Show more images in Tensorboard - Tensorflow object detection

Question

I am using Tensorflow's object detection framework. Training and evaluation jobs are going well, but in tensorboard I am only able to see 10 images for the evaluation job. Is there a way to increase this number to look at more images? I tried changing the config file:

eval_config: {
  num_examples: 1000
  max_evals: 50
}

eval_input_reader: {
  tf_record_input_reader {
    input_path: "xxx/eval.record"
  }
  label_map_path: "xxx/label_map.pbtxt"
  shuffle: false
  num_readers: 1
}

I thought the max_eval parameter would change this but it doesn't.

This is the command i'm running for the evaluation job:

python ../models/research/object_detection/eval.py \
    --logtostderr \
    --pipeline_config_path=xxx/ssd.config \
    --checkpoint_dir="xxx/train/" \
    --eval_dir="xxx/eval"

Answer 1

It should be the num_visualizations parameter in your eval_config (cf. eval.proto code).

Answer 2

I've been able to get this to work in Tensorboard 1.11.0 by editing the object_detection/protos/eval.proto file, then re-running protoc (see the Tensorflow docs). For example, this line in eval.proto would enable 100 examples (instead of the default 10):

optional uint32 num_visualizations = 1 [default=100];

This probably has an impact on system memory, browser performance, eval performance, etc.. so use with caution.

Answer 3

Probably the easiest way is to add command line argument --samples_per_plugin

Full example

tensorboard --logdir . --samples_per_plugin=images=100

https://github.com/tensorflow/tensorboard/issues/1012