def train_pose(args):
    torch.set_num_threads(1)

    if args.resume_run_id:
        resume_dir = EXP_DIR / args.resume_run_id
        resume_args = yaml.load((resume_dir / 'config.yaml').read_text())
        keep_fields = set([
            'resume_run_id',
            'epoch_size',
        ])
        vars(args).update({
            k: v
            for k, v in vars(resume_args).items() if k not in keep_fields
        })

    args.train_refiner = args.TCO_input_generator == 'gt+noise'
    args.train_coarse = not args.train_refiner
    args.save_dir = EXP_DIR / args.run_id

    logger.info(f"{'-'*80}")
    for k, v in args.__dict__.items():
        logger.info(f"{k}: {v}")
    logger.info(f"{'-'*80}")

    # Initialize distributed
    device = torch.cuda.current_device()
    init_distributed_mode()
    world_size = get_world_size()
    args.n_gpus = world_size
    args.global_batch_size = world_size * args.batch_size
    logger.info(f'Connection established with {world_size} gpus.')

    # Make train/val datasets
    def make_datasets(dataset_names):
        datasets = []
        for (ds_name, n_repeat) in dataset_names:
            assert 'test' not in ds_name
            ds = make_scene_dataset(ds_name)
            logger.info(f'Loaded {ds_name} with {len(ds)} images.')
            for _ in range(n_repeat):
                datasets.append(ds)
        return ConcatDataset(datasets)

    # tracking dataset
    scene_ds_train = make_datasets(args.train_ds_names)
    scene_ds_val = make_datasets(args.val_ds_names)

    ds_kwargs = dict(
        resize=args.input_resize,
        rgb_augmentation=args.rgb_augmentation,
        background_augmentation=args.background_augmentation,
        min_area=args.min_area,
        gray_augmentation=args.gray_augmentation,
    )
    ds_train = PoseTrackingDataset(scene_ds_train, **ds_kwargs)
    ds_val = PoseTrackingDataset(scene_ds_val, **ds_kwargs)

    train_sampler = PartialSampler(ds_train, epoch_size=args.epoch_size)
    ds_iter_train = DataLoader(ds_train,
                               sampler=train_sampler,
                               batch_size=args.batch_size,
                               num_workers=args.n_dataloader_workers,
                               collate_fn=ds_train.collate_fn,
                               drop_last=False,
                               pin_memory=True)
    ds_iter_train = MultiEpochDataLoader(ds_iter_train)

    val_sampler = PartialSampler(ds_val, epoch_size=int(0.1 * args.epoch_size))
    ds_iter_val = DataLoader(ds_val,
                             sampler=val_sampler,
                             batch_size=args.batch_size,
                             num_workers=args.n_dataloader_workers,
                             collate_fn=ds_val.collate_fn,
                             drop_last=False,
                             pin_memory=True)
    ds_iter_val = MultiEpochDataLoader(ds_iter_val)

    # Make model
    # renderer = BulletBatchRenderer(object_set=args.urdf_ds_name, n_workers=args.n_rendering_workers)
    object_ds = make_object_dataset(args.object_ds_name)
    mesh_db = MeshDataBase.from_object_ds(object_ds).batched(
        n_sym=args.n_symmetries_batch).cuda().float()

    model = create_model_pose_custom(cfg=args, mesh_db=mesh_db).cuda()

    eval_bundle = make_eval_bundle(args, model)

    if args.resume_run_id:
        resume_dir = EXP_DIR / args.resume_run_id
        path = resume_dir / 'checkpoint.pth.tar'
        logger.info(f'Loading checkpoing from {path}')
        save = torch.load(path)
        state_dict = save['state_dict']
        model.load_state_dict(state_dict)
        start_epoch = save['epoch'] + 1
    else:
        start_epoch = 0
    end_epoch = args.n_epochs

    if args.run_id_pretrain is not None:
        pretrain_path = EXP_DIR / args.run_id_pretrain / 'checkpoint.pth.tar'
        logger.info(f'Using pretrained model from {pretrain_path}.')
        model.load_state_dict(torch.load(pretrain_path)['state_dict'])

    # Synchronize models across processes.
    model = sync_model(model)
    model = torch.nn.parallel.DistributedDataParallel(model,
                                                      device_ids=[device],
                                                      output_device=device)

    # Optimizer
    optimizer = torch.optim.Adam(model.parameters(),
                                 lr=args.lr,
                                 weight_decay=args.weight_decay)

    # Warmup
    if args.n_epochs_warmup == 0:
        lambd = lambda epoch: 1
    else:
        n_batches_warmup = args.n_epochs_warmup * (args.epoch_size //
                                                   args.batch_size)
        lambd = lambda batch: (batch + 1) / n_batches_warmup
    lr_scheduler_warmup = torch.optim.lr_scheduler.LambdaLR(optimizer, lambd)
    lr_scheduler_warmup.last_epoch = start_epoch * args.epoch_size // args.batch_size

    # LR schedulers
    # Divide LR by 10 every args.lr_epoch_decay
    lr_scheduler = torch.optim.lr_scheduler.StepLR(
        optimizer,
        step_size=args.lr_epoch_decay,
        gamma=0.1,
    )
    lr_scheduler.last_epoch = start_epoch - 1
    lr_scheduler.step()

    for epoch in range(start_epoch, end_epoch):
        meters_train = defaultdict(lambda: AverageValueMeter())
        meters_val = defaultdict(lambda: AverageValueMeter())
        meters_time = defaultdict(lambda: AverageValueMeter())

        h = functools.partial(h_pose_custom,
                              model=model,
                              cfg=args,
                              n_iterations=args.n_iterations,
                              mesh_db=mesh_db,
                              input_generator=args.TCO_input_generator)

        def train_epoch():
            model.train()
            iterator = tqdm(ds_iter_train, ncols=80)
            t = time.time()
            for n, sample in enumerate(iterator):
                if n > 0:
                    meters_time['data'].add(time.time() - t)

                optimizer.zero_grad()

                t = time.time()
                loss = h(data=sample, meters=meters_train)
                meters_time['forward'].add(time.time() - t)
                iterator.set_postfix(loss=loss.item())
                meters_train['loss_total'].add(loss.item())

                t = time.time()
                loss.backward()
                total_grad_norm = torch.nn.utils.clip_grad_norm_(
                    model.parameters(),
                    max_norm=args.clip_grad_norm,
                    norm_type=2)
                meters_train['grad_norm'].add(
                    torch.as_tensor(total_grad_norm).item())

                optimizer.step()
                meters_time['backward'].add(time.time() - t)
                meters_time['memory'].add(torch.cuda.max_memory_allocated() /
                                          1024.**2)

                if epoch < args.n_epochs_warmup:
                    lr_scheduler_warmup.step()
                t = time.time()
            if epoch >= args.n_epochs_warmup:
                lr_scheduler.step()

        @torch.no_grad()
        def validation():
            model.eval()
            for sample in tqdm(ds_iter_val, ncols=80):
                loss = h(data=sample, meters=meters_val)
                meters_val['loss_total'].add(loss.item())

        @torch.no_grad()
        def test():
            model.eval()
            return run_eval(eval_bundle, epoch=epoch)

        train_epoch()
        if epoch % args.val_epoch_interval == 0:
            validation()

        test_dict = None
        if epoch % args.test_epoch_interval == 0:
            test_dict = test()

        log_dict = dict()
        log_dict.update({
            'grad_norm':
            meters_train['grad_norm'].mean,
            'grad_norm_std':
            meters_train['grad_norm'].std,
            'learning_rate':
            optimizer.param_groups[0]['lr'],
            'time_forward':
            meters_time['forward'].mean,
            'time_backward':
            meters_time['backward'].mean,
            'time_data':
            meters_time['data'].mean,
            'gpu_memory':
            meters_time['memory'].mean,
            'time':
            time.time(),
            'n_iterations': (epoch + 1) * len(ds_iter_train),
            'n_datas':
            (epoch + 1) * args.global_batch_size * len(ds_iter_train),
        })

        for string, meters in zip(('train', 'val'),
                                  (meters_train, meters_val)):
            for k in dict(meters).keys():
                log_dict[f'{string}_{k}'] = meters[k].mean

        log_dict = reduce_dict(log_dict)
        if get_rank() == 0:
            log(config=args,
                model=model,
                epoch=epoch,
                log_dict=log_dict,
                test_dict=test_dict)
        dist.barrier()
Beispiel #2
0
def train_detector(args):
    torch.set_num_threads(1)

    if args.resume_run_id:
        resume_dir = EXP_DIR / args.resume_run_id
        resume_args = yaml.load((resume_dir / 'config.yaml').read_text())
        keep_fields = set([
            'resume_run_id',
            'epoch_size',
        ])
        vars(args).update({
            k: v
            for k, v in vars(resume_args).items() if k not in keep_fields
        })

    args = check_update_config(args)
    args.save_dir = EXP_DIR / args.run_id

    logger.info(f"{'-'*80}")
    for k, v in args.__dict__.items():
        logger.info(f"{k}: {v}")
    logger.info(f"{'-'*80}")

    # Initialize distributed
    device = torch.cuda.current_device()
    init_distributed_mode()
    world_size = get_world_size()
    args.n_gpus = world_size
    args.global_batch_size = world_size * args.batch_size
    logger.info(f'Connection established with {world_size} gpus.')

    # Make train/val datasets
    def make_datasets(dataset_names):
        datasets = []
        all_labels = set()
        for (ds_name, n_repeat) in dataset_names:
            assert 'test' not in ds_name
            ds = make_scene_dataset(ds_name)
            logger.info(f'Loaded {ds_name} with {len(ds)} images.')
            all_labels = all_labels.union(set(ds.all_labels))
            for _ in range(n_repeat):
                datasets.append(ds)
        return ConcatDataset(datasets), all_labels

    scene_ds_train, train_labels = make_datasets(args.train_ds_names)
    scene_ds_val, _ = make_datasets(args.val_ds_names)
    label_to_category_id = dict()
    label_to_category_id['background'] = 0
    for n, label in enumerate(sorted(list(train_labels)), 1):
        label_to_category_id[label] = n
    logger.info(
        f'Training with {len(label_to_category_id)} categories: {label_to_category_id}'
    )
    args.label_to_category_id = label_to_category_id

    ds_kwargs = dict(
        resize=args.input_resize,
        rgb_augmentation=args.rgb_augmentation,
        background_augmentation=args.background_augmentation,
        gray_augmentation=args.gray_augmentation,
        label_to_category_id=label_to_category_id,
    )
    ds_train = DetectionDataset(scene_ds_train, **ds_kwargs)
    ds_val = DetectionDataset(scene_ds_val, **ds_kwargs)

    train_sampler = PartialSampler(ds_train, epoch_size=args.epoch_size)
    ds_iter_train = DataLoader(ds_train,
                               sampler=train_sampler,
                               batch_size=args.batch_size,
                               num_workers=args.n_dataloader_workers,
                               collate_fn=collate_fn,
                               drop_last=False,
                               pin_memory=True)
    ds_iter_train = MultiEpochDataLoader(ds_iter_train)

    val_sampler = PartialSampler(ds_val, epoch_size=int(0.1 * args.epoch_size))
    ds_iter_val = DataLoader(ds_val,
                             sampler=val_sampler,
                             batch_size=args.batch_size,
                             num_workers=args.n_dataloader_workers,
                             collate_fn=collate_fn,
                             drop_last=False,
                             pin_memory=True)
    ds_iter_val = MultiEpochDataLoader(ds_iter_val)

    model = create_model_detector(cfg=args,
                                  n_classes=len(
                                      args.label_to_category_id)).cuda()

    if args.resume_run_id:
        resume_dir = EXP_DIR / args.resume_run_id
        path = resume_dir / 'checkpoint.pth.tar'
        logger.info(f'Loading checkpoing from {path}')
        save = torch.load(path)
        state_dict = save['state_dict']
        model.load_state_dict(state_dict)
        start_epoch = save['epoch'] + 1
    else:
        start_epoch = 0
    end_epoch = args.n_epochs

    if args.run_id_pretrain is not None:
        pretrain_path = EXP_DIR / args.run_id_pretrain / 'checkpoint.pth.tar'
        logger.info(f'Using pretrained model from {pretrain_path}.')
        model.load_state_dict(torch.load(pretrain_path)['state_dict'])
    elif args.pretrain_coco:
        state_dict = load_state_dict_from_url(
            model_urls['maskrcnn_resnet50_fpn_coco'])
        keep = lambda k: 'box_predictor' not in k and 'mask_predictor' not in k
        state_dict = {k: v for k, v in state_dict.items() if keep(k)}
        model.load_state_dict(state_dict, strict=False)
        logger.info(
            'Using model pre-trained on coco. Removed predictor heads.')
    else:
        logger.info('Training MaskRCNN from scratch.')

    # Synchronize models across processes.
    model = sync_model(model)
    model = torch.nn.parallel.DistributedDataParallel(model,
                                                      device_ids=[device],
                                                      output_device=device)

    # Optimizer
    params = [p for p in model.parameters() if p.requires_grad]
    if args.optimizer.lower() == 'sgd':
        optimizer = torch.optim.SGD(params,
                                    lr=args.lr,
                                    weight_decay=args.weight_decay,
                                    momentum=args.momentum)
    elif args.optimizer.lower() == 'adam':
        optimizer = torch.optim.Adam(params,
                                     lr=args.lr,
                                     weight_decay=args.weight_decay)
    else:
        raise ValueError(f'Unknown optimizer {args.optimizer}')

    # Warmup
    if args.n_epochs_warmup == 0:
        lambd = lambda epoch: 1
    else:
        n_batches_warmup = args.n_epochs_warmup * (args.epoch_size //
                                                   args.batch_size)
        lambd = lambda batch: (batch + 1) / n_batches_warmup
    lr_scheduler_warmup = torch.optim.lr_scheduler.LambdaLR(optimizer, lambd)
    lr_scheduler_warmup.last_epoch = start_epoch * args.epoch_size // args.batch_size

    # LR schedulers
    # Divide LR by 10 every args.lr_epoch_decay
    lr_scheduler = torch.optim.lr_scheduler.StepLR(
        optimizer,
        step_size=args.lr_epoch_decay,
        gamma=0.1,
    )
    lr_scheduler.last_epoch = start_epoch - 1
    lr_scheduler.step()

    for epoch in range(start_epoch, end_epoch):
        meters_train = defaultdict(AverageValueMeter)
        meters_val = defaultdict(AverageValueMeter)
        meters_time = defaultdict(AverageValueMeter)

        h = functools.partial(h_maskrcnn, model=model, cfg=args)

        def train_epoch():
            model.train()
            iterator = tqdm(ds_iter_train, ncols=80)
            t = time.time()
            for n, sample in enumerate(iterator):
                if n > 0:
                    meters_time['data'].add(time.time() - t)

                optimizer.zero_grad()

                t = time.time()
                loss = h(data=sample, meters=meters_train)
                meters_time['forward'].add(time.time() - t)
                iterator.set_postfix(loss=loss.item())
                meters_train['loss_total'].add(loss.item())

                t = time.time()
                loss.backward()
                total_grad_norm = torch.nn.utils.clip_grad_norm_(
                    model.parameters(), max_norm=np.inf, norm_type=2)
                meters_train['grad_norm'].add(
                    torch.as_tensor(total_grad_norm).item())

                optimizer.step()
                meters_time['backward'].add(time.time() - t)
                meters_time['memory'].add(torch.cuda.max_memory_allocated() /
                                          1024.**2)

                if epoch < args.n_epochs_warmup:
                    lr_scheduler_warmup.step()
                t = time.time()
            if epoch >= args.n_epochs_warmup:
                lr_scheduler.step()

        @torch.no_grad()
        def validation():
            model.train()
            for sample in tqdm(ds_iter_val, ncols=80):
                loss = h(data=sample, meters=meters_val)
                meters_val['loss_total'].add(loss.item())

        train_epoch()
        if epoch % args.val_epoch_interval == 0:
            validation()

        test_dict = None
        if epoch % args.test_epoch_interval == 0:
            model.eval()
            test_dict = run_eval(args, model, epoch)

        log_dict = dict()
        log_dict.update({
            'grad_norm':
            meters_train['grad_norm'].mean,
            'grad_norm_std':
            meters_train['grad_norm'].std,
            'learning_rate':
            optimizer.param_groups[0]['lr'],
            'time_forward':
            meters_time['forward'].mean,
            'time_backward':
            meters_time['backward'].mean,
            'time_data':
            meters_time['data'].mean,
            'gpu_memory':
            meters_time['memory'].mean,
            'time':
            time.time(),
            'n_iterations': (epoch + 1) * len(ds_iter_train),
            'n_datas':
            (epoch + 1) * args.global_batch_size * len(ds_iter_train),
        })

        for string, meters in zip(('train', 'val'),
                                  (meters_train, meters_val)):
            for k in dict(meters).keys():
                log_dict[f'{string}_{k}'] = meters[k].mean

        log_dict = reduce_dict(log_dict)
        if get_rank() == 0:
            log(config=args,
                model=model,
                epoch=epoch,
                log_dict=log_dict,
                test_dict=test_dict)
        dist.barrier()