Hello I am getting this error “ValueError: steps_per_epoch must be > 0” when trying to train a custom dataset with TLT. I’ve tried without success: adjusting the batch_size_per_gpu adjusting num_epochs I am running this command: tlt-train detectnet_v2 -e /workspace/path_spec -r /workspace/pat…

Ok, figured it out. I was pointing to a specific tfrecord when needed to use “*” in tfrecords/*. Updated to the code below in the spec file and now runs good. dataset_config { data_sources: { tfrecords_path: “/workspace/home/label_images/tfrecords/*”

NVIDIA Developer Forums

Custom dataset -- ValueError: steps_per_epoch must be > 0

Accelerated Computing Intelligent Video Analytics TAO Toolkit

Morganh November 25, 2020, 2:18am 3

Refer to TLT Custom data set size question?