train_cvpr19.py

#  #!/usr/bin/env python
#   Copyright (c) 2019. ShiJie Sun at the Chang'an University
#   This work is licensed under the terms of the Creative Commons Attribution-NonCommercial-ShareAlike 3.0 License.
#   For a copy, see <http://creativecommons.org/licenses/by-nc-sa/3.0/>.
#   Author: shijie Sun
#   Email: shijieSun@chd.edu.cn
#   Github: www.github.com/shijieS
#
#

import os
import torch
import torch.optim as optim
import argparse
from torch.autograd import Variable
import torch.utils.data as data
from dataset import CVPR19TrainDataset
from config import config
from layers.dmmn import DMMN, DMMNLoss
import time
from dataset import collate_fn
from dataset.utils import Transforms
from draw_utils import show_bboxes
import torch.backends.cudnn as cudnn

# torch.multiprocessing.set_start_method('spawn', force=True)


str2bool = lambda v: v.lower() in ("yes", "true", "t", "1")

cfg = config["train"]

parser = argparse.ArgumentParser(description='Single Shot Detector and Tracker Train')
parser.add_argument('--version', default='v1', help='current version')

parser.add_argument('--basenet', default=cfg['base_net_weights'], help='resnet weights file')
parser.add_argument('--batch_size', default=cfg['batch_size'], type=int, help='Batch size for training')
parser.add_argument('--resume', default=cfg['resume'], type=str, help='Resume from checkpoint')
parser.add_argument('--num_workers', default=cfg['num_workers'], type=int, help='Number of workers used in dataloading')
parser.add_argument('--start_epoch', default=cfg['start_epoch'], type=int, help='end of iteration')
parser.add_argument('--end_epoch', default=cfg['end_epoch'], type=int, help='begin of iteration')
parser.add_argument('--lr_decay_per_epoch', default=cfg['lr_decay_per_epoch'], type=list, help='learning rate decay')
parser.add_argument('--cuda', default=config['cuda'], type=str2bool, help='Use cuda to train motion_model')
parser.add_argument('--lr', '--learning-rate', default=cfg['learning_rate'], type=float, help='initial learning rate')
parser.add_argument('--momentum', default=cfg['momentum'], type=float, help='momentum')
parser.add_argument('--weight_decay', default=cfg['weight_decay'], type=float, help='Weight decay for SGD')
parser.add_argument('--gamma', default=cfg['gamma'], type=float, help='Gamma update for SGD')
parser.add_argument('--log_iters', default=cfg['log_iters'], type=bool, help='Print the loss at each iteration')
parser.add_argument('--tensorboard', default=cfg['tensorboard'], type=str2bool, help='Use tensor board x for loss visualization')
parser.add_argument('--port', default=cfg['port'], type=int, help='set vidom port')
parser.add_argument('--send_images', default=cfg['send_images'], type=str2bool, help='send the generated images to tensorboard')
parser.add_argument('--log_save_folder', default=cfg['log_save_folder'], help='Location to save checkpoint models')
parser.add_argument('--weights_save_folder', default=cfg['weights_save_folder'], help='Location to save network weights')
parser.add_argument('--save_weight_per_epoch', default=cfg['save_weight_per_epoch'], help='Every n epoch to save weights')
parser.add_argument('--dataset_path', default=config['dataset_path'], help='ua dataset root folder')
parser.add_argument('--run_mode', default=config["train"]["run_mode"], help="ua run mode, 'debug' mode will save print more message, otherwise, you should select 'run' motion_model")

args = parser.parse_args()

# load dataset
dataset = CVPR19TrainDataset(transform=Transforms())

epoch_size = len(dataset) // args.batch_size
start_iter = args.start_epoch * epoch_size
end_iter = args.end_epoch * epoch_size + 10

save_weights_iteration = int(epoch_size * args.save_weight_per_epoch)

# re-calculate the learning rate
step_values = [i*epoch_size for i in args.lr_decay_per_epoch]

# init tensorboard
if args.tensorboard:
    from tensorboardX import SummaryWriter
    if not os.path.exists(args.log_save_folder):
        os.makedirs(args.log_save_folder)
    writer = SummaryWriter(log_dir=args.log_save_folder)

# cuda configure
if torch.cuda.is_available():
    if args.cuda:
        torch.set_default_tensor_type('torch.cuda.FloatTensor')
    if not args.cuda:
        print("WARNING: It looks like you have a CUDA device, but aren't " +
              "using CUDA.\nRun with --cuda for optimal training speed.")
        torch.set_default_tensor_type('torch.FloatTensor')
else:
    torch.set_default_tensor_type('torch.FloatTensor')

# check saving directory
if not os.path.exists(args.weights_save_folder):
    os.mkdir(args.weights_save_folder)

# creat the network
dmmn = DMMN.build("train")
net = dmmn

if args.cuda:
    net = torch.nn.DataParallel(dmmn)
    cudnn.benchmark = True
    net = net.cuda()

if args.resume:
    print("Resuming training, loading {}...".format(args.resume))
    dmmn.load_weights(args.resume)
elif args.basenet is not None:
    print("Loading base network...")
    dmmn.load_base_weights(args.basenet)

# create optimizer
optimizer = optim.Adam(net.parameters(), args.lr, weight_decay=args.weight_decay)
# optimizer = optim.SGD(net.parameters(), lr=args.lr, momentum=args.momentum, weight_decay=args.weight_decay)

# create loss criterion
criterion = DMMNLoss()


# train function
def train():
    # do some init operation
    if args.run_mode == "debug":
        print_iteration = 10
        save_image_iteration = 10
        add_scalar_iteration = 1
        add_histogram_iteration = 10

    else:
        print_iteration = 100
        add_scalar_iteration = 10
        save_image_iteration = 1000
        add_histogram_iteration = 1000


    # change the network mode
    net.train()
    batch_iterator = None

    data_loader = data.DataLoader(dataset=dataset, batch_size=args.batch_size,
                                  num_workers=args.num_workers,
                                  shuffle=True,
                                  collate_fn=collate_fn,
                                  pin_memory=False)

    step_index = 0
    current_lr = args.lr
    # init learning rate for the resume
    for iteration in range(start_iter):
        if iteration in step_values:
            step_index += 1
            # current_lr = adjust_learning_rate(optimizer, args.gamma, step_index)

    # start training
    batch_iterator = None
    for iteration in range(start_iter, end_iter):
        if (not batch_iterator) or (iteration % epoch_size == 0):
            # create batch iterator
            batch_iterator = iter(data_loader)
            all_epoch_loss = []

        # adjus t learning rate
        if iteration in step_values:
            step_index += 1
            # current_lr = adjust_learning_rate(optimizer, args.gamma, step_index)

        if batch_iterator is None:
            continue
        # reading item
        frames_1, target_1, times_1 = next(batch_iterator)
        if frames_1 is None:
            continue

        # print(iteration)
        # continue
        if frames_1 is None or target_1 is None or times_1 is None:
            continue

        if args.cuda:
            frames_1 = Variable(frames_1.cuda())
            with torch.no_grad():
                target_1 = [
                    [Variable(target[j].cuda()) for j in range(len(target))]
                for target in target_1]
                times_1 = Variable(times_1.cuda())
        else:
            pass

        # forward
        t0 = time.time()
        param, p_c, p_e = net(frames_1)

        # loss
        optimizer.zero_grad()
        loss_l, loss_c, loss_e = criterion(
            (param, p_c, p_e, dmmn.priors),
            target_1,
            times_1)

        loss = loss_l + loss_c + loss_e

        if torch.isnan(loss):
            print("nan loss ignored")
            continue

        loss.backward()
        optimizer.step()
        all_epoch_loss += [loss.data.cpu()]

        t1 = time.time()

        # console logs
        if iteration % print_iteration == 0:
            print('Timer: %.4f sec.' % (t1 - t0))
            print('iter ' + str(iteration) + ', ' + str(epoch_size) + ' || epoch: %.4f ' % (iteration / (float)(epoch_size)) + ' || Loss: %.4f ||' % all_epoch_loss[-1], end=' ')

        # tensorboard logs
        if args.tensorboard and iteration % add_scalar_iteration == 0:
            writer.add_scalar('data/learning_rate', current_lr, iteration)
            writer.add_scalar('loss/loss', loss.data.cpu(), iteration)
            writer.add_scalar('loss-location', loss_l.data.cpu(), iteration)
            writer.add_scalar('loss-classification', loss_c.data.cpu(), iteration)
            writer.add_scalar('loss-exists', loss_e.data.cpu(), iteration)


        # if args.tensorboard and iteration % add_histogram_iteration == 0:
        #         # add weights
        #         for name, param in net.named_parameters():
        #             writer.add_histogram(name, param.cpu().data.numpy(), iteration, bins='fd')

        # save the result image
        show_bboxes(frames_1, target_1, is_save=True, iteration=iteration)

        # weights save
        if iteration % save_weights_iteration == 0:
            print('Saving weights, iter: {}'.format(iteration))
            torch.save(dmmn.state_dict(),
                       os.path.join(args.weights_save_folder,
                                    'dmmn' + repr(iteration) + '.pth'))
    torch.save(dmmn.state_dict(), args.weights_save_folder + '' + args.version + '.pth')


def adjust_learning_rate(optimizer, gamma, step):
    """Sets the learning rate to the initial LR decayed by 10 at every specified step
    # Adapted from PyTorch Imagenet example:
    # https://github.com/pytorch/examples/blob/master/imagenet/main.py
    """
    lr = args.lr * (gamma ** (step))
    for param_group in optimizer.param_groups:
        param_group['lr'] = lr
    return lr


if __name__ == '__main__':
    train()