Validation accuracy for object detection on custom dataset (SSD model)

I’m trying to finetune an SSD model to detect a certain object in a custom dataset.

I downloaded the images, created the .lst files, and the corresponding train.rec and test.rec files following the tutorial. I first used a slightly modified version of and everything works ok, but this way the accuracy on the validation set is not provided.

I would like to track it in order to decide when to stop the training, and I seem to understand that the validation accuracy is calculated when using the full script.
However, I’m having trouble figuring out how to set the right arguments to run it on my custom dataset.

I tried something in the line of:

python --data-shape=512  --dataset-root=custom_dataset_path/ --epochs=10

but I’m getting the error:

OSError: ~/.mxnet/datasets/voc is not a valid dir. Did you forget to initialize datasets described in [...]

Any suggestions? Thanks!