CoCalc -- test_scheduler

GitHub Repository: shivamshrirao/diffusers
Path: blob/main/tests/schedulers/test_scheduler_ddim.py
¹⁴⁴⁸ views
1
import torch
2

3
from diffusers import DDIMScheduler
4

5
from .test_schedulers import SchedulerCommonTest
6

7

8
class DDIMSchedulerTest(SchedulerCommonTest):
9
    scheduler_classes = (DDIMScheduler,)
10
    forward_default_kwargs = (("eta", 0.0), ("num_inference_steps", 50))
11

12
    def get_scheduler_config(self, **kwargs):
13
        config = {
14
            "num_train_timesteps": 1000,
15
            "beta_start": 0.0001,
16
            "beta_end": 0.02,
17
            "beta_schedule": "linear",
18
            "clip_sample": True,
19
        }
20

21
        config.update(**kwargs)
22
        return config
23

24
    def full_loop(self, **config):
25
        scheduler_class = self.scheduler_classes[0]
26
        scheduler_config = self.get_scheduler_config(**config)
27
        scheduler = scheduler_class(**scheduler_config)
28

29
        num_inference_steps, eta = 10, 0.0
30

31
        model = self.dummy_model()
32
        sample = self.dummy_sample_deter
33

34
        scheduler.set_timesteps(num_inference_steps)
35

36
        for t in scheduler.timesteps:
37
            residual = model(sample, t)
38
            sample = scheduler.step(residual, t, sample, eta).prev_sample
39

40
        return sample
41

42
    def test_timesteps(self):
43
        for timesteps in [100, 500, 1000]:
44
            self.check_over_configs(num_train_timesteps=timesteps)
45

46
    def test_steps_offset(self):
47
        for steps_offset in [0, 1]:
48
            self.check_over_configs(steps_offset=steps_offset)
49

50
        scheduler_class = self.scheduler_classes[0]
51
        scheduler_config = self.get_scheduler_config(steps_offset=1)
52
        scheduler = scheduler_class(**scheduler_config)
53
        scheduler.set_timesteps(5)
54
        assert torch.equal(scheduler.timesteps, torch.LongTensor([801, 601, 401, 201, 1]))
55

56
    def test_betas(self):
57
        for beta_start, beta_end in zip([0.0001, 0.001, 0.01, 0.1], [0.002, 0.02, 0.2, 2]):
58
            self.check_over_configs(beta_start=beta_start, beta_end=beta_end)
59

60
    def test_schedules(self):
61
        for schedule in ["linear", "squaredcos_cap_v2"]:
62
            self.check_over_configs(beta_schedule=schedule)
63

64
    def test_prediction_type(self):
65
        for prediction_type in ["epsilon", "v_prediction"]:
66
            self.check_over_configs(prediction_type=prediction_type)
67

68
    def test_clip_sample(self):
69
        for clip_sample in [True, False]:
70
            self.check_over_configs(clip_sample=clip_sample)
71

72
    def test_thresholding(self):
73
        self.check_over_configs(thresholding=False)
74
        for threshold in [0.5, 1.0, 2.0]:
75
            for prediction_type in ["epsilon", "v_prediction"]:
76
                self.check_over_configs(
77
                    thresholding=True,
78
                    prediction_type=prediction_type,
79
                    sample_max_value=threshold,
80
                )
81

82
    def test_time_indices(self):
83
        for t in [1, 10, 49]:
84
            self.check_over_forward(time_step=t)
85

86
    def test_inference_steps(self):
87
        for t, num_inference_steps in zip([1, 10, 50], [10, 50, 500]):
88
            self.check_over_forward(time_step=t, num_inference_steps=num_inference_steps)
89

90
    def test_eta(self):
91
        for t, eta in zip([1, 10, 49], [0.0, 0.5, 1.0]):
92
            self.check_over_forward(time_step=t, eta=eta)
93

94
    def test_variance(self):
95
        scheduler_class = self.scheduler_classes[0]
96
        scheduler_config = self.get_scheduler_config()
97
        scheduler = scheduler_class(**scheduler_config)
98

99
        assert torch.sum(torch.abs(scheduler._get_variance(0, 0) - 0.0)) < 1e-5
100
        assert torch.sum(torch.abs(scheduler._get_variance(420, 400) - 0.14771)) < 1e-5
101
        assert torch.sum(torch.abs(scheduler._get_variance(980, 960) - 0.32460)) < 1e-5
102
        assert torch.sum(torch.abs(scheduler._get_variance(0, 0) - 0.0)) < 1e-5
103
        assert torch.sum(torch.abs(scheduler._get_variance(487, 486) - 0.00979)) < 1e-5
104
        assert torch.sum(torch.abs(scheduler._get_variance(999, 998) - 0.02)) < 1e-5
105

106
    def test_full_loop_no_noise(self):
107
        sample = self.full_loop()
108

109
        result_sum = torch.sum(torch.abs(sample))
110
        result_mean = torch.mean(torch.abs(sample))
111

112
        assert abs(result_sum.item() - 172.0067) < 1e-2
113
        assert abs(result_mean.item() - 0.223967) < 1e-3
114

115
    def test_full_loop_with_v_prediction(self):
116
        sample = self.full_loop(prediction_type="v_prediction")
117

118
        result_sum = torch.sum(torch.abs(sample))
119
        result_mean = torch.mean(torch.abs(sample))
120

121
        assert abs(result_sum.item() - 52.5302) < 1e-2
122
        assert abs(result_mean.item() - 0.0684) < 1e-3
123

124
    def test_full_loop_with_set_alpha_to_one(self):
125
        # We specify different beta, so that the first alpha is 0.99
126
        sample = self.full_loop(set_alpha_to_one=True, beta_start=0.01)
127
        result_sum = torch.sum(torch.abs(sample))
128
        result_mean = torch.mean(torch.abs(sample))
129

130
        assert abs(result_sum.item() - 149.8295) < 1e-2
131
        assert abs(result_mean.item() - 0.1951) < 1e-3
132

133
    def test_full_loop_with_no_set_alpha_to_one(self):
134
        # We specify different beta, so that the first alpha is 0.99
135
        sample = self.full_loop(set_alpha_to_one=False, beta_start=0.01)
136
        result_sum = torch.sum(torch.abs(sample))
137
        result_mean = torch.mean(torch.abs(sample))
138

139
        assert abs(result_sum.item() - 149.0784) < 1e-2
140
        assert abs(result_mean.item() - 0.1941) < 1e-3
141

142
Product

Resources

Company
