How to save intermediate tensors that are used during the application lifecycle

Basically, I am running the evaluation script using:

python3 object_detection/legacy/eval.py \
    --pipeline_config_path=../../faster_rcnn_resnet101_voc12.pbtxt\
    --checkpoint_dir=../../trainedmodel_voc07 \
    --eval_dir=../../eval_results \
    --run_once=True \
    --alsologtostderr

Now, this script outputs the bounding boxes on image along with prediction scores. But what I need is, the metrics calculated.

When I looked into the eval.py, I found that, it returns nothing. It run the evaluation, calculates the metrics but doesn't return it. So, I just changed the code to return the metrics and print it on console.

metrics = evaluator.evaluate(
        create_input_dict_fn,
        model_fn,
        eval_config,
        categories,
        FLAGS.checkpoint_dir,
        FLAGS.eval_dir,
        graph_hook_fn=graph_rewriter_fn,
    )
print(metrics)

But, as we know, this won't print the values inside the metric. I have tried several things like metrics.numpy() which won't work when eager execution is turned off. I have also tried using metrics.eval() by creating a session but this just hangs up the PC and does nothing.

Also, tf.print() won't work as it runs in graph mode.

Then I found that TF stores summaries of metrics and we can visualize it using tensorboard. That's okay.. but what I wanted to extract values from other tensors?

So, basically, how can I access the intermediate tensors in the script?