scheduler.ets = dummy_past_residuals[:]

625
626
            output = scheduler.step_prk(residual, time_step, sample, **kwargs).prev_sample
            new_output = new_scheduler.step_prk(residual, time_step, sample, **kwargs).prev_sample
627

628
            assert torch.sum(torch.abs(output - new_output)) < 1e-5, "Scheduler outputs are not identical"
629

630
631
            output = scheduler.step_plms(residual, time_step, sample, **kwargs).prev_sample
            new_output = new_scheduler.step_plms(residual, time_step, sample, **kwargs).prev_sample
Patrick von Platen's avatar
Patrick von Platen committed
632

633
            assert torch.sum(torch.abs(output - new_output)) < 1e-5, "Scheduler outputs are not identical"
Patrick von Platen's avatar
Patrick von Platen committed
634

635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
    def full_loop(self, **config):
        scheduler_class = self.scheduler_classes[0]
        scheduler_config = self.get_scheduler_config(**config)
        scheduler = scheduler_class(**scheduler_config)

        num_inference_steps = 10
        model = self.dummy_model()
        sample = self.dummy_sample_deter
        scheduler.set_timesteps(num_inference_steps)

        for i, t in enumerate(scheduler.prk_timesteps):
            residual = model(sample, t)
            sample = scheduler.step_prk(residual, t, sample).prev_sample

        for i, t in enumerate(scheduler.plms_timesteps):
            residual = model(sample, t)
            sample = scheduler.step_plms(residual, t, sample).prev_sample

        return sample

655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
    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

            if num_inference_steps is not None and hasattr(scheduler, "set_timesteps"):
                scheduler.set_timesteps(num_inference_steps)
            elif num_inference_steps is not None and not hasattr(scheduler, "set_timesteps"):
                kwargs["num_inference_steps"] = num_inference_steps

Nathan Lambert's avatar
Nathan Lambert committed
672
673
674
675
            # copy over dummy past residuals (must be done after set_timesteps)
            dummy_past_residuals = [residual + 0.2, residual + 0.15, residual + 0.1, residual + 0.05]
            scheduler.ets = dummy_past_residuals[:]

676
677
            output_0 = scheduler.step_prk(residual, 0, sample, **kwargs).prev_sample
            output_1 = scheduler.step_prk(residual, 1, sample, **kwargs).prev_sample
678
679
680
681

            self.assertEqual(output_0.shape, sample.shape)
            self.assertEqual(output_0.shape, output_1.shape)

682
683
            output_0 = scheduler.step_plms(residual, 0, sample, **kwargs).prev_sample
            output_1 = scheduler.step_plms(residual, 1, sample, **kwargs).prev_sample
684
685
686
687

            self.assertEqual(output_0.shape, sample.shape)
            self.assertEqual(output_0.shape, output_1.shape)

Patrick von Platen's avatar
Patrick von Platen committed
688
689
    def test_timesteps(self):
        for timesteps in [100, 1000]:
Nathan Lambert's avatar
Nathan Lambert committed
690
            self.check_over_configs(num_train_timesteps=timesteps)
Patrick von Platen's avatar
Patrick von Platen committed
691

692
693
694
695
696
697
698
699
    def test_steps_offset(self):
        for steps_offset in [0, 1]:
            self.check_over_configs(steps_offset=steps_offset)

        scheduler_class = self.scheduler_classes[0]
        scheduler_config = self.get_scheduler_config(steps_offset=1)
        scheduler = scheduler_class(**scheduler_config)
        scheduler.set_timesteps(10)
700
        assert torch.equal(
701
            scheduler.timesteps,
702
703
704
705
            torch.LongTensor(
                [901, 851, 851, 801, 801, 751, 751, 701, 701, 651, 651, 601, 601, 501, 401, 301, 201, 101, 1]
            ),
        )
706

Patrick von Platen's avatar
Patrick von Platen committed
707
    def test_betas(self):
708
        for beta_start, beta_end in zip([0.0001, 0.001], [0.002, 0.02]):
Patrick von Platen's avatar
Patrick von Platen committed
709
710
711
712
713
714
715
716
717
718
719
720
            self.check_over_configs(beta_start=beta_start, beta_end=beta_end)

    def test_schedules(self):
        for schedule in ["linear", "squaredcos_cap_v2"]:
            self.check_over_configs(beta_schedule=schedule)

    def test_time_indices(self):
        for t in [1, 5, 10]:
            self.check_over_forward(time_step=t)

    def test_inference_steps(self):
        for t, num_inference_steps in zip([1, 5, 10], [10, 50, 100]):
721
            self.check_over_forward(num_inference_steps=num_inference_steps)
Patrick von Platen's avatar
Patrick von Platen committed
722

723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
    def test_pow_of_3_inference_steps(self):
        # earlier version of set_timesteps() caused an error indexing alpha's with inference steps as power of 3
        num_inference_steps = 27

        for scheduler_class in self.scheduler_classes:
            sample = self.dummy_sample
            residual = 0.1 * sample

            scheduler_config = self.get_scheduler_config()
            scheduler = scheduler_class(**scheduler_config)

            scheduler.set_timesteps(num_inference_steps)

            # before power of 3 fix, would error on first step, so we only need to do two
            for i, t in enumerate(scheduler.prk_timesteps[:2]):
                sample = scheduler.step_prk(residual, t, sample).prev_sample

740
    def test_inference_plms_no_past_residuals(self):
Patrick von Platen's avatar
Patrick von Platen committed
741
742
743
744
745
        with self.assertRaises(ValueError):
            scheduler_class = self.scheduler_classes[0]
            scheduler_config = self.get_scheduler_config()
            scheduler = scheduler_class(**scheduler_config)

746
            scheduler.step_plms(self.dummy_sample, 1, self.dummy_sample).prev_sample
Patrick von Platen's avatar
Patrick von Platen committed
747
748

    def test_full_loop_no_noise(self):
749
750
751
        sample = self.full_loop()
        result_sum = torch.sum(torch.abs(sample))
        result_mean = torch.mean(torch.abs(sample))
Patrick von Platen's avatar
Patrick von Platen committed
752

753
754
        assert abs(result_sum.item() - 198.1318) < 1e-2
        assert abs(result_mean.item() - 0.2580) < 1e-3
Patrick von Platen's avatar
Patrick von Platen committed
755

756
757
758
759
760
    def test_full_loop_with_set_alpha_to_one(self):
        # We specify different beta, so that the first alpha is 0.99
        sample = self.full_loop(set_alpha_to_one=True, beta_start=0.01)
        result_sum = torch.sum(torch.abs(sample))
        result_mean = torch.mean(torch.abs(sample))
Patrick von Platen's avatar
Patrick von Platen committed
761

762
763
        assert abs(result_sum.item() - 230.0399) < 1e-2
        assert abs(result_mean.item() - 0.2995) < 1e-3
Patrick von Platen's avatar
Patrick von Platen committed
764

765
766
767
    def test_full_loop_with_no_set_alpha_to_one(self):
        # We specify different beta, so that the first alpha is 0.99
        sample = self.full_loop(set_alpha_to_one=False, beta_start=0.01)
768
769
        result_sum = torch.sum(torch.abs(sample))
        result_mean = torch.mean(torch.abs(sample))
Patrick von Platen's avatar
Patrick von Platen committed
770

771
772
        assert abs(result_sum.item() - 186.9482) < 1e-2
        assert abs(result_mean.item() - 0.2434) < 1e-3
Nathan Lambert's avatar
Nathan Lambert committed
773
774


775
776
class ScoreSdeVeSchedulerTest(unittest.TestCase):
    # TODO adapt with class SchedulerCommonTest (scheduler needs Numpy Integration)
Nathan Lambert's avatar
Nathan Lambert committed
777
    scheduler_classes = (ScoreSdeVeScheduler,)
778
    forward_default_kwargs = ()
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810

    @property
    def dummy_sample(self):
        batch_size = 4
        num_channels = 3
        height = 8
        width = 8

        sample = torch.rand((batch_size, num_channels, height, width))

        return sample

    @property
    def dummy_sample_deter(self):
        batch_size = 4
        num_channels = 3
        height = 8
        width = 8

        num_elems = batch_size * num_channels * height * width
        sample = torch.arange(num_elems)
        sample = sample.reshape(num_channels, height, width, batch_size)
        sample = sample / num_elems
        sample = sample.permute(3, 0, 1, 2)

        return sample

    def dummy_model(self):
        def model(sample, t, *args):
            return sample * t / (t + 1)

        return model
Nathan Lambert's avatar
Nathan Lambert committed
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837

    def get_scheduler_config(self, **kwargs):
        config = {
            "num_train_timesteps": 2000,
            "snr": 0.15,
            "sigma_min": 0.01,
            "sigma_max": 1348,
            "sampling_eps": 1e-5,
        }

        config.update(**kwargs)
        return config

    def check_over_configs(self, time_step=0, **config):
        kwargs = dict(self.forward_default_kwargs)

        for scheduler_class in self.scheduler_classes:
            sample = self.dummy_sample
            residual = 0.1 * sample

            scheduler_config = self.get_scheduler_config(**config)
            scheduler = scheduler_class(**scheduler_config)

            with tempfile.TemporaryDirectory() as tmpdirname:
                scheduler.save_config(tmpdirname)
                new_scheduler = scheduler_class.from_config(tmpdirname)

838
839
840
841
842
843
            output = scheduler.step_pred(
                residual, time_step, sample, generator=torch.manual_seed(0), **kwargs
            ).prev_sample
            new_output = new_scheduler.step_pred(
                residual, time_step, sample, generator=torch.manual_seed(0), **kwargs
            ).prev_sample
Nathan Lambert's avatar
Nathan Lambert committed
844

845
            assert torch.sum(torch.abs(output - new_output)) < 1e-5, "Scheduler outputs are not identical"
Nathan Lambert's avatar
Nathan Lambert committed
846

847
848
849
850
            output = scheduler.step_correct(residual, sample, generator=torch.manual_seed(0), **kwargs).prev_sample
            new_output = new_scheduler.step_correct(
                residual, sample, generator=torch.manual_seed(0), **kwargs
            ).prev_sample
Nathan Lambert's avatar
Nathan Lambert committed
851

852
            assert torch.sum(torch.abs(output - new_output)) < 1e-5, "Scheduler correction are not identical"
Nathan Lambert's avatar
Nathan Lambert committed
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868

    def check_over_forward(self, time_step=0, **forward_kwargs):
        kwargs = dict(self.forward_default_kwargs)
        kwargs.update(forward_kwargs)

        for scheduler_class in self.scheduler_classes:
            sample = self.dummy_sample
            residual = 0.1 * sample

            scheduler_config = self.get_scheduler_config()
            scheduler = scheduler_class(**scheduler_config)

            with tempfile.TemporaryDirectory() as tmpdirname:
                scheduler.save_config(tmpdirname)
                new_scheduler = scheduler_class.from_config(tmpdirname)

869
870
871
872
873
874
            output = scheduler.step_pred(
                residual, time_step, sample, generator=torch.manual_seed(0), **kwargs
            ).prev_sample
            new_output = new_scheduler.step_pred(
                residual, time_step, sample, generator=torch.manual_seed(0), **kwargs
            ).prev_sample
Nathan Lambert's avatar
Nathan Lambert committed
875

876
            assert torch.sum(torch.abs(output - new_output)) < 1e-5, "Scheduler outputs are not identical"
Nathan Lambert's avatar
Nathan Lambert committed
877

878
879
880
881
            output = scheduler.step_correct(residual, sample, generator=torch.manual_seed(0), **kwargs).prev_sample
            new_output = new_scheduler.step_correct(
                residual, sample, generator=torch.manual_seed(0), **kwargs
            ).prev_sample
Nathan Lambert's avatar
Nathan Lambert committed
882

883
            assert torch.sum(torch.abs(output - new_output)) < 1e-5, "Scheduler correction are not identical"
Nathan Lambert's avatar
Nathan Lambert committed
884
885
886
887
888
889
890
891
892
893

    def test_timesteps(self):
        for timesteps in [10, 100, 1000]:
            self.check_over_configs(num_train_timesteps=timesteps)

    def test_sigmas(self):
        for sigma_min, sigma_max in zip([0.0001, 0.001, 0.01], [1, 100, 1000]):
            self.check_over_configs(sigma_min=sigma_min, sigma_max=sigma_max)

    def test_time_indices(self):
894
        for t in [0.1, 0.5, 0.75]:
Nathan Lambert's avatar
Nathan Lambert committed
895
896
897
            self.check_over_forward(time_step=t)

    def test_full_loop_no_noise(self):
898
899
        kwargs = dict(self.forward_default_kwargs)

Nathan Lambert's avatar
Nathan Lambert committed
900
901
902
903
904
905
906
907
908
909
        scheduler_class = self.scheduler_classes[0]
        scheduler_config = self.get_scheduler_config()
        scheduler = scheduler_class(**scheduler_config)

        num_inference_steps = 3

        model = self.dummy_model()
        sample = self.dummy_sample_deter

        scheduler.set_sigmas(num_inference_steps)
910
        scheduler.set_timesteps(num_inference_steps)
911
        generator = torch.manual_seed(0)
Nathan Lambert's avatar
Nathan Lambert committed
912
913
914
915

        for i, t in enumerate(scheduler.timesteps):
            sigma_t = scheduler.sigmas[i]

916
            for _ in range(scheduler.config.correct_steps):
Nathan Lambert's avatar
Nathan Lambert committed
917
                with torch.no_grad():
918
                    model_output = model(sample, sigma_t)
919
                sample = scheduler.step_correct(model_output, sample, generator=generator, **kwargs).prev_sample
Nathan Lambert's avatar
Nathan Lambert committed
920
921

            with torch.no_grad():
922
                model_output = model(sample, sigma_t)
Patrick von Platen's avatar
Patrick von Platen committed
923

924
            output = scheduler.step_pred(model_output, t, sample, generator=generator, **kwargs)
925
            sample, _ = output.prev_sample, output.prev_sample_mean
Patrick von Platen's avatar
Patrick von Platen committed
926

927
928
        result_sum = torch.sum(torch.abs(sample))
        result_mean = torch.mean(torch.abs(sample))
Patrick von Platen's avatar
Patrick von Platen committed
929

930
931
        assert np.isclose(result_sum.item(), 14372758528.0)
        assert np.isclose(result_mean.item(), 18714530.0)
Patrick von Platen's avatar
Patrick von Platen committed
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949

    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

            if num_inference_steps is not None and hasattr(scheduler, "set_timesteps"):
                scheduler.set_timesteps(num_inference_steps)
            elif num_inference_steps is not None and not hasattr(scheduler, "set_timesteps"):
                kwargs["num_inference_steps"] = num_inference_steps

950
951
            output_0 = scheduler.step_pred(residual, 0, sample, generator=torch.manual_seed(0), **kwargs).prev_sample
            output_1 = scheduler.step_pred(residual, 1, sample, generator=torch.manual_seed(0), **kwargs).prev_sample
Patrick von Platen's avatar
Patrick von Platen committed
952
953
954

            self.assertEqual(output_0.shape, sample.shape)
            self.assertEqual(output_0.shape, output_1.shape)
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977


class LMSDiscreteSchedulerTest(SchedulerCommonTest):
    scheduler_classes = (LMSDiscreteScheduler,)
    num_inference_steps = 10

    def get_scheduler_config(self, **kwargs):
        config = {
            "num_train_timesteps": 1100,
            "beta_start": 0.0001,
            "beta_end": 0.02,
            "beta_schedule": "linear",
            "trained_betas": None,
        }

        config.update(**kwargs)
        return config

    def test_timesteps(self):
        for timesteps in [10, 50, 100, 1000]:
            self.check_over_configs(num_train_timesteps=timesteps)

    def test_betas(self):
978
        for beta_start, beta_end in zip([0.00001, 0.0001, 0.001], [0.0002, 0.002, 0.02]):
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
            self.check_over_configs(beta_start=beta_start, beta_end=beta_end)

    def test_schedules(self):
        for schedule in ["linear", "scaled_linear"]:
            self.check_over_configs(beta_schedule=schedule)

    def test_time_indices(self):
        for t in [0, 500, 800]:
            self.check_over_forward(time_step=t)

    def test_full_loop_no_noise(self):
        scheduler_class = self.scheduler_classes[0]
        scheduler_config = self.get_scheduler_config()
        scheduler = scheduler_class(**scheduler_config)

        scheduler.set_timesteps(self.num_inference_steps)

        model = self.dummy_model()
997
        sample = self.dummy_sample_deter * scheduler.init_noise_sigma
998
999

        for i, t in enumerate(scheduler.timesteps):
1000
            sample = scheduler.scale_model_input(sample, t)
1001
1002
1003

            model_output = model(sample, t)

1004
            output = scheduler.step(model_output, t, sample)
1005
1006
1007
1008
1009
            sample = output.prev_sample

        result_sum = torch.sum(torch.abs(sample))
        result_mean = torch.mean(torch.abs(sample))

1010
        assert abs(result_sum.item() - 1006.388) < 1e-2
1011
        assert abs(result_mean.item() - 1.31) < 1e-3
1012
1013


hlky's avatar
hlky committed
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
class EulerDiscreteSchedulerTest(SchedulerCommonTest):
    scheduler_classes = (EulerDiscreteScheduler,)
    num_inference_steps = 10

    def get_scheduler_config(self, **kwargs):
        config = {
            "num_train_timesteps": 1100,
            "beta_start": 0.0001,
            "beta_end": 0.02,
            "beta_schedule": "linear",
            "trained_betas": None,
        }

        config.update(**kwargs)
        return config

    def test_timesteps(self):
        for timesteps in [10, 50, 100, 1000]:
            self.check_over_configs(num_train_timesteps=timesteps)

    def test_betas(self):
        for beta_start, beta_end in zip([0.00001, 0.0001, 0.001], [0.0002, 0.002, 0.02]):
            self.check_over_configs(beta_start=beta_start, beta_end=beta_end)

    def test_schedules(self):
        for schedule in ["linear", "scaled_linear"]:
            self.check_over_configs(beta_schedule=schedule)

    def test_full_loop_no_noise(self):
        scheduler_class = self.scheduler_classes[0]
        scheduler_config = self.get_scheduler_config()
        scheduler = scheduler_class(**scheduler_config)

        scheduler.set_timesteps(self.num_inference_steps)

        generator = torch.Generator().manual_seed(0)

        model = self.dummy_model()
        sample = self.dummy_sample_deter * scheduler.init_noise_sigma

        for i, t in enumerate(scheduler.timesteps):
            sample = scheduler.scale_model_input(sample, t)

            model_output = model(sample, t)

            output = scheduler.step(model_output, t, sample, generator=generator)
            sample = output.prev_sample

        result_sum = torch.sum(torch.abs(sample))
        result_mean = torch.mean(torch.abs(sample))
        print(result_sum, result_mean)

        assert abs(result_sum.item() - 10.0807) < 1e-2
        assert abs(result_mean.item() - 0.0131) < 1e-3


class EulerAncestralDiscreteSchedulerTest(SchedulerCommonTest):
    scheduler_classes = (EulerAncestralDiscreteScheduler,)
    num_inference_steps = 10

    def get_scheduler_config(self, **kwargs):
        config = {
            "num_train_timesteps": 1100,
            "beta_start": 0.0001,
            "beta_end": 0.02,
            "beta_schedule": "linear",
            "trained_betas": None,
        }

        config.update(**kwargs)
        return config

    def test_timesteps(self):
        for timesteps in [10, 50, 100, 1000]:
            self.check_over_configs(num_train_timesteps=timesteps)

    def test_betas(self):
        for beta_start, beta_end in zip([0.00001, 0.0001, 0.001], [0.0002, 0.002, 0.02]):
            self.check_over_configs(beta_start=beta_start, beta_end=beta_end)

    def test_schedules(self):
        for schedule in ["linear", "scaled_linear"]:
            self.check_over_configs(beta_schedule=schedule)

    def test_full_loop_no_noise(self):
        scheduler_class = self.scheduler_classes[0]
        scheduler_config = self.get_scheduler_config()
        scheduler = scheduler_class(**scheduler_config)

        scheduler.set_timesteps(self.num_inference_steps)

        generator = torch.Generator().manual_seed(0)

        model = self.dummy_model()
        sample = self.dummy_sample_deter * scheduler.init_noise_sigma

        for i, t in enumerate(scheduler.timesteps):
            sample = scheduler.scale_model_input(sample, t)

            model_output = model(sample, t)

            output = scheduler.step(model_output, t, sample, generator=generator)
            sample = output.prev_sample

        result_sum = torch.sum(torch.abs(sample))
        result_mean = torch.mean(torch.abs(sample))
        print(result_sum, result_mean)
        assert abs(result_sum.item() - 152.3192) < 1e-2
        assert abs(result_mean.item() - 0.1983) < 1e-3


1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
class IPNDMSchedulerTest(SchedulerCommonTest):
    scheduler_classes = (IPNDMScheduler,)
    forward_default_kwargs = (("num_inference_steps", 50),)

    def get_scheduler_config(self, **kwargs):
        config = {"num_train_timesteps": 1000}
        config.update(**kwargs)
        return config

    def check_over_configs(self, time_step=0, **config):
        kwargs = dict(self.forward_default_kwargs)
        num_inference_steps = kwargs.pop("num_inference_steps", None)
        sample = self.dummy_sample
        residual = 0.1 * sample
        dummy_past_residuals = [residual + 0.2, residual + 0.15, residual + 0.1, residual + 0.05]

        for scheduler_class in self.scheduler_classes:
            scheduler_config = self.get_scheduler_config(**config)
            scheduler = scheduler_class(**scheduler_config)
            scheduler.set_timesteps(num_inference_steps)
            # copy over dummy past residuals
            scheduler.ets = dummy_past_residuals[:]

            if time_step is None:
                time_step = scheduler.timesteps[len(scheduler.timesteps) // 2]

            with tempfile.TemporaryDirectory() as tmpdirname:
                scheduler.save_config(tmpdirname)
                new_scheduler = scheduler_class.from_config(tmpdirname)
                new_scheduler.set_timesteps(num_inference_steps)
                # copy over dummy past residuals
                new_scheduler.ets = dummy_past_residuals[:]

            output = scheduler.step(residual, time_step, sample, **kwargs).prev_sample
            new_output = new_scheduler.step(residual, time_step, sample, **kwargs).prev_sample

            assert torch.sum(torch.abs(output - new_output)) < 1e-5, "Scheduler outputs are not identical"

            output = scheduler.step(residual, time_step, sample, **kwargs).prev_sample
            new_output = new_scheduler.step(residual, time_step, sample, **kwargs).prev_sample

            assert torch.sum(torch.abs(output - new_output)) < 1e-5, "Scheduler outputs are not identical"

    def test_from_pretrained_save_pretrained(self):
        pass

    def check_over_forward(self, time_step=0, **forward_kwargs):
        kwargs = dict(self.forward_default_kwargs)
        num_inference_steps = kwargs.pop("num_inference_steps", None)
        sample = self.dummy_sample
        residual = 0.1 * sample
        dummy_past_residuals = [residual + 0.2, residual + 0.15, residual + 0.1, residual + 0.05]

        for scheduler_class in self.scheduler_classes:
            scheduler_config = self.get_scheduler_config()
            scheduler = scheduler_class(**scheduler_config)
            scheduler.set_timesteps(num_inference_steps)

            # copy over dummy past residuals (must be after setting timesteps)
            scheduler.ets = dummy_past_residuals[:]

            if time_step is None:
                time_step = scheduler.timesteps[len(scheduler.timesteps) // 2]

            with tempfile.TemporaryDirectory() as tmpdirname:
                scheduler.save_config(tmpdirname)
                new_scheduler = scheduler_class.from_config(tmpdirname)
                # copy over dummy past residuals
                new_scheduler.set_timesteps(num_inference_steps)

                # copy over dummy past residual (must be after setting timesteps)
                new_scheduler.ets = dummy_past_residuals[:]

            output = scheduler.step(residual, time_step, sample, **kwargs).prev_sample
            new_output = new_scheduler.step(residual, time_step, sample, **kwargs).prev_sample

            assert torch.sum(torch.abs(output - new_output)) < 1e-5, "Scheduler outputs are not identical"

            output = scheduler.step(residual, time_step, sample, **kwargs).prev_sample
            new_output = new_scheduler.step(residual, time_step, sample, **kwargs).prev_sample

            assert torch.sum(torch.abs(output - new_output)) < 1e-5, "Scheduler outputs are not identical"

    def full_loop(self, **config):
        scheduler_class = self.scheduler_classes[0]
        scheduler_config = self.get_scheduler_config(**config)
        scheduler = scheduler_class(**scheduler_config)

        num_inference_steps = 10
        model = self.dummy_model()
        sample = self.dummy_sample_deter
        scheduler.set_timesteps(num_inference_steps)

        for i, t in enumerate(scheduler.timesteps):
            residual = model(sample, t)
            sample = scheduler.step(residual, t, sample).prev_sample

        for i, t in enumerate(scheduler.timesteps):
            residual = model(sample, t)
            sample = scheduler.step(residual, t, sample).prev_sample

        return sample

    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

            if num_inference_steps is not None and hasattr(scheduler, "set_timesteps"):
                scheduler.set_timesteps(num_inference_steps)
            elif num_inference_steps is not None and not hasattr(scheduler, "set_timesteps"):
                kwargs["num_inference_steps"] = num_inference_steps

            # copy over dummy past residuals (must be done after set_timesteps)
            dummy_past_residuals = [residual + 0.2, residual + 0.15, residual + 0.1, residual + 0.05]
            scheduler.ets = dummy_past_residuals[:]

            time_step_0 = scheduler.timesteps[5]
            time_step_1 = scheduler.timesteps[6]

            output_0 = scheduler.step(residual, time_step_0, sample, **kwargs).prev_sample
            output_1 = scheduler.step(residual, time_step_1, sample, **kwargs).prev_sample

            self.assertEqual(output_0.shape, sample.shape)
            self.assertEqual(output_0.shape, output_1.shape)

            output_0 = scheduler.step(residual, time_step_0, sample, **kwargs).prev_sample
            output_1 = scheduler.step(residual, time_step_1, sample, **kwargs).prev_sample

            self.assertEqual(output_0.shape, sample.shape)
            self.assertEqual(output_0.shape, output_1.shape)

    def test_timesteps(self):
        for timesteps in [100, 1000]:
            self.check_over_configs(num_train_timesteps=timesteps, time_step=None)

    def test_inference_steps(self):
        for t, num_inference_steps in zip([1, 5, 10], [10, 50, 100]):
            self.check_over_forward(num_inference_steps=num_inference_steps, time_step=None)

    def test_full_loop_no_noise(self):
        sample = self.full_loop()
        result_mean = torch.mean(torch.abs(sample))

        assert abs(result_mean.item() - 2540529) < 10
Will Berman's avatar
Will Berman committed
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326


class VQDiffusionSchedulerTest(SchedulerCommonTest):
    scheduler_classes = (VQDiffusionScheduler,)

    def get_scheduler_config(self, **kwargs):
        config = {
            "num_vec_classes": 4097,
            "num_train_timesteps": 100,
        }

        config.update(**kwargs)
        return config

    def dummy_sample(self, num_vec_classes):
        batch_size = 4
        height = 8
        width = 8

        sample = torch.randint(0, num_vec_classes, (batch_size, height * width))

        return sample

    @property
    def dummy_sample_deter(self):
        assert False

    def dummy_model(self, num_vec_classes):
        def model(sample, t, *args):
            batch_size, num_latent_pixels = sample.shape
            logits = torch.rand((batch_size, num_vec_classes - 1, num_latent_pixels))
            return_value = F.log_softmax(logits.double(), dim=1).float()
            return return_value

        return model

    def test_timesteps(self):
        for timesteps in [2, 5, 100, 1000]:
            self.check_over_configs(num_train_timesteps=timesteps)

    def test_num_vec_classes(self):
        for num_vec_classes in [5, 100, 1000, 4000]:
            self.check_over_configs(num_vec_classes=num_vec_classes)

    def test_time_indices(self):
        for t in [0, 50, 99]:
            self.check_over_forward(time_step=t)

    def test_add_noise_device(self):
        pass