SSCBM/main.py at main · Skyyyy0920/SSCBM · GitHub

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
import yaml
import zipfile
from collections import defaultdict

from configs.basic_config import *
from utils import *
from train.training import *
from train.evaluate import *
import interventions.utils as intervention_utils

if __name__ == '__main__':
    # ==================================================================================================
    # 1. Get experiment args and seed
    # ==================================================================================================
    args = get_args()
    logging_time = time.strftime('%H-%M', time.localtime())
    save_dir = os.path.join(args.save_path, f"{args.dataset}_{logging_time}")
    logging_config(save_dir)
    logging.info(f"args: {args}")
    logging.info(f"Saving path: {save_dir}")
    with open(f"configs/{args.dataset}.yaml", "r") as f:
        experiment_config = yaml.load(f, Loader=yaml.FullLoader)

    logging.info(f"GPU number: {torch.cuda.device_count()}")
    # logging.info(f"current GPU: {torch.cuda.current_device()}")

    # ==================================================================================================
    # 2. Save codes and settings
    # ==================================================================================================
    zipf = zipfile.ZipFile(file=os.path.join(save_dir, 'codes.zip'), mode='a', compression=zipfile.ZIP_DEFLATED)
    zipdir(Path().absolute(), zipf, include_format=['.py'])
    zipf.close()
    with open(os.path.join(save_dir, 'args.yml'), 'a') as f:
        yaml.dump(vars(args), f, sort_keys=False)
    with open(os.path.join(save_dir, f"experiment_config.yaml"), "w") as f:
        yaml.dump(experiment_config, f)

    # ==================================================================================================
    # 3. Prepare data
    # ==================================================================================================
    (
        train_dl,
        val_dl,
        test_dl,
        imbalance,
        concept_map,
        intervened_groups,
        task_class_weights,
        acquisition_costs
    ) = generate_dataset_and_update_config(experiment_config, args)

    # ==================================================================================================
    # 4. Build models, define overall loss and optimizer. Then training
    # ==================================================================================================
    results = defaultdict(dict)
    for current_config in experiment_config['runs']:
        run_name = current_config['architecture']
        trial_config = copy.deepcopy(experiment_config)
        trial_config.update(current_config)

        for run_config in generate_hyper_param_configs(trial_config):
            run_config = copy.deepcopy(run_config)
            run_config['result_dir'] = save_dir
            run_config["c_extractor_arch"] = args.image_encoder
            evaluate_expressions(run_config, soft=True)

            old_results = None
            model, model_results = train_end_to_end_model(
                run_name=run_name,
                task_class_weights=task_class_weights,
                accelerator=args.device,
                devices='auto',
                n_concepts=run_config['n_concepts'],
                n_tasks=run_config['n_tasks'],
                config=run_config,
                train_dl=train_dl,
                val_dl=val_dl,
                test_dl=test_dl,
                result_dir=save_dir,
                seed=args.seed,
                imbalance=imbalance,
                old_results=old_results,
                gradient_clip_val=run_config.get('gradient_clip_val', 0),
                activation_freq=args.activation_freq,
                single_frequency_epochs=args.single_frequency_epochs,
            )
            continue
            if 'intervention_config' in run_config:
                intervention_config = run_config['intervention_config']
                test_int_args = dict(
                    task_class_weights=task_class_weights,
                    run_name=run_name,
                    train_dl=train_dl,
                    val_dl=val_dl,
                    test_dl=test_dl,
                    imbalance=imbalance,
                    config=run_config,
                    n_tasks=run_config['n_tasks'],
                    n_concepts=run_config['n_concepts'],
                    acquisition_costs=None,
                    result_dir=save_dir,
                    concept_map=concept_map,
                    intervened_groups=intervened_groups,
                    accelerator=args.device,
                    devices='auto',
                    split=0,
                    rerun=False,
                    old_results=old_results,
                    group_level_competencies=intervention_config.get("group_level_competencies", False),
                    competence_levels=intervention_config.get('competence_levels', [1]),
                )
                if "real_competencies" in intervention_config:
                    for real_comp in intervention_config['real_competencies']:
                        def _real_competence_generator(x):
                            if real_comp == "same":
                                return x
                            if real_comp == "complement":
                                return 1 - x
                            if test_int_args['group_level_competencies']:
                                if real_comp == "unif":
                                    batch_group_level_competencies = np.zeros((x.shape[0], len(concept_map)))
                                    for batch_idx in range(x.shape[0]):
                                        for group_idx, (_, concept_members) in enumerate(concept_map.items()):
                                            batch_group_level_competencies[
                                                batch_idx,
                                                group_idx,
                                            ] = np.random.uniform(1 / len(concept_members), 1)
                                else:
                                    batch_group_level_competencies = np.ones((x.shape[0], len(concept_map))) * real_comp
                                return batch_group_level_competencies

                            if real_comp == "unif":
                                return np.random.uniform(0.5, 1, size=x.shape)
                            return np.ones(x.shape) * real_comp


                        if real_comp == "same":
                            # Then we will just run what we normally run as the provided competency matches the level
                            # of competency of the user
                            test_int_args.pop("real_competence_generator", None)
                            test_int_args.pop("extra_suffix", None)
                            test_int_args.pop("real_competence_level", None)
                        else:
                            test_int_args['real_competence_generator'] = _real_competence_generator
                            test_int_args['extra_suffix'] = f"_real_comp_{real_comp}_"
                            test_int_args["real_competence_level"] = real_comp

                update_statistics(
                    aggregate_results=results[run_name],
                    run_config=run_config,
                    model=model,
                    test_results=intervention_utils.test_interventions(**test_int_args),
                    run_name=run_name,
                    prefix="",
                )

            update_statistics(
                aggregate_results=results[run_name],
                run_config=run_config,
                model=model,
                test_results=evaluate_representation_metrics(
                    config=run_config,
                    n_concepts=run_config['n_concepts'],
                    n_tasks=run_config['n_tasks'],
                    test_dl=test_dl,
                    run_name=run_name,
                    imbalance=imbalance,
                    result_dir=save_dir,
                    task_class_weights=task_class_weights,
                    accelerator=args.device,
                    devices='auto',
                    seed=args.seed,
                    old_results=old_results,
                ),
                run_name=run_name,
                prefix="",
            )
            results[run_name][f'num_trainable_params'] = \
                sum(p.numel() for p in model.parameters() if p.requires_grad)
            results[run_name][f'num_non_trainable_params'] = \
                sum(p.numel() for p in model.parameters() if not p.requires_grad)

        with open(f'{save_dir}/results.txt', 'w') as f:
            for key, value in results[run_name].items():
                f.write(f"{key}: {value}\n")

    print(f"========================finish========================")