bilevel_design_opt/train_model.py at main · song970407/bilevel_design_opt · 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
import os
import yaml
import pickle
import argparse

from os.path import join
from time import perf_counter

import torch
import wandb
import dgl
from torch.utils.data import TensorDataset, DataLoader

from src.utils.preprocess_data import preprocess_data
from src.utils.generate_decaying_coefficient import generate_decaying_coefficient
from src.model.get_model import get_model
from src.utils.penalty_utils import get_nonnegative_penalty, project_params
from src.utils.fix_seed import fix_seed

device = 'cuda:0' if torch.cuda.is_available() else 'cpu'


def train_model(model_name, train_data, val_data, data_preprocessing_config):
    if not os.path.exists('saved_model/{}'.format(model_name)):
        os.makedirs('saved_model/{}'.format(model_name))
    model_config = yaml.safe_load(open('config/model/{}/model_config.yaml'.format(model_name), 'r'))
    train_config = yaml.safe_load(open('config/model/{}/train_config.yaml'.format(model_name), 'r'))

    model_saved_path = model_config['model_saved_path']
    history_len = train_config['history_len']
    receding_horizon = train_config['receding_horizon']
    decaying_gammas = train_config['decaying_gammas']
    epoch = train_config['epoch']
    bs = train_config['bs']
    test_every = train_config['test_every']
    opt_name = train_config['opt_name']
    opt_config = train_config['opt_config']
    seed_num = train_config['seed_num']
    fix_seed(seed_num)

    data_preprocessing_config['history_len'] = history_len
    data_preprocessing_config['receding_horizon'] = receding_horizon
    data_preprocessing_config['device'] = device

    train_hist_xs, train_hist_us, train_future_us, train_future_xs, train_gs, train_idxs = preprocess_data(train_data,
                                                                                                           data_preprocessing_config)
    val_hist_xs, val_hist_us, val_future_us, val_future_xs, val_gs, val_idxs = preprocess_data(val_data,
                                                                                               data_preprocessing_config)
    with torch.no_grad():
        vg = dgl.batch([val_gs[idx[0]] for idx in val_idxs])
        vhx = torch.cat([val_hist_xs[idx[0]][idx[1]] for idx in val_idxs])
        vhu = torch.cat([val_hist_us[idx[0]][idx[1]] for idx in val_idxs])
        vfu = torch.cat([val_future_us[idx[0]][idx[1]] for idx in val_idxs])
        vfx = torch.cat([val_future_xs[idx[0]][idx[1]] for idx in val_idxs])
    train_dataset = TensorDataset(train_idxs)
    train_dl = DataLoader(train_dataset, batch_size=bs, shuffle=True)
    iters = len(train_dl)

    m = get_model(model_name, model_config).to(device)
    m.train()

    config = {
        'model_name': model_name,
        'model_config': model_config,
        'train_config': train_config,
        'data': {
            'train_data': train_data,
            'val_data': val_data,
            'data_preprocessing_config': data_preprocessing_config
        }
    }
    run = wandb.init(config=config,
                     project='Bilevel_Design_Opt',
                     entity='55mong',
                     reinit=True,
                     name=model_name)

    num_updates = 0
    val_best_loss = float('inf')
    val_crit = torch.nn.SmoothL1Loss()

    for decaying_gamma in decaying_gammas:
        opt = getattr(torch.optim, opt_name)(m.parameters(), lr=opt_config['lr'])
        decaying_coefficient = generate_decaying_coefficient(decaying_gamma, receding_horizon, device)

        def crit(x, y):
            loss_fn = torch.nn.SmoothL1Loss(reduction='none')
            loss = loss_fn(x, y).mean(dim=(0, 2))
            return (decaying_coefficient * loss).mean()

        for ep in range(epoch):
            if ep % 10 == 0:
                print('Epoch [{}] / [{}]'.format(ep, epoch))
            for i, (train_idx,) in enumerate(train_dl):
                start_time = perf_counter()
                tg = dgl.batch([train_gs[idx[0]] for idx in train_idx])
                thx = torch.cat([train_hist_xs[idx[0]][idx[1]] for idx in train_idx])
                thu = torch.cat([train_hist_us[idx[0]][idx[1]] for idx in train_idx])
                tfu = torch.cat([train_future_us[idx[0]][idx[1]] for idx in train_idx])
                tfx = torch.cat([train_future_xs[idx[0]][idx[1]] for idx in train_idx])
                pfx = m.multistep_prediction(tg, thx, thu, tfu)
                train_loss = crit(tfx, pfx)
                opt.zero_grad()
                train_loss.backward()
                opt.step()
                project_params(m)
                num_updates += 1
                log = {
                    'Epoch': ep + i / iters,
                    'fit_time': perf_counter() - start_time,
                    'train_loss': train_loss.item()
                }
                if num_updates % test_every == 0:
                    with torch.no_grad():
                        val_loss = val_crit(m.multistep_prediction(vg, vhx, vhu, vfu), vfx)
                    if val_loss.item() < val_best_loss:
                        val_best_loss = val_loss.item()
                        torch.save(m.state_dict(), model_saved_path)
                    torch.save(m.state_dict(), join(wandb.run.dir, 'model.pt'))
                    log['val_loss'] = val_loss.item()
                    log['val_best_loss'] = val_best_loss
                wandb.log(log)
    run.finish()


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--model_name', default='Linear')
    args = parser.parse_args()

    data_generation_config = yaml.safe_load(open('config/data/data_generation_config.yaml', 'r'))
    data_preprocessing_config = yaml.safe_load(open('config/data/data_preprocessing_config.yaml', 'r'))
    train_data = pickle.load(open(data_generation_config['train_data_saved_path'], 'rb'))
    val_data = pickle.load(open(data_generation_config['val_data_saved_path'], 'rb'))
    train_model(args.model_name, train_data, val_data, data_preprocessing_config)