File size: 14,651 Bytes
87d40d2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
import tempfile

import torch

from diffusers import (
    DEISMultistepScheduler,
    DPMSolverMultistepScheduler,
    DPMSolverSinglestepScheduler,
    UniPCMultistepScheduler,
)

from .test_schedulers import SchedulerCommonTest


class DPMSolverSinglestepSchedulerTest(SchedulerCommonTest):
    scheduler_classes = (DPMSolverSinglestepScheduler,)
    forward_default_kwargs = (("num_inference_steps", 25),)

    def get_scheduler_config(self, **kwargs):
        config = {
            "num_train_timesteps": 1000,
            "beta_start": 0.0001,
            "beta_end": 0.02,
            "beta_schedule": "linear",
            "solver_order": 2,
            "prediction_type": "epsilon",
            "thresholding": False,
            "sample_max_value": 1.0,
            "algorithm_type": "dpmsolver++",
            "solver_type": "midpoint",
            "lambda_min_clipped": -float("inf"),
            "variance_type": None,
            "final_sigmas_type": "sigma_min",
        }

        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 + 0.10]

        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.model_outputs = dummy_past_residuals[: scheduler.config.solver_order]

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

            output, new_output = sample, sample
            for t in range(time_step, time_step + scheduler.config.solver_order + 1):
                t = scheduler.timesteps[t]
                output = scheduler.step(residual, t, output, **kwargs).prev_sample
                new_output = new_scheduler.step(residual, t, new_output, **kwargs).prev_sample

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

    def test_from_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 + 0.10]

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

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

            with tempfile.TemporaryDirectory() as tmpdirname:
                scheduler.save_config(tmpdirname)
                new_scheduler = scheduler_class.from_pretrained(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.model_outputs = dummy_past_residuals[: new_scheduler.config.solver_order]

            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, scheduler=None, **config):
        if scheduler is None:
            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

        return sample

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

        num_inference_steps = 10
        scheduler.set_timesteps(num_inference_steps)
        timesteps = scheduler.timesteps
        # reset the timesteps using`timesteps`
        scheduler = scheduler_class(**scheduler_config)
        scheduler.set_timesteps(num_inference_steps=None, timesteps=timesteps)

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

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

        return sample

    def test_full_uneven_loop(self):
        scheduler = DPMSolverSinglestepScheduler(**self.get_scheduler_config())
        num_inference_steps = 50
        model = self.dummy_model()
        sample = self.dummy_sample_deter
        scheduler.set_timesteps(num_inference_steps)

        # make sure that the first t is uneven
        for i, t in enumerate(scheduler.timesteps[3:]):
            residual = model(sample, t)
            sample = scheduler.step(residual, t, sample).prev_sample

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

        assert abs(result_mean.item() - 0.2574) < 1e-3

    def test_timesteps(self):
        for timesteps in [25, 50, 100, 999, 1000]:
            self.check_over_configs(num_train_timesteps=timesteps)

    def test_switch(self):
        # make sure that iterating over schedulers with same config names gives same results
        # for defaults
        scheduler = DPMSolverSinglestepScheduler(**self.get_scheduler_config())
        sample = self.full_loop(scheduler=scheduler)
        result_mean = torch.mean(torch.abs(sample))

        assert abs(result_mean.item() - 0.2791) < 1e-3

        scheduler = DEISMultistepScheduler.from_config(scheduler.config)
        scheduler = DPMSolverMultistepScheduler.from_config(scheduler.config)
        scheduler = UniPCMultistepScheduler.from_config(scheduler.config)
        scheduler = DPMSolverSinglestepScheduler.from_config(scheduler.config)

        sample = self.full_loop(scheduler=scheduler)
        result_mean = torch.mean(torch.abs(sample))

        assert abs(result_mean.item() - 0.2791) < 1e-3

    def test_thresholding(self):
        self.check_over_configs(thresholding=False)
        for order in [1, 2, 3]:
            for solver_type in ["midpoint", "heun"]:
                for threshold in [0.5, 1.0, 2.0]:
                    for prediction_type in ["epsilon", "sample"]:
                        self.check_over_configs(
                            thresholding=True,
                            prediction_type=prediction_type,
                            sample_max_value=threshold,
                            algorithm_type="dpmsolver++",
                            solver_order=order,
                            solver_type=solver_type,
                        )

    def test_prediction_type(self):
        for prediction_type in ["epsilon", "v_prediction"]:
            self.check_over_configs(prediction_type=prediction_type)

    def test_solver_order_and_type(self):
        for algorithm_type in ["dpmsolver", "dpmsolver++"]:
            for solver_type in ["midpoint", "heun"]:
                for order in [1, 2, 3]:
                    for prediction_type in ["epsilon", "sample"]:
                        self.check_over_configs(
                            solver_order=order,
                            solver_type=solver_type,
                            prediction_type=prediction_type,
                            algorithm_type=algorithm_type,
                        )
                        sample = self.full_loop(
                            solver_order=order,
                            solver_type=solver_type,
                            prediction_type=prediction_type,
                            algorithm_type=algorithm_type,
                        )
                        assert not torch.isnan(sample).any(), "Samples have nan numbers"

    def test_lower_order_final(self):
        self.check_over_configs(lower_order_final=True)
        self.check_over_configs(lower_order_final=False)

    def test_lambda_min_clipped(self):
        self.check_over_configs(lambda_min_clipped=-float("inf"))
        self.check_over_configs(lambda_min_clipped=-5.1)

    def test_variance_type(self):
        self.check_over_configs(variance_type=None)
        self.check_over_configs(variance_type="learned_range")

    def test_inference_steps(self):
        for num_inference_steps in [1, 2, 3, 5, 10, 50, 100, 999, 1000]:
            self.check_over_forward(num_inference_steps=num_inference_steps, time_step=0)

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

        assert abs(result_mean.item() - 0.2791) < 1e-3

    def test_full_loop_with_karras(self):
        sample = self.full_loop(use_karras_sigmas=True)
        result_mean = torch.mean(torch.abs(sample))

        assert abs(result_mean.item() - 0.2248) < 1e-3

    def test_full_loop_with_v_prediction(self):
        sample = self.full_loop(prediction_type="v_prediction")
        result_mean = torch.mean(torch.abs(sample))

        assert abs(result_mean.item() - 0.1453) < 1e-3

    def test_full_loop_with_karras_and_v_prediction(self):
        sample = self.full_loop(prediction_type="v_prediction", use_karras_sigmas=True)
        result_mean = torch.mean(torch.abs(sample))

        assert abs(result_mean.item() - 0.0649) < 1e-3

    def test_fp16_support(self):
        scheduler_class = self.scheduler_classes[0]
        scheduler_config = self.get_scheduler_config(thresholding=True, dynamic_thresholding_ratio=0)
        scheduler = scheduler_class(**scheduler_config)

        num_inference_steps = 10
        model = self.dummy_model()
        sample = self.dummy_sample_deter.half()
        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

        assert sample.dtype == torch.float16

    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 + 0.10]
            scheduler.model_outputs = dummy_past_residuals[: scheduler.config.solver_order]

            time_step_0 = scheduler.timesteps[0]
            time_step_1 = scheduler.timesteps[1]

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

        num_inference_steps = 10
        t_start = 5

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

        # add noise
        noise = self.dummy_noise_deter
        timesteps = scheduler.timesteps[t_start * scheduler.order :]
        sample = scheduler.add_noise(sample, noise, timesteps[:1])

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

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

        assert abs(result_sum.item() - 269.2187) < 1e-2, f" expected result sum  269.2187, but get {result_sum}"
        assert abs(result_mean.item() - 0.3505) < 1e-3, f" expected result mean 0.3505, but get {result_mean}"

    def test_custom_timesteps(self):
        for prediction_type in ["epsilon", "sample", "v_prediction"]:
            for lower_order_final in [True, False]:
                for final_sigmas_type in ["sigma_min", "zero"]:
                    sample = self.full_loop(
                        prediction_type=prediction_type,
                        lower_order_final=lower_order_final,
                        final_sigmas_type=final_sigmas_type,
                    )
                    sample_custom_timesteps = self.full_loop_custom_timesteps(
                        prediction_type=prediction_type,
                        lower_order_final=lower_order_final,
                        final_sigmas_type=final_sigmas_type,
                    )
                    assert (
                        torch.sum(torch.abs(sample - sample_custom_timesteps)) < 1e-5
                    ), f"Scheduler outputs are not identical for prediction_type: {prediction_type}, lower_order_final: {lower_order_final} and final_sigmas_type: {final_sigmas_type}"