def main(): """ Main Function """ # Set up the Arguments, Tensorboard Writer, Dataloader, Loss Fn, Optimizer assert_and_infer_cfg(args) writer = prep_experiment(args, parser) train_loader, val_loader, train_obj = datasets.setup_loaders(args) criterion, criterion_val = loss.get_loss(args) net = network.get_net(args, criterion) optim, scheduler = optimizer.get_optimizer(args, net) if args.fix_bn: net.apply(set_bn_eval) print("Fix bn for finetuning") if args.fp16: net, optim = amp.initialize(net, optim, opt_level="O1") net = network.wrap_network_in_dataparallel(net, args.apex) if args.snapshot: optimizer.load_weights(net, optim, args.snapshot, args.restore_optimizer) if args.evaluateF: assert args.snapshot is not None, "must load weights for evaluation" evaluate(val_loader, net, args) return torch.cuda.empty_cache() # Main Loop for epoch in range(args.start_epoch, args.max_epoch): # Update EPOCH CTR cfg.immutable(False) cfg.EPOCH = epoch cfg.immutable(True) scheduler.step() train(train_loader, net, optim, epoch, writer) if args.apex: train_loader.sampler.set_epoch(epoch + 1) validate(val_loader, net, criterion_val, optim, epoch, writer) if args.class_uniform_pct: if epoch >= args.max_cu_epoch: train_obj.build_epoch(cut=True) if args.apex: train_loader.sampler.set_num_samples() else: train_obj.build_epoch()
def main(): """ Main Function """ if AutoResume: AutoResume.init() assert args.result_dir is not None, 'need to define result_dir arg' logx.initialize(logdir=args.result_dir, tensorboard=True, hparams=vars(args), global_rank=args.global_rank) # Set up the Arguments, Tensorboard Writer, Dataloader, Loss Fn, Optimizer assert_and_infer_cfg(args) prep_experiment(args) train_loader, val_loader, train_obj = \ datasets.setup_loaders(args) criterion, criterion_val = get_loss(args) auto_resume_details = None if AutoResume: auto_resume_details = AutoResume.get_resume_details() if auto_resume_details: checkpoint_fn = auto_resume_details.get("RESUME_FILE", None) checkpoint = torch.load(checkpoint_fn, map_location=torch.device('cpu')) args.result_dir = auto_resume_details.get("TENSORBOARD_DIR", None) args.start_epoch = int(auto_resume_details.get("EPOCH", None)) + 1 args.restore_net = True args.restore_optimizer = True msg = ("Found details of a requested auto-resume: checkpoint={}" " tensorboard={} at epoch {}") logx.msg(msg.format(checkpoint_fn, args.result_dir, args.start_epoch)) elif args.resume: checkpoint = torch.load(args.resume, map_location=torch.device('cpu')) args.arch = checkpoint['arch'] args.start_epoch = int(checkpoint['epoch']) + 1 args.restore_net = True args.restore_optimizer = True msg = "Resuming from: checkpoint={}, epoch {}, arch {}" logx.msg(msg.format(args.resume, args.start_epoch, args.arch)) elif args.snapshot: if 'ASSETS_PATH' in args.snapshot: args.snapshot = args.snapshot.replace('ASSETS_PATH', cfg.ASSETS_PATH) checkpoint = torch.load(args.snapshot, map_location=torch.device('cpu')) args.restore_net = True msg = "Loading weights from: checkpoint={}".format(args.snapshot) logx.msg(msg) #define the NASA optimizer parameter iter_tot = len(train_loader) * args.max_epoch # tau = args.tau_factor/sqrt(iter_tot) tau = 1 net = network.get_net(args, criterion) k = 1 # optim, scheduler = get_optimizer(args, net) optim, scheduler = get_optimizer(args, net, tau, k) # Visualize feature maps #activation = {} #def get_activation(name): #def hook(model, input, output): #activation[name] = output.detach() #return hook #net.layer[0].register_forward_hook(get_activation('conv1')) #data, _ = dataset[0] #data.unsqueeze_(0) #output = model(data) #act = activation['conv1'].squeeze() #fig, axarr = plt.subplots(act.size(0)) #for idx in range(act.size(0)): #axarr[idx].imshow(act[idx]) if args.fp16: net, optim = amp.initialize(net, optim, opt_level=args.amp_opt_level) net = network.wrap_network_in_dataparallel(net, args.apex) if args.summary: from thop import profile img = torch.randn(1, 3, 640, 640).cuda() mask = torch.randn(1, 1, 640, 640).cuda() macs, params = profile(net, inputs={'images': img, 'gts': mask}) print(f'macs {macs} params {params}') sys.exit() if args.restore_optimizer: restore_opt(optim, checkpoint) if args.restore_net: restore_net(net, checkpoint) if args.init_decoder: net.module.init_mods() torch.cuda.empty_cache() if args.start_epoch != 0: scheduler.step(args.start_epoch) # There are 4 options for evaluation: # --eval val just run validation # --eval val --dump_assets dump all images and assets # --eval folder just dump all basic images # --eval folder --dump_assets dump all images and assets if args.eval == 'test': validate(val_loader, net, criterion=None, optim=None, epoch=0, calc_metrics=False, dump_assets=args.dump_assets, dump_all_images=True, testing=True, grid=city) return 0 if args.eval == 'val': if args.dump_topn: validate_topn(val_loader, net, criterion_val, optim, 0, args) else: validate(val_loader, net, criterion=criterion_val, optim=optim, epoch=0, dump_assets=args.dump_assets, dump_all_images=args.dump_all_images, calc_metrics=not args.no_metrics) return 0 elif args.eval == 'folder': # Using a folder for evaluation means to not calculate metrics validate(val_loader, net, criterion=criterion_val, optim=optim, epoch=0, calc_metrics=False, dump_assets=args.dump_assets, dump_all_images=True) return 0 elif args.eval is not None: raise 'unknown eval option {}'.format(args.eval) for epoch in range(args.start_epoch, args.max_epoch): update_epoch(epoch) if args.only_coarse: train_obj.only_coarse() train_obj.build_epoch() if args.apex: train_loader.sampler.set_num_samples() elif args.class_uniform_pct: if epoch >= args.max_cu_epoch: train_obj.disable_coarse() train_obj.build_epoch() if args.apex: train_loader.sampler.set_num_samples() else: train_obj.build_epoch() else: pass train(train_loader, net, optim, epoch) if args.apex: train_loader.sampler.set_epoch(epoch + 1) if epoch % args.val_freq == 0: validate(val_loader, net, criterion_val, optim, epoch) scheduler.step() if check_termination(epoch): return 0
def main(): """ Main Function """ if AutoResume: AutoResume.init() assert args.result_dir is not None, 'need to define result_dir arg' logx.initialize(logdir=args.result_dir, tensorboard=True, hparams=vars(args), global_rank=args.global_rank) # Set up the Arguments, Tensorboard Writer, Dataloader, Loss Fn, Optimizer assert_and_infer_cfg(args) prep_experiment(args) train_loader, val_loader, train_obj = \ datasets.setup_loaders(args) criterion, criterion_val = get_loss(args) auto_resume_details = None if AutoResume: auto_resume_details = AutoResume.get_resume_details() if auto_resume_details: checkpoint_fn = auto_resume_details.get("RESUME_FILE", None) checkpoint = torch.load(checkpoint_fn, map_location=torch.device('cpu')) args.result_dir = auto_resume_details.get("TENSORBOARD_DIR", None) args.start_epoch = int(auto_resume_details.get("EPOCH", None)) + 1 args.restore_net = True args.restore_optimizer = True msg = ("Found details of a requested auto-resume: checkpoint={}" " tensorboard={} at epoch {}") logx.msg(msg.format(checkpoint_fn, args.result_dir, args.start_epoch)) elif args.resume: checkpoint = torch.load(args.resume, map_location=torch.device('cpu')) args.arch = checkpoint['arch'] args.start_epoch = int(checkpoint['epoch']) + 1 args.restore_net = True args.restore_optimizer = True msg = "Resuming from: checkpoint={}, epoch {}, arch {}" logx.msg(msg.format(args.resume, args.start_epoch, args.arch)) elif args.snapshot: if 'ASSETS_PATH' in args.snapshot: args.snapshot = args.snapshot.replace('ASSETS_PATH', cfg.ASSETS_PATH) checkpoint = torch.load(args.snapshot, map_location=torch.device('cpu')) args.restore_net = True msg = "Loading weights from: checkpoint={}".format(args.snapshot) logx.msg(msg) net = network.get_net(args, criterion) optim, scheduler = get_optimizer(args, net) if args.fp16: net, optim = amp.initialize(net, optim, opt_level=args.amp_opt_level) net = network.wrap_network_in_dataparallel(net, args.apex) if args.summary: print(str(net)) from pytorchOpCounter.thop import profile img = torch.randn(1, 3, 1024, 2048).cuda() mask = torch.randn(1, 1, 1024, 2048).cuda() macs, params = profile(net, inputs={'images': img, 'gts': mask}) print(f'macs {macs} params {params}') sys.exit() if args.restore_optimizer: restore_opt(optim, checkpoint) if args.restore_net: restore_net(net, checkpoint) if args.init_decoder: net.module.init_mods() torch.cuda.empty_cache() if args.start_epoch != 0: scheduler.step(args.start_epoch) # There are 4 options for evaluation: # --eval val just run validation # --eval val --dump_assets dump all images and assets # --eval folder just dump all basic images # --eval folder --dump_assets dump all images and assets if args.eval == 'val': if args.dump_topn: validate_topn(val_loader, net, criterion_val, optim, 0, args) else: validate(val_loader, net, criterion=criterion_val, optim=optim, epoch=0, dump_assets=args.dump_assets, dump_all_images=args.dump_all_images, calc_metrics=not args.no_metrics) return 0 elif args.eval == 'folder': # Using a folder for evaluation means to not calculate metrics validate(val_loader, net, criterion=None, optim=None, epoch=0, calc_metrics=False, dump_assets=args.dump_assets, dump_all_images=True) return 0 elif args.eval is not None: raise 'unknown eval option {}'.format(args.eval) for epoch in range(args.start_epoch, args.max_epoch): update_epoch(epoch) if args.only_coarse: train_obj.only_coarse() train_obj.build_epoch() if args.apex: train_loader.sampler.set_num_samples() elif args.class_uniform_pct: if epoch >= args.max_cu_epoch: train_obj.disable_coarse() train_obj.build_epoch() if args.apex: train_loader.sampler.set_num_samples() else: train_obj.build_epoch() else: pass train(train_loader, net, optim, epoch) if args.apex: train_loader.sampler.set_epoch(epoch + 1) if epoch % args.val_freq == 0: validate(val_loader, net, criterion_val, optim, epoch) scheduler.step() if check_termination(epoch): return 0
def main(): """ Main Function """ if AutoResume: AutoResume.init() assert args.result_dir is not None, 'need to define result_dir arg' logx.initialize(logdir=args.result_dir, tensorboard=False, hparams=vars(args), global_rank=args.global_rank) # Set up the Arguments, Tensorboard Writer, Dataloader, Loss Fn, Optimizer assert_and_infer_cfg(args) prep_experiment(args) train_loader, val_loader, train_obj = datasets.setup_loaders(args) criterion, criterion_val = get_loss(args) auto_resume_details = None if AutoResume: auto_resume_details = AutoResume.get_resume_details() if auto_resume_details: checkpoint_fn = auto_resume_details.get("RESUME_FILE", None) checkpoint = torch.load(checkpoint_fn, map_location=torch.device('cpu')) args.result_dir = auto_resume_details.get("TENSORBOARD_DIR", None) args.start_epoch = int(auto_resume_details.get("EPOCH", None)) + 1 args.restore_net = True args.restore_optimizer = True msg = ("Found details of a requested auto-resume: checkpoint={}" " tensorboard={} at epoch {}") logx.msg(msg.format(checkpoint_fn, args.result_dir, args.start_epoch)) elif args.resume: checkpoint = torch.load(args.resume, map_location=torch.device('cpu')) args.arch = checkpoint['arch'] args.start_epoch = int(checkpoint['epoch']) + 1 args.restore_net = True args.restore_optimizer = True msg = "Resuming from: checkpoint={}, epoch {}, arch {}" logx.msg(msg.format(args.resume, args.start_epoch, args.arch)) elif args.snapshot: if 'ASSETS_PATH' in args.snapshot: args.snapshot = args.snapshot.replace('ASSETS_PATH', cfg.ASSETS_PATH) checkpoint = torch.load(args.snapshot, map_location=torch.device('cpu')) args.restore_net = True msg = "Loading weights from: checkpoint={}".format(args.snapshot) logx.msg(msg) net = network.get_net(args, criterion) optim, scheduler = get_optimizer(args, net) net = network.wrap_network_in_dataparallel(net, args.apex) if args.restore_optimizer: restore_opt(optim, checkpoint) if args.restore_net: restore_net(net, checkpoint) if args.init_decoder: net.module.init_mods() torch.cuda.empty_cache() if args.start_epoch != 0: scheduler.step(args.start_epoch) if args.eval == 'folder': # Using a folder for evaluation means to not calculate metrics # validate(val_loader, net, criterion=None, optim=None, epoch=0, # calc_metrics=False, dump_assets=args.dump_assets, # dump_all_images=True) if not os.path.exists(args.result_dir + 'image_2/'): os.mkdir(args.result_dir + 'image_2/') if not os.path.exists(args.result_dir + 'image_3/'): os.mkdir(args.result_dir + 'image_3/') num_image = 7481 for idx in tqdm(range(num_image)): sample_idx = "%06d" % idx eval_minibatch(sample_idx, "image_2/", net, args) eval_minibatch(sample_idx, "image_3/", net, args) return 0 elif args.eval is not None: raise 'unknown eval option {}'.format(args.eval)
def main(): """ Main Function """ # Set up the Arguments, Tensorboard Writer, Dataloader, Loss Fn, Optimizer # args2 = copy.deepcopy(args) assert_and_infer_cfg(args) # assert_and_infer_cfg(args2) # args2.dataset = 'kitti_trav' # print(args.dataset) # print(args2.dataset) writer = prep_experiment(args, parser) # writer = prep_experiment(args2, parser) # Dataset train_loader, val_loader, train_obj = datasets.setup_loaders(args) # train_loader2, val_loader2, train_obj2 = datasets.setup_loaders(args2) criterion, criterion_val = loss.get_loss(args, data_type='semantic') criterion2, criterion_val2 = loss.get_loss(args, data_type='trav') net = network.get_net(args, criterion, criterion2) #parameters list # param1_lists = list(net.mod1.parameters()) + list(net.mod2.parameters()) + list(net.mod3.parameters()) + list(net.mod4.parameters()) + list(net.mod5.parameters()) + list(net.mod6.parameters()) + list(net.mod7.parameters()) + list(net.pool2.parameters()) + list(net.pool3.parameters()) + list(net.aspp.parameters()) + list(net.bot_fine.parameters()) + list(net.bot_aspp.parameters()) + list(net.final.parameters()) + [log_sigma_A] # param2_lists = list(net.mod1.parameters()) + list(net.mod2.parameters()) + list(net.mod3.parameters()) + list(net.mod4.parameters()) + list(net.mod5.parameters()) + list(net.mod6.parameters()) + list(net.mod7.parameters()) + list(net.pool2.parameters()) + list(net.pool3.parameters()) + list(net.aspp.parameters()) + list(net.bot_fine.parameters()) + list(net.bot_aspp.parameters()) + list(net.final2.parameters()) + [log_sigma_B] #optimizers optim, scheduler = optimizer.get_optimizer(args, net) # optim2, scheduler2 = optimizer.get_optimizer(args, param2_lists) if args.fp16: net, optim = amp.initialize(net, optim, opt_level="O1") net = network.wrap_network_in_dataparallel(net, args.apex) if args.snapshot: optimizer.load_weights(net, optim, args.snapshot, args.snapshot2, args.restore_optimizer) # optimizer.load_weights(net, optim2, # args.snapshot, args.snapshot2, args.restore_optimizer) torch.cuda.empty_cache() # Main Loop for epoch in range(args.start_epoch, args.max_epoch): # Update EPOCH CTR cfg.immutable(False) cfg.EPOCH = epoch cfg.immutable(True) scheduler.step() train(train_loader, net, optim, epoch, writer) if args.apex: train_loader.sampler.set_epoch(epoch + 1) # train_loader2.sampler.set_epoch(epoch + 1) validate(val_loader, net, criterion_val, criterion_val2, optim, epoch, writer) if args.class_uniform_pct: if epoch >= args.max_cu_epoch: train_obj.build_epoch(cut=True) # train_obj2.build_epoch(cut=True) if args.apex: train_loader.sampler.set_num_samples() # train_loader2.sampler.set_num_samples() else: train_obj.build_epoch()