解析复杂深度学习项目构建

从tea出发

    start_time = datetime.datetime.now()
    args = parser.parse_args()

导入相关参数

 num_class, args.train_list, args.val_list, prefix = dataset_config.return_dataset(args.dataset,
                                                                                      args.modality)

从data_config.py中导出数据集的相关信息

full_arch_name = args.arch
    if args.shift:
        full_arch_name += '_shift{}_{}'.format(args.shift_div, args.shift_place)
    args.store_name = '_'.join(
        [args.experiment_name, args.dataset, args.modality, full_arch_name, args.consensus_type, 'segment%d' % args.num_segments,
         'e{}'.format(args.epochs)])
    if args.pretrain != 'imagenet':
        args.store_name += '_{}'.format(args.pretrain)
    if args.lr_type != 'step':
        args.store_name += '_{}'.format(args.lr_type)
    if args.dense_sample:
        args.store_name += '_dense'
    if args.suffix is not None:
        args.store_name += '_{}'.format(args.suffix)
    print('storing name: ' + args.store_name)

训练结果存储

model = TSN(num_class, args.num_segments, args.modality,
                base_model=args.arch,
                consensus_type=args.consensus_type,
                dropout=args.dropout,
                img_feature_dim=args.img_feature_dim,
                partial_bn=not args.no_partialbn,
                pretrain=args.pretrain,
                is_shift=args.shift, shift_div=args.shift_div,shift_place=args.shift_place,
                fc_lr5=not (args.tune_from and args.dataset in args.tune_from),)

导入模型

    crop_size = model.crop_size
    scale_size = model.scale_size
    input_mean = model.input_mean
    input_std = model.input_std
    policies = model.get_optim_policies()
    train_augmentation = model.get_augmentation(flip=False if 'something' in args.dataset else True)
    import pdb; pdb.set_trace()
    with torch.no_grad():
         model = torch.nn.DataParallel(model, device_ids=[0, 1]).cuda()#并行计算

相关参数导出及设置并行计算，其中GPU的数量由实际情况设置

 # Add specific initialized lr and weight_decay for each group
    for param_group in policies:
        param_group['lr'] = args.lr * param_group['lr_mult']
        param_group['weight_decay'] = args.weight_decay * param_group['decay_mult']

设置初始化的学习率

 optimizer = torch.optim.SGD(policies,momentum=args.momentum)

设置优化器

    if args.resume:#用来设置是否从断点出继续训练
        if os.path.isfile(args.resume):
            print(("=> loading checkpoint '{}'".format(args.resume)))
            checkpoint = torch.load(args.resume)
            args.start_epoch = checkpoint['epoch']
            best_prec1 = checkpoint['best_prec1']
            model.load_state_dict(checkpoint['state_dict'])
            optimizer.load_state_dict(checkpoint['optimizer'])
            print(("=> loaded checkpoint '{}' (epoch {})"
                   .format(args.evaluate, checkpoint['epoch'])))
        else:
            print(("=> no checkpoint found at '{}'".format(args.resume)))

通过预训练模型训练
…