Deci-AI
/
super-gradients
connected to https://github.com/Deci-AI/super-gradients.git


  
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

	
            import unittest
from typing import Union

import pkg_resources
from hydra import initialize_config_dir
from hydra.core.global_hydra import GlobalHydra
from hydra import compose
from omegaconf import OmegaConf, open_dict, DictConfig
from super_gradients import Trainer, init_trainer
from super_gradients.common.registry.registry import register_pre_launch_callback
from super_gradients.training.pre_launch_callbacks import PreLaunchCallback
from super_gradients.common.environment.cfg_utils import normalize_path


@register_pre_launch_callback()
class PreLaunchTrainBatchSizeVerificationCallback(PreLaunchCallback):
    def __init__(self, batch_size):
        self.batch_size = batch_size

    def __call__(self, cfg: Union[dict, DictConfig]) -> Union[dict, DictConfig]:
        if cfg.dataset_params.train_dataloader_params.batch_size != self.batch_size:
            raise RuntimeError(f"Final selected batch size is {cfg.dataset_params.train_dataloader_params.batch_size}, expected: {self.batch_size}")
        return cfg


@register_pre_launch_callback()
class PreLaunchLRVerificationCallback(PreLaunchCallback):
    def __init__(self, lr):
        self.lr = lr

    def __call__(self, cfg: Union[dict, DictConfig]) -> Union[dict, DictConfig]:
        if cfg.training_hyperparams.initial_lr != self.lr:
            raise RuntimeError(f"Final selected lr is {cfg.training_hyperparams.initial_lr }, expected: {self.lr}")
        return cfg


class TestAutoBatchSelectionSingleGPU(unittest.TestCase):
    def test_auto_batch_size_no_max_no_lr_adaptation(self):
        GlobalHydra.instance().clear()
        sg_recipes_dir = pkg_resources.resource_filename("super_gradients.recipes", "")
        init_trainer()
        with initialize_config_dir(config_dir=normalize_path(sg_recipes_dir), version_base="1.2"):
            cfg = compose(config_name="cifar10_resnet")
            cfg.experiment_name = "batch_size_selection_test_no_max"
            cfg.training_hyperparams.max_epochs = 1
            OmegaConf.set_struct(cfg, True)
            with open_dict(cfg):
                cfg.pre_launch_callbacks_list = [
                    OmegaConf.create(
                        {"AutoTrainBatchSizeSelectionCallback": {"min_batch_size": 64, "size_step": 10000, "num_forward_passes": 3, "scale_lr": False}}
                    ),
                    OmegaConf.create({"PreLaunchTrainBatchSizeVerificationCallback": {"batch_size": 64}}),
                    OmegaConf.create({"PreLaunchLRVerificationCallback": {"lr": cfg.training_hyperparams.initial_lr}}),
                ]
        Trainer.train_from_config(cfg)

    def test_auto_batch_size_with_upper_limit_no_lr_adaptation(self):
        GlobalHydra.instance().clear()
        sg_recipes_dir = pkg_resources.resource_filename("super_gradients.recipes", "")
        init_trainer()
        with initialize_config_dir(config_dir=normalize_path(sg_recipes_dir), version_base="1.2"):
            cfg = compose(config_name="cifar10_resnet")
            cfg.experiment_name = "batch_size_selection_test_with_upper_limit"
            cfg.training_hyperparams.max_epochs = 1
            OmegaConf.set_struct(cfg, True)
            with open_dict(cfg):
                cfg.pre_launch_callbacks_list = [
                    OmegaConf.create(
                        {
                            "AutoTrainBatchSizeSelectionCallback": {
                                "min_batch_size": 32,
                                "size_step": 32,
                                "max_batch_size": 64,
                                "num_forward_passes": 3,
                                "scale_lr": False,
                                "mode": "largest",
                            }
                        }
                    ),
                    OmegaConf.create({"PreLaunchTrainBatchSizeVerificationCallback": {"batch_size": 64}}),
                    OmegaConf.create({"PreLaunchLRVerificationCallback": {"lr": cfg.training_hyperparams.initial_lr}}),
                    OmegaConf.create({"PreLaunchLRVerificationCallback": {"lr": cfg.training_hyperparams.initial_lr}}),
                ]
        Trainer.train_from_config(cfg)

    def test_auto_batch_size_no_max_with_lr_adaptation(self):
        GlobalHydra.instance().clear()
        sg_recipes_dir = pkg_resources.resource_filename("super_gradients.recipes", "")
        init_trainer()
        with initialize_config_dir(config_dir=normalize_path(sg_recipes_dir), version_base="1.2"):
            cfg = compose(config_name="cifar10_resnet")
            cfg.experiment_name = "batch_size_selection_test_no_max"
            cfg.training_hyperparams.max_epochs = 1
            OmegaConf.set_struct(cfg, True)
            with open_dict(cfg):
                cfg.pre_launch_callbacks_list = [
                    OmegaConf.create(
                        {"AutoTrainBatchSizeSelectionCallback": {"min_batch_size": 64, "size_step": 10000, "num_forward_passes": 3, "mode": "largest"}}
                    ),
                    OmegaConf.create({"PreLaunchTrainBatchSizeVerificationCallback": {"batch_size": 64}}),
                    OmegaConf.create(
                        {
                            "PreLaunchLRVerificationCallback": {
                                "lr": cfg.training_hyperparams.initial_lr * 64 / cfg.dataset_params.train_dataloader_params.batch_size
                            }
                        }
                    ),
                ]
        Trainer.train_from_config(cfg)

    def test_auto_batch_size_with_upper_limit_with_lr_adaptation(self):
        GlobalHydra.instance().clear()
        sg_recipes_dir = pkg_resources.resource_filename("super_gradients.recipes", "")
        init_trainer()
        with initialize_config_dir(config_dir=normalize_path(sg_recipes_dir), version_base="1.2"):
            cfg = compose(config_name="cifar10_resnet")
            cfg.experiment_name = "batch_size_selection_test_with_upper_limit"
            cfg.training_hyperparams.max_epochs = 1
            OmegaConf.set_struct(cfg, True)
            with open_dict(cfg):
                cfg.pre_launch_callbacks_list = [
                    OmegaConf.create(
                        {
                            "AutoTrainBatchSizeSelectionCallback": {
                                "min_batch_size": 32,
                                "size_step": 32,
                                "max_batch_size": 64,
                                "num_forward_passes": 3,
                                "mode": "largest",
                            }
                        }
                    ),
                    OmegaConf.create({"PreLaunchTrainBatchSizeVerificationCallback": {"batch_size": 64}}),
                    OmegaConf.create(
                        {
                            "PreLaunchLRVerificationCallback": {
                                "lr": cfg.training_hyperparams.initial_lr * 64 / cfg.dataset_params.train_dataloader_params.batch_size
                            }
                        }
                    ),
                ]
        Trainer.train_from_config(cfg)


if __name__ == "__main__":
    unittest.main()