NameError: name 'unet' is not defined

Gaming and Visualization Technologies APIs DirectX, DXR, DirectCompute
1

I try to use the Transfer Learning Toolkit-3.0 to train the Unet model. But I got error:
NameError: name ‘unet’ is not defined

Can you please advise how to fix this problem?

Thanks
Jimmy

  1. the TLT-3.0 docker image is use to run first

docker run --runtime=nvidia -it -v /home/jiande/workspace:/workspace nvcr.io/nvidia/tlt-streamanalytics:v3.0-dp-py3

  1. here is the spec.txt file I used
    checkpoint: “resnet101/resnet_101.hdf5” - Here I have downloaded the Unet model fro NGC
    model_config {
    num_layers: 18
    all_projections: true
    arch: “resnet”
    freeze_blocks: 0
    freeze_blocks: 1
    use_batch_norm: true
    training_precision {
    backend_floatx: FLOAT32
    }
    model_input_height: 320
    model_input_width: 320
    model_input_channels: 3
    }
    raining_config {
    batch_size: 2
    epochs: 3
    log_summary_steps: 10
    checkpoint_interval: 1
    loss: “cross_dice_sum”
    learning_rate:0.0001
    regularizer {
    type: L2
    weight: 3.00000002618e-09
    }
    optimizer {
    adam {
    epsilon: 9.99999993923e-09
    beta1: 0.899999976158
    beta2: 0.999000012875
    }
    }
    }
    data_config{
    image_size: “(832, 1344)”
    augment_input_data: True
    eval_samples: 5000
    training_file_pattern: "
    $DATA_DIR/train*.tfrecord"
    validation_file_pattern: “$DATA_DIR/val*.tfrecord”
    val_json_file: “$DATA_DIR/annotations/instances_val2017.json”
    num_classes: 91
    skip_crowd_during_training: True
    }

  2. command to traing
    tlt-train unet -e spec.txt -d ./workspace -k key --gpus 1