XAI-diffusion/tune_ddpm.py at master · LSX-UniWue/XAI-diffusion · 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
import torch
import lib
import os
import optuna
import shutil
import argparse
from pathlib import Path

from train import train
from pipeline import save_file

parser = argparse.ArgumentParser()
parser.add_argument('ds_name', type=str)
parser.add_argument('prefix', type=str)

args = parser.parse_args()
ds_name = args.ds_name
prefix = str(args.prefix)

pipeline = f'pipeline.py'
base_config_path = f'exp/{ds_name}/config.toml'
parent_path = Path(f'exp/{ds_name}/')
exps_path = Path(f'exp/{ds_name}/many-exps/')

os.makedirs(exps_path, exist_ok=True)
os.makedirs(parent_path / prefix, exist_ok=True)

param_score_path = parent_path / prefix / 'val_loss_params_score.json'
params_and_score = {}
lib.dump_json(params_and_score, param_score_path)


def _suggest_mlp_layers(trial):
    def suggest_dim(name):
        t = trial.suggest_int(name, d_min, d_max)
        return 2 ** t

    min_n_layers, max_n_layers, d_min, d_max = 1, 4, 7, 10
    n_layers = 2 * trial.suggest_int('n_layers', min_n_layers, max_n_layers)
    d_first = [suggest_dim('d_first')] if n_layers else []
    d_middle = (
        [suggest_dim('d_middle')] * (n_layers - 2)
        if n_layers > 2
        else []
    )
    d_last = [suggest_dim('d_last')] if n_layers > 1 else []
    d_layers = d_first + d_middle + d_last
    return d_layers


def objective(trial):
    lr = trial.suggest_loguniform('lr', 0.00001, 0.003)
    d_layers = _suggest_mlp_layers(trial)
    steps = trial.suggest_categorical('steps', [5000, 20000, 30000])
    num_timesteps = trial.suggest_categorical('num_timesteps', [100, 250, 1000])

    params_and_score = lib.load_json(param_score_path)
    params_and_score[trial.number] = {
        'params': {
            'lr': lr,
            'steps': steps,
            'num_timesteps_train': num_timesteps,
        },
        'val_loss': -1
    }
    lib.dump_json(params_and_score, param_score_path)

    base_config = lib.load_config(base_config_path)

    base_config['train']['main']['lr'] = lr
    base_config['train']['main']['steps'] = steps
    base_config['model_params']['rtdl_params']['d_layers'] = d_layers
    base_config['diffusion_params']['num_timesteps'] = num_timesteps

    base_config['parent_dir'] = str(exps_path / f"{trial.number}")

    trial.set_user_attr("config", base_config)

    lib.dump_config(base_config, exps_path / 'config.toml')

    raw_config = lib.load_config(f'{exps_path / "config.toml"}')
    device = torch.device(raw_config['device']) if 'device' in raw_config else torch.device('cuda')
    save_file(os.path.join(raw_config['parent_dir'], 'config.toml'), f'{exps_path / "config.toml"}')

    train(**raw_config['train']['main'],
          **raw_config['diffusion_params'],
          parent_dir=raw_config['parent_dir'],
          data_name=raw_config['data_name'],
          real_data_path=raw_config['real_data_path'],
          model_type=raw_config['model_type'],
          model_params=raw_config['model_params'],
          device=device,
          validation=True,
          cat_encoding=raw_config['cat_encoding'],
          num_encoding=raw_config['num_encoding'])

    save_file(os.path.join(raw_config['parent_dir'], 'info.json'),
              os.path.join(raw_config['real_data_path'], 'info.json'))

    report_path = str(Path(base_config['parent_dir']) / 'val_loss.json')
    report = lib.load_json(report_path)

    shutil.rmtree(exps_path / f"{trial.number}")

    loss = report['multi'] + report['gauss']

    params_and_score = lib.load_json(param_score_path)
    params_and_score[str(trial.number)]['val_loss'] = loss
    lib.dump_json(params_and_score, param_score_path)

    return loss


study = optuna.create_study(
    direction="minimize",
    sampler=optuna.samplers.TPESampler(seed=0),
)

study.optimize(objective, n_trials=100, show_progress_bar=True)

best_config_path = parent_path / f'{prefix}_best/config.toml'
best_config = study.best_trial.user_attrs['config']
best_config["parent_dir"] = str(parent_path / f'{prefix}_best/')

os.makedirs(parent_path / f'{prefix}_best', exist_ok=True)
lib.dump_config(best_config, best_config_path)

raw_config = lib.load_config(f'{best_config_path}')
device = torch.device(raw_config['device']) if 'device' in raw_config else torch.device('cuda')
save_file(os.path.join(raw_config['parent_dir'], 'config.toml'), f'{best_config_path}')

train(**raw_config['train']['main'],
      **raw_config['diffusion_params'],
      parent_dir=raw_config['parent_dir'],
      data_name=raw_config['data_name'],
      real_data_path=raw_config['real_data_path'],
      model_type=raw_config['model_type'],
      model_params=raw_config['model_params'],
      device=device,
      validation=False,
      cat_encoding=raw_config['cat_encoding'],
      num_encoding=raw_config['num_encoding'])

save_file(os.path.join(raw_config['parent_dir'], 'info.json'),
          os.path.join(raw_config['real_data_path'], 'info.json'))