MC-Net/train_obsnet.py at main · QQQYang/MC-Net · 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
188
189
#encoding=utf-8

import torch
from torch import nn
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
import os
import numpy as np
import json
import shutil

from model.obs_net import ObsNet, ObsAverageNet
from multimodal_data import MeaIdealDataset, MeaIdealLargeDataset
from sklearn.metrics import r2_score

from arguments import args
from util.logger import setup_logger
from util.config import configs

logger = setup_logger(args.log_dir, name='measurement')

def train(model, loader, loader_test, args, board):
    model.train()
    device = "cuda" if torch.cuda.is_available() else "cpu"
    if args.train.cuda and torch.cuda.is_available():
        model.cuda()

    loss_fn = nn.MSELoss()

    params = [
        {'params': model.parameters(), 'initial_lr': args.train.lr, 'lr': args.train.lr},
    ]

    optimizer = torch.optim.Adam(params, betas=(0.5, 0.999), weight_decay=1e-4)
    scheduler = torch.optim.lr_scheduler.StepLR(optimizer, args.train.decay_step, gamma=0.5, last_epoch=-1)
    res_test = {}
    res_train = {}
    for i in range(args.train.n_epoch):
        if device == 'cuda':
            model.cuda()

        for mea_vec1, mea_vec2, label, _ in loader:
            if args.dataset.obs_feat_dim == 1:
                feat1 = model(mea_vec1.unsqueeze(-1).to(device))
                feat2 = model(mea_vec2.unsqueeze(-1).to(device))
            else:
                feat1 = model(mea_vec1.to(device))
                feat2 = model(mea_vec2.to(device))
            sim = (torch.sum(torch.mul(feat1, feat2), dim=1) + 1)/2

            loss = loss_fn(sim, label.to(device))

            optimizer.zero_grad()
            loss.backward()
            optimizer.step()
        lr = optimizer.param_groups[0]['lr']
        logger.info(f'Epoch={i}, loss={loss.item()}, lr={lr}')
        board.add_scalar('loss', loss.item(), i+1)

        scheduler.step()
        if (i+1) % args.train.save_freq == 0:
            preds, labels = [], []
            with torch.no_grad():
                model.eval()
                for mea_vec1, mea_vec2, label, _ in loader_test:
                    feat1 = model(mea_vec1.to(device))
                    feat2 = model(mea_vec2.to(device))
                    sim = (torch.sum(torch.mul(feat1, feat2), dim=1)+1)/2
                    preds.append(sim.cpu().detach().numpy())
                    labels.append(label.cpu().detach().numpy())
                preds = np.concatenate(preds, axis=0)
                labels = np.concatenate(labels, axis=0)
                mse = np.mean((preds-labels)**2)
                mae = np.mean(np.abs(preds-labels))
                logger.info(f'Test mse = {mse}, mae={mae}, r2={r2_score(labels, preds)}')
                model.train()
                res_test[str(i+1)] = mse.tolist()

                preds, labels = [], []
                for mea_vec1, mea_vec2, label, _ in loader:
                    feat1 = model(mea_vec1.to(device))
                    feat2 = model(mea_vec2.to(device))
                    sim = (torch.sum(torch.mul(feat1, feat2), dim=1)+1)/2
                    preds.append(sim.cpu().detach().numpy())
                    labels.append(label.cpu().detach().numpy())
                preds = np.concatenate(preds, axis=0)
                labels = np.concatenate(labels, axis=0)
                mse = np.mean((preds-labels)**2)
                res_train[str(i+1)] = mse.tolist()

    torch.save(model.cpu().state_dict(), os.path.join(args.log_dir, f'{args.model.net_type}_c{args.dataset.n_circuit}_s{args.dataset.shadow_size}_e{args.train.n_epoch}.pth'))
    return res_test, res_train

def test(model, loader):
    model.eval()
    device = "cuda" if torch.cuda.is_available() else "cpu"
    if args.cuda and torch.cuda.is_available():
        model.cuda()

    preds, labels, keys = [], [], []
    feat_query, feat_ref = [], []
    with torch.no_grad():
        for mea_vec1, mea_vec2, label, key in loader:
            if args.obs_feat_dim == 1:
                feat1 = model(mea_vec1.unsqueeze(-1).to(device))
                feat2 = model(mea_vec2.unsqueeze(-1).to(device))
            else:
                feat1 = model(mea_vec1.to(device))
                feat2 = model(mea_vec2.to(device))
            sim = (torch.sum(torch.mul(feat1, feat2), dim=1)+1)/2
            preds.append(sim.cpu().detach().numpy())
            labels.append(label.cpu().detach().numpy())
            feat_query.append(feat1.cpu().detach().numpy())
            feat_ref.append(feat2.cpu().detach().numpy())
            keys = keys + key
        return np.concatenate(preds, axis=0), np.concatenate(labels, axis=0), keys, np.concatenate(feat_query, axis=0), np.concatenate(feat_ref, axis=0)


if __name__ == '__main__':
    torch.manual_seed(args.seed)
    if torch.cuda.is_available():
        if not args.cuda:
            logger.info("WARNING: You have a CUDA device, so you should probably run with --cuda")

    configs.load(args.config, recursive=True)

    logger.info('begin')
    logger.info(f'arguments = {configs}')
    args.log_dir = os.path.join(args.log_dir, '/'.join(args.config.split('/')[1:])[:-5])
    configs.log_dir = args.log_dir
    if not os.path.exists(args.log_dir):
        os.makedirs(args.log_dir)
    shutil.copy(args.config, args.log_dir)

    if configs.dataset.average:
        model = ObsAverageNet(configs.dataset.obs_feat_dim, configs.model.dims)
    else:
        model = ObsNet(configs.dataset.obs_feat_dim, configs.model.dims)

    if args.phase == 'train':
        if args.retrain:
            model.load_state_dict(torch.load(os.path.join(args.log_dir, f'{configs.model.net_type}_c{configs.dataset.n_circuit}_s{configs.dataset.shadow_size}_e{configs.train.n_epoch}.pth')))
        names = os.listdir(args.log_dir)
        for name in names:
            if 'event' in name:
                os.remove(os.path.join(args.log_dir, name))
        board = SummaryWriter(log_dir=args.log_dir)

        logger.info('---------build dataset----------------')
        if configs.model.loss_type == 'rl':
            if configs.dataset.split:
                train_dataset = MeaIdealLargeDataset(configs.dataset, test=False)
            else:
                train_dataset = MeaIdealDataset(configs.dataset, test=False)
            test_file_name = configs.dataset.mea_data.replace('train', 'test')
            test_file_name = test_file_name.replace('p1000', 'p100')
            configs.dataset.mea_data = test_file_name
            if configs.dataset.split:
                test_dataset = MeaIdealLargeDataset(configs.dataset, test=True)
            else:
                test_dataset = MeaIdealDataset(configs.dataset, test=True)
            logger.info(f'Train={len(train_dataset)}, test={len(test_dataset)}')
        train_loader =  DataLoader(train_dataset, batch_size=configs.train.bs, shuffle=True, num_workers=configs.train.n_worker, pin_memory=True)
        test_loader =  DataLoader(test_dataset, batch_size=configs.train.bs, shuffle=True, num_workers=configs.train.n_worker, pin_memory=True)

        res_test, res_train = train(model, train_loader, test_loader, configs, board)
        with open(os.path.join(args.log_dir, f'test_mea_epoch.json'), 'w') as f:
            json.dump(res_test, f)
        with open(os.path.join(args.log_dir, f'train_mea_epoch.json'), 'w') as f:
            json.dump(res_train, f)
    elif args.phase == 'test':
        model.load_state_dict(torch.load(os.path.join(args.log_dir, f'{configs.model.net_type}_c{configs.dataset.n_circuit}_s{configs.dataset.shadow_size}_e{configs.train.n_epoch}.pth')))

        test_file_name = configs.dataset.mea_data.replace('train', 'test')
        test_file_name = test_file_name.replace('p1000', 'p100')
        configs.dataset.mea_data = test_file_name
        if configs.dataset.split:
            test_dataset = MeaIdealLargeDataset(configs.dataset, test=True)
        else:
            test_dataset = MeaIdealDataset(configs.dataset, test=True)
        test_loader =  DataLoader(test_dataset, batch_size=configs.train.bs, shuffle=False, num_workers=configs.train.n_worker, pin_memory=True)
        preds, labels, keys, feat_query, feat_ref = test(model, test_loader)
        mse = np.mean((preds-labels)**2)
        mae = np.mean(np.abs(preds-labels))
        logger.info(f'{len(preds)} samples, MSE={mse}, MAE={mae}, r2={r2_score(labels, preds)}')

        res = {}
        for i, key in enumerate(keys):
            res[key] = {'pred': preds[i].tolist(), 'label': labels[i].tolist()}