File size: 9,640 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
# Copyright 2024 ParaDiGMS authors and The HuggingFace Team. All rights reserved.
#
# 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.

import torch

from diffusers import DDPMParallelScheduler

from .test_schedulers import SchedulerCommonTest


class DDPMParallelSchedulerTest(SchedulerCommonTest):
    scheduler_classes = (DDPMParallelScheduler,)

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

        config.update(**kwargs)
        return config

    def test_timesteps(self):
        for timesteps in [1, 5, 100, 1000]:
            self.check_over_configs(num_train_timesteps=timesteps)

    def test_betas(self):
        for beta_start, beta_end in zip([0.0001, 0.001, 0.01, 0.1], [0.002, 0.02, 0.2, 2]):
            self.check_over_configs(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)

    def test_clip_sample(self):
        for clip_sample in [True, False]:
            self.check_over_configs(clip_sample=clip_sample)

    def test_thresholding(self):
        self.check_over_configs(thresholding=False)
        for threshold in [0.5, 1.0, 2.0]:
            for prediction_type in ["epsilon", "sample", "v_prediction"]:
                self.check_over_configs(
                    thresholding=True,
                    prediction_type=prediction_type,
                    sample_max_value=threshold,
                )

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

    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)

        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

    def test_rescale_betas_zero_snr(self):
        for rescale_betas_zero_snr in [True, False]:
            self.check_over_configs(rescale_betas_zero_snr=rescale_betas_zero_snr)

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

        num_trained_timesteps = len(scheduler)

        model = self.dummy_model()
        sample1 = self.dummy_sample_deter
        sample2 = self.dummy_sample_deter + 0.1
        sample3 = self.dummy_sample_deter - 0.1

        per_sample_batch = sample1.shape[0]
        samples = torch.stack([sample1, sample2, sample3], dim=0)
        timesteps = torch.arange(num_trained_timesteps)[0:3, None].repeat(1, per_sample_batch)

        residual = model(samples.flatten(0, 1), timesteps.flatten(0, 1))
        pred_prev_sample = scheduler.batch_step_no_noise(residual, timesteps.flatten(0, 1), samples.flatten(0, 1))

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

        assert abs(result_sum.item() - 1153.1833) < 1e-2
        assert abs(result_mean.item() - 0.5005) < 1e-3

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

        num_trained_timesteps = len(scheduler)

        model = self.dummy_model()
        sample = self.dummy_sample_deter
        generator = torch.manual_seed(0)

        for t in reversed(range(num_trained_timesteps)):
            # 1. predict noise residual
            residual = model(sample, t)

            # 2. predict previous mean of sample x_t-1
            pred_prev_sample = scheduler.step(residual, t, sample, generator=generator).prev_sample

            sample = pred_prev_sample

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

        assert abs(result_sum.item() - 258.9606) < 1e-2
        assert abs(result_mean.item() - 0.3372) < 1e-3

    def test_full_loop_with_v_prediction(self):
        scheduler_class = self.scheduler_classes[0]
        scheduler_config = self.get_scheduler_config(prediction_type="v_prediction")
        scheduler = scheduler_class(**scheduler_config)

        num_trained_timesteps = len(scheduler)

        model = self.dummy_model()
        sample = self.dummy_sample_deter
        generator = torch.manual_seed(0)

        for t in reversed(range(num_trained_timesteps)):
            # 1. predict noise residual
            residual = model(sample, t)

            # 2. predict previous mean of sample x_t-1
            pred_prev_sample = scheduler.step(residual, t, sample, generator=generator).prev_sample

            sample = pred_prev_sample

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

        assert abs(result_sum.item() - 202.0296) < 1e-2
        assert abs(result_mean.item() - 0.2631) < 1e-3

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

        timesteps = [100, 87, 50, 1, 0]

        scheduler.set_timesteps(timesteps=timesteps)

        scheduler_timesteps = scheduler.timesteps

        for i, timestep in enumerate(scheduler_timesteps):
            if i == len(timesteps) - 1:
                expected_prev_t = -1
            else:
                expected_prev_t = timesteps[i + 1]

            prev_t = scheduler.previous_timestep(timestep)
            prev_t = prev_t.item()

            self.assertEqual(prev_t, expected_prev_t)

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

        timesteps = [100, 87, 50, 51, 0]

        with self.assertRaises(ValueError, msg="`custom_timesteps` must be in descending order."):
            scheduler.set_timesteps(timesteps=timesteps)

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

        timesteps = [100, 87, 50, 1, 0]
        num_inference_steps = len(timesteps)

        with self.assertRaises(ValueError, msg="Can only pass one of `num_inference_steps` or `custom_timesteps`."):
            scheduler.set_timesteps(num_inference_steps=num_inference_steps, timesteps=timesteps)

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

        timesteps = [scheduler.config.num_train_timesteps]

        with self.assertRaises(
            ValueError,
            msg="`timesteps` must start before `self.config.train_timesteps`: {scheduler.config.num_train_timesteps}}",
        ):
            scheduler.set_timesteps(timesteps=timesteps)

    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_trained_timesteps = len(scheduler)
        t_start = num_trained_timesteps - 2

        model = self.dummy_model()
        sample = self.dummy_sample_deter
        generator = torch.manual_seed(0)

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

        for t in timesteps:
            # 1. predict noise residual
            residual = model(sample, t)

            # 2. predict previous mean of sample x_t-1
            pred_prev_sample = scheduler.step(residual, t, sample, generator=generator).prev_sample
            sample = pred_prev_sample

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

        assert abs(result_sum.item() - 387.9466) < 1e-2, f" expected result sum 387.9466, but get {result_sum}"
        assert abs(result_mean.item() - 0.5051) < 1e-3, f" expected result mean 0.5051, but get {result_mean}"