TFrecords for validation_data_sources do not work with SSD and DSSD

Are TFrecords supported for the validation set using SSD and DSSD?

Using the nvcr.io/nvidia/tao/tao-toolkit:4.0.0-tf1.15.5 Docker image directly. With the below dataset configurations the output of the following command ssd train -e /Config/experiment_config.txt -r /Training/ -k encryption_key -m /Pretrained/mobilenet_v2.hdf5 is as shown below the dataset config. The rest runs as expected just the validation step is missing.

dataset_config:
  data_sources:
    tfrecords_path: "/Data/train_val-fold-001*"
  include_difficult_in_training: true
  validation_data_sources:
    tfrecords_path: "/Data/train_val-fold-000*"
2023-02-15 08:22:48,649 [INFO] __main__: Number of images in the training dataset:	    41
2023-02-15 08:22:48,649 [INFO] __main__: Number of images in the validation dataset:	     0

Flipping the folds around gives the same problem, indicating that the records are valid:

dataset_config:
  data_sources:
    tfrecords_path: "/Data/train_val-fold-000*"
  include_difficult_in_training: true
  validation_data_sources:
    tfrecords_path: "/Data/train_val-fold-001*"
2023-02-15 08:23:15,246 [INFO] __main__: Number of images in the training dataset:	    10
2023-02-15 08:23:15,246 [INFO] __main__: Number of images in the validation dataset:	     0

Currently, SSD and DSSD do not support running evaluation against tfrecords file during training.
Please use below way.
validation_data_sources: {
label_directory_path: “/path/to/val/labels”
image_directory_path: “/path/to/val/images”

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.