File size: 10,298 Bytes
fd43906
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import tempfile

import torch

from diffusers import PNDMScheduler

from .test_schedulers import SchedulerCommonTest


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

    def get_scheduler_config(self, **kwargs):
        config = {
            "num_train_timesteps": 1000,
            "beta_start": 0.0001,
            "beta_end": 0.02,
            "beta_schedule": "linear",
        }

        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[:]

            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.ets = dummy_past_residuals[:]

            output = scheduler.step_prk(residual, time_step, sample, **kwargs).prev_sample
            new_output = new_scheduler.step_prk(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_plms(residual, time_step, sample, **kwargs).prev_sample
            new_output = new_scheduler.step_plms(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_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[:]

            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.ets = dummy_past_residuals[:]

            output = scheduler.step_prk(residual, time_step, sample, **kwargs).prev_sample
            new_output = new_scheduler.step_prk(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_plms(residual, time_step, sample, **kwargs).prev_sample
            new_output = new_scheduler.step_plms(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.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

    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[:]

            output_0 = scheduler.step_prk(residual, 0, sample, **kwargs).prev_sample
            output_1 = scheduler.step_prk(residual, 1, sample, **kwargs).prev_sample

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

            output_0 = scheduler.step_plms(residual, 0, sample, **kwargs).prev_sample
            output_1 = scheduler.step_plms(residual, 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)

    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)
        assert torch.equal(
            scheduler.timesteps,
            torch.LongTensor(
                [901, 851, 851, 801, 801, 751, 751, 701, 701, 651, 651, 601, 601, 501, 401, 301, 201, 101, 1]
            ),
        )

    def test_betas(self):
        for beta_start, beta_end in zip([0.0001, 0.001], [0.002, 0.02]):
            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_prediction_type(self):
        for prediction_type in ["epsilon", "v_prediction"]:
            self.check_over_configs(prediction_type=prediction_type)

    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]):
            self.check_over_forward(num_inference_steps=num_inference_steps)

    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

    def test_inference_plms_no_past_residuals(self):
        with self.assertRaises(ValueError):
            scheduler_class = self.scheduler_classes[0]
            scheduler_config = self.get_scheduler_config()
            scheduler = scheduler_class(**scheduler_config)

            scheduler.step_plms(self.dummy_sample, 1, self.dummy_sample).prev_sample

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

        assert abs(result_sum.item() - 198.1318) < 1e-2
        assert abs(result_mean.item() - 0.2580) < 1e-3

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

        assert abs(result_sum.item() - 67.3986) < 1e-2
        assert abs(result_mean.item() - 0.0878) < 1e-3

    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))

        assert abs(result_sum.item() - 230.0399) < 1e-2
        assert abs(result_mean.item() - 0.2995) < 1e-3

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

        assert abs(result_sum.item() - 186.9482) < 1e-2
        assert abs(result_mean.item() - 0.2434) < 1e-3