Is it technically possible to give one image directory and/or one coco annotation file in the data config?

According to the official example, the config used to transform a dataset to tfrecord look like this:

coco_config {
  root_directory_path: "/workspace/tao-experiments/data/coco"
  img_dir_names: ["val2017", "train2017"]
  annotation_files: ["annotations/instances_val2017.json", "annotations/instances_train2017.json"]
  num_partitions: 2
  num_shards: [32, 256]
image_directory_path: "/workspace/tao-experiments/data/coco"

In the above example, the training and validation images are in two seperate directories. Also, the training and validation annotations are in two seperate json files.

However, there might be two other scenarios:

  • Scenario 1: The train and validation images are in a common directory:
    In this case, is it possible to configure the img_dir_names as follows?
    img_dir_names: [“common2017”, “common2017”]

  • Scenario 2: The annotations are not already splitted:
    For instance, there would only be one annotation file containing all the annotations (train and val). Is there a way that tao model <model_name> dataset-convert takes care of this splitting out-of-the-box?

For above two scenarios, the train and validation images are in only one folder “common2017”, and also their annotations are not splitted. The dataset-convert would not split them to train and validation.
You can set

img_dir_names: ["common2017"]
annotation_files: ["annotations/common2017.json"]
num_partitions: 1
num_shards: [256]

Then, after conversion, you can split 256 tfrecords files to train part and validation part.

Would the dataset-convert put the data in a distributed manner in the tfrecords? are they somehow stratified by default or should I add some stratification?

Refer to DetectNet_v2 - NVIDIA Docs.
For kitti_config, we can set partition_mode: "random" in order to put the data randomly in the tfrecords. More info can be found in, .
For coco_config, details about the partition and its shard can be found in

But this option is only available for the kitti config, not for a coco based config. In such case, I need to transform my dataset to kitti.

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