DeepMETv2/train.py at master · erdemyigit/DeepMETv2 · 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 json
import os.path as osp
import os
import numpy as np
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch_geometric.utils import to_undirected
from torch_cluster import radius_graph, knn_graph
from torch_geometric.datasets import MNISTSuperpixels
import torch_geometric.transforms as T
from torch_geometric.data import DataLoader
from tqdm import tqdm
import argparse
import utils
import model.net as net
import model.data_loader as data_loader
from evaluate import evaluate
import warnings
warnings.simplefilter('ignore')
from time import strftime, gmtime

parser = argparse.ArgumentParser()
parser.add_argument('--restore_file', default=None,
                    help="Optional, name of the file in --model_dir containing weights to reload before \
                    training")  # 'best' or 'train'
parser.add_argument('--data', default='data',
                    help="Name of the data folder")
parser.add_argument('--ckpts', default='ckpts',
                    help="Name of the ckpts folder")


def train(model, device, optimizer, scheduler, loss_fn, dataloader, epoch):
    model.train()
    loss_avg_arr = []
    loss_avg = utils.RunningAverage()
    with tqdm(total=len(dataloader)) as t:
        for data in dataloader:
            optimizer.zero_grad()
            data = data.to(device)
            x_cont = data.x[:,:8] #include puppi
            #x_cont = data.x[:,:7] #remove puppi
            x_cat = data.x[:,8:].long()
            phi = torch.atan2(data.x[:,1], data.x[:,0])
            etaphi = torch.cat([data.x[:,3][:,None], phi[:,None]], dim=1)
            # NB: there is a problem right now for comparing hits at the +/- pi boundary
            edge_index = radius_graph(etaphi, r=deltaR, batch=data.batch, loop=True, max_num_neighbors=255)
            result = model(x_cont, x_cat, edge_index, data.batch)
            loss = loss_fn(result, data.x, data.y, data.batch)
            loss.backward()
            optimizer.step()
            # update the average loss
            loss_avg_arr.append(loss.item())
            loss_avg.update(loss.item())
            t.set_postfix(loss='{:05.3f}'.format(loss_avg()))
            t.update()
    scheduler.step(np.mean(loss_avg_arr))
    print('Training epoch: {:02d}, MSE: {:.4f}'.format(epoch, np.mean(loss_avg_arr)))
    return np.mean(loss_avg_arr)

if __name__ == '__main__':
    args = parser.parse_args()

    dataloaders = data_loader.fetch_dataloader(data_dir=osp.join(os.environ['PWD'],args.data),
                                               batch_size=6,
                                               validation_split=.2)
    train_dl = dataloaders['train']
    test_dl = dataloaders['test']

    print(len(train_dl), len(test_dl))
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    model = net.Net(8, 3).to(device) #include puppi
    #model = net.Net(7, 3).to(device) #remove puppi
    optimizer = torch.optim.AdamW(model.parameters(),lr=0.001)
    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, factor=0.5, patience=500, threshold=0.05)
    first_epoch = 0
    best_validation_loss = 10e7
    deltaR = 0.4
    deltaR_dz = 0.3

    loss_fn = net.loss_fn
    metrics = net.metrics

    model_dir = osp.join(os.environ['PWD'],args.ckpts)
    loss_log = open(model_dir+'/loss.log', 'w')
    loss_log.write('# loss log for training starting in '+strftime("%Y-%m-%d %H:%M:%S", gmtime()) + '\n')
    loss_log.write('epoch, loss, val_loss\n')
    loss_log.flush()

    # reload weights from restore_file if specified
    if args.restore_file is not None:
        restore_ckpt = osp.join(model_dir, args.restore_file + '.pth.tar')
        ckpt = utils.load_checkpoint(restore_ckpt, model, optimizer, scheduler)
        first_epoch = ckpt['epoch']
        print('Restarting training from epoch',first_epoch)
        with open(osp.join(model_dir, 'metrics_val_best.json')) as restore_metrics:
            best_validation_loss = json.load(restore_metrics)['loss']

    for epoch in range(first_epoch+1,101):

        print('Current best loss:', best_validation_loss)
        if '_last_lr' in scheduler.state_dict():
            print('Learning rate:', scheduler.state_dict()['_last_lr'][0])

        # compute number of batches in one epoch (one full pass over the training set)
        train_loss = train(model, device, optimizer, scheduler, loss_fn, train_dl, epoch)

        # Save weights
        utils.save_checkpoint({'epoch': epoch,
                               'state_dict': model.state_dict(),
                               'optim_dict': optimizer.state_dict(),
                               'sched_dict': scheduler.state_dict()},
                              is_best=False,
                              checkpoint=model_dir)

        # Evaluate for one epoch on validation set
        test_metrics, resolutions = evaluate(model, device, loss_fn, test_dl, metrics, deltaR,deltaR_dz, model_dir)

        validation_loss = test_metrics['loss']
        loss_log.write('%d,%.2f,%.2f\n'%(epoch,train_loss, validation_loss))
        loss_log.flush()
        is_best = (validation_loss<=best_validation_loss)

        # If best_eval, best_save_path
        if is_best:
            print('Found new best loss!')
            best_validation_loss=validation_loss

            # Save weights
            utils.save_checkpoint({'epoch': epoch,
                                   'state_dict': model.state_dict(),
                                   'optim_dict': optimizer.state_dict(),
                                   'sched_dict': scheduler.state_dict()},
                                  is_best=True,
                                  checkpoint=model_dir)

            # Save best val metrics in a json file in the model directory
            utils.save_dict_to_json(test_metrics, osp.join(model_dir, 'metrics_val_best.json'))
            utils.save(resolutions, osp.join(model_dir, 'best.resolutions'))

        utils.save_dict_to_json(test_metrics, osp.join(model_dir, 'metrics_val_last.json'))
        utils.save(resolutions, osp.join(model_dir, 'last.resolutions'))

    loss_log.close()