從tea出發

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

導入相關參數

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

從data_config.py中導出數據集的相關信息

full_arch_name = args.archif 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)

訓練結果存儲

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),)

導入模型

crop_size = model.crop_sizescale_size = model.scale_sizeinput_mean = model.input_meaninput_std = model.input_stdpolicies = 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()#并行計算

相關參數導出及設置并行計算，其中GPU的數量由實際情況設置

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

設置初始化的學習率

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

設置優化器

if args.resume:#用來設置是否從斷點出繼續訓練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)))

通過預訓練模型訓練
…

總結

以上是生活随笔為你收集整理的解析复杂深度学习项目构建的全部內容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網站內容還不錯，歡迎將生活随笔推薦給好友。