示例#1
0
def training_loops(net, train_loader, valid_loader, use_gpu, num_epochs, lr_schedule, label_names, label_names_temporal,
                   path_out, temporal_annotation_training=False, log_fn=print, confmat_event=None):
    criterion = nn.CrossEntropyLoss()
    optimizer = optim.Adam(net.parameters(), lr=0.0001)

    best_state_dict = None
    best_top1 = -1.
    best_loss = float('inf')

    for epoch in range(0, num_epochs):  # loop over the dataset multiple times
        new_lr = lr_schedule.get(epoch)
        if new_lr:
            log_fn(f"update lr to {new_lr}")
            for param_group in optimizer.param_groups:
                param_group['lr'] = new_lr

        net.train()
        train_loss, train_top1, cnf_matrix = run_epoch(train_loader, net, criterion, label_names_temporal,
                                                       optimizer, use_gpu,
                                                       temporal_annotation_training=temporal_annotation_training)
        net.eval()
        valid_loss, valid_top1, cnf_matrix = run_epoch(valid_loader, net, criterion, label_names_temporal,
                                                       None, use_gpu,
                                                       temporal_annotation_training=temporal_annotation_training)

        log_fn('[%d] train loss: %.3f train top1: %.3f valid loss: %.3f top1: %.3f'
               % (epoch + 1, train_loss, train_top1, valid_loss, valid_top1))

        if not temporal_annotation_training:
            if valid_top1 > best_top1:
                best_top1 = valid_top1
                best_state_dict = net.state_dict().copy()
                save_confusion_matrix(path_out, cnf_matrix, label_names, confmat_event=confmat_event)
        else:
            if valid_loss < best_loss:
                best_loss = valid_loss
                best_state_dict = net.state_dict().copy()
                save_confusion_matrix(path_out, cnf_matrix, label_names_temporal, confmat_event=confmat_event)

        # save the last checkpoint
        model_state_dict = net.state_dict().copy()
        model_state_dict = {clean_pipe_state_dict_key(key): value
                            for key, value in model_state_dict.items()}
        torch.save(model_state_dict, os.path.join(path_out, "last_classifier.checkpoint"))

    log_fn('Finished Training')
    return best_state_dict
示例#2
0
def train_model(path_in,
                path_out,
                model_name,
                model_version,
                num_layers_to_finetune,
                epochs,
                use_gpu=True,
                overwrite=True,
                temporal_training=None,
                resume=False,
                log_fn=print,
                confmat_event=None):
    os.makedirs(path_out, exist_ok=True)

    # Check for existing files
    saved_files = [
        "last_classifier.checkpoint", "best_classifier.checkpoint",
        "config.json", "label2int.json", "confusion_matrix.png",
        "confusion_matrix.npy"
    ]

    if not overwrite and any(
            os.path.exists(os.path.join(path_out, file))
            for file in saved_files):
        print(f"Warning: This operation will overwrite files in {path_out}")

        while True:
            confirmation = input(
                "Are you sure? Add --overwrite to hide this warning. (Y/N) ")
            if confirmation.lower() == "y":
                break
            elif confirmation.lower() == "n":
                sys.exit()
            else:
                print('Invalid input')

    # Load weights
    selected_config, weights = get_relevant_weights(
        SUPPORTED_MODEL_CONFIGURATIONS,
        model_name,
        model_version,
        log_fn,
    )
    backbone_weights = weights['backbone']

    if resume:
        # Load the last classifier
        checkpoint_classifier = torch.load(
            os.path.join(path_out, 'last_classifier.checkpoint'))

        # Update original weights in case some intermediate layers have been finetuned
        update_backbone_weights(backbone_weights, checkpoint_classifier)

    # Load backbone network
    backbone_network = build_backbone_network(selected_config,
                                              backbone_weights)

    # Get the required temporal dimension of feature tensors in order to
    # finetune the provided number of layers
    if num_layers_to_finetune > 0:
        num_timesteps = backbone_network.num_required_frames_per_layer.get(
            -num_layers_to_finetune)
        if not num_timesteps:
            # Remove 1 because we added 0 to temporal_dependencies
            num_layers = len(
                backbone_network.num_required_frames_per_layer) - 1
            msg = (f'ERROR - Num of layers to finetune not compatible. '
                   f'Must be an integer between 0 and {num_layers}')
            log_fn(msg)
            raise IndexError(msg)
    else:
        num_timesteps = 1

    # Extract layers to finetune
    if num_layers_to_finetune > 0:
        fine_tuned_layers = backbone_network.cnn[-num_layers_to_finetune:]
        backbone_network.cnn = backbone_network.cnn[0:-num_layers_to_finetune]

    # finetune the model
    extract_features(path_in,
                     selected_config,
                     backbone_network,
                     num_layers_to_finetune,
                     use_gpu,
                     num_timesteps=num_timesteps,
                     log_fn=log_fn)

    # Find label names
    label_names = os.listdir(directories.get_videos_dir(path_in, 'train'))
    label_names = [x for x in label_names if not x.startswith('.')]
    label_names_temporal = ['background']

    project_config = load_project_config(path_in)
    if project_config:
        for temporal_tags in project_config['classes'].values():
            label_names_temporal.extend(temporal_tags)
    else:
        for label in label_names:
            label_names_temporal.extend([f'{label}_tag1', f'{label}_tag2'])

    label_names_temporal = sorted(set(label_names_temporal))

    label2int_temporal_annotation = {
        name: index
        for index, name in enumerate(label_names_temporal)
    }
    label2int = {name: index for index, name in enumerate(label_names)}

    extractor_stride = backbone_network.num_required_frames_per_layer_padding[
        0]

    # Create the data loaders
    features_dir = directories.get_features_dir(path_in, 'train',
                                                selected_config,
                                                num_layers_to_finetune)
    tags_dir = directories.get_tags_dir(path_in, 'train')
    train_loader = generate_data_loader(
        project_config,
        features_dir,
        tags_dir,
        label_names,
        label2int,
        label2int_temporal_annotation,
        num_timesteps=num_timesteps,
        stride=extractor_stride,
        temporal_annotation_only=temporal_training,
    )

    features_dir = directories.get_features_dir(path_in, 'valid',
                                                selected_config,
                                                num_layers_to_finetune)
    tags_dir = directories.get_tags_dir(path_in, 'valid')
    valid_loader = generate_data_loader(
        project_config,
        features_dir,
        tags_dir,
        label_names,
        label2int,
        label2int_temporal_annotation,
        num_timesteps=None,
        batch_size=1,
        shuffle=False,
        stride=extractor_stride,
        temporal_annotation_only=temporal_training,
    )

    # Check if the data is loaded fully
    if not train_loader or not valid_loader:
        log_fn(
            "ERROR - \n "
            "\tMissing annotations for train or valid set.\n"
            "\tHint: Check if tags_train and tags_valid directories exist.\n")
        return

    # Modify the network to generate the training network on top of the features
    if temporal_training:
        num_output = len(label_names_temporal)
    else:
        num_output = len(label_names)

    # modify the network to generate the training network on top of the features
    gesture_classifier = LogisticRegression(
        num_in=backbone_network.feature_dim,
        num_out=num_output,
        use_softmax=False)

    if resume:
        gesture_classifier.load_state_dict(checkpoint_classifier)

    if num_layers_to_finetune > 0:
        # remove internal padding for training
        fine_tuned_layers.apply(set_internal_padding_false)
        net = Pipe(fine_tuned_layers, gesture_classifier)
    else:
        net = gesture_classifier
    net.train()

    if use_gpu:
        net = net.cuda()

    lr_schedule = {
        0: 0.0001,
        int(epochs / 2): 0.00001
    } if epochs > 1 else {
        0: 0.0001
    }
    num_epochs = epochs

    # Save training config and label2int dictionary
    config = {
        'backbone_name': selected_config.model_name,
        'backbone_version': selected_config.version,
        'num_layers_to_finetune': num_layers_to_finetune,
        'classifier': str(gesture_classifier),
        'temporal_training': temporal_training,
        'lr_schedule': lr_schedule,
        'num_epochs': num_epochs,
        'start_time': str(datetime.datetime.now()),
        'end_time': '',
    }
    with open(os.path.join(path_out, 'config.json'), 'w') as f:
        json.dump(config, f, indent=2)

    with open(os.path.join(path_out, 'label2int.json'), 'w') as f:
        json.dump(
            label2int_temporal_annotation if temporal_training else label2int,
            f,
            indent=2)

    # Train model
    best_model_state_dict = training_loops(
        net,
        train_loader,
        valid_loader,
        use_gpu,
        num_epochs,
        lr_schedule,
        label_names,
        path_out,
        temporal_annotation_training=temporal_training,
        log_fn=log_fn,
        confmat_event=confmat_event)

    # Save best model
    if isinstance(net, Pipe):
        best_model_state_dict = {
            clean_pipe_state_dict_key(key): value
            for key, value in best_model_state_dict.items()
        }
    torch.save(best_model_state_dict,
               os.path.join(path_out, "best_classifier.checkpoint"))

    config['end_time'] = str(datetime.datetime.now())
    with open(os.path.join(path_out, 'config.json'), 'w') as f:
        json.dump(config, f, indent=2)
示例#3
0
        net = net.cuda()

    lr_schedule = {0: 0.0001, 40: 0.00001}
    num_epochs = 80
    best_model_state_dict = training_loops(
        net,
        train_loader,
        valid_loader,
        use_gpu,
        num_epochs,
        lr_schedule,
        label_names,
        path_out,
        temporal_annotation_training=temporal_training)

    # Save best model
    if isinstance(net, Pipe):
        best_model_state_dict = {
            clean_pipe_state_dict_key(key): value
            for key, value in best_model_state_dict.items()
        }
    torch.save(best_model_state_dict,
               os.path.join(path_out, "best_classifier.checkpoint"))

    if temporal_training:
        json.dump(label2int_temporal_annotation,
                  open(os.path.join(path_out, "label2int.json"), "w"))
    else:
        json.dump(label2int, open(os.path.join(path_out, "label2int.json"),
                                  "w"))