Understanding Detection API config file

Question

I want to use "coco_detection_metrics". I read in forums that I should add metrics_set: "coco_detection_metrics" to eval_config:

eval_config: {
  num_examples:2000
  max_evals: 10
  eval_interval_secs: 5
  metrics_set: "coco_detection_metrics"
}

But there are two config files for each model and I see "eval_config" in both of them, for example for "ssd_mobilenet_v1_coco":

1- ssd_mobilenet_v1_coco.config

(located in: **samples/configs/**)

2- ssd_mobilenet_v1_coco_2018_01_28/pipeline.config

(located in: **ssd_mobilenet_v1_coco_2018_01_28.tar.gz**)

Which one should be modified? What is the difference of these two files? Which one will be used during training or evaluation?

Thank you!

score 0 · Accepted Answer · answered Jul 24 '18 at 18:05

Modify the one that you pass to the train.py script as a flag:

python3 object_detection/train.py --logtostderr --pipeline_config_path=/path/to/your/config_file.config --train_dir=/your/train/dir

So, create your config file wherever on your file system, modify it as you want, and then pass it to the train or eval scripts as shown above.

1 Answers1