Training Retinanet

Accelerated Computing Intelligent Video Analytics TAO Toolkit
#1

Hi,

I am facing some difficulties while trying to train Retinanet with my own Dataset.
Here the snapshot describing the error encountered:

image
image1848×794 129 KB

In the training specification txt file, I changed the size of input image such that it corresponds to my data (640x352).

I would be pleased if someone could help me to fix this.

#3

There should be some issues from your tfrecords.
Please share the full log and spec when you run tlt-dataset-convert.

#4

The specification file retinanet_tfrecords_kitti_trainval.txt:

  root_directory_path: "/workspace/tlt-experiments/data/training"
  image_dir_name: "image_2"
  label_dir_name: "label_2"
  image_extension: ".jpg"
  partition_mode: "random"
  num_partitions: 2
  val_split: 14
  num_shards: 10
}
image_directory_path: "/workspace/tlt-experiments/data/training"

And the log of the cmd tlt-dataset-convert:

image
image1910×825 174 KB

#5

Your val image is too little, only 6.
Please refer to Training detectnet_v2 Issue

  1. val_images >= num_shards
  2. train_images >= num_shards
#6

Thanks a lot, that just works!

#8