Saki Shinoda

Tensorflow summary event files as numpy arrays

2017-02-13 01:58:00 +0000

Pass this function a directory path, and it returns the path to the most recently modified file in that directory. This can be used to get the most recent event file in your logdir.

import glob, os
def get_event(dir_path):
    return max(
      glob.glob('{}/*'.format(dir_path)),
                key=os.path.getctime)

I usually organise my logdirs into logs/{}/train/ and logs/{}/test/ where {} is filled by the string ID of whatever task/exercise the events are for. This makes for nice categorisation in Tensorboard as well as ease of extraction using the function above, which we might use to return the tuple of most recent logs for training and test respectively.

def get_events(ex_id):
    return (
      max(
        glob.glob('logs/{}/train/*'.format(ex_id)),
                  key=os.path.getctime),
      max(
        glob.glob('logs/{}/test/*'.format(ex_id)),
                  key=os.path.getctime))

Let scalar_str be the string name of the scalar summary value you want to retrieve, e.g. in the following example, the first argument to tf.summary.scalar is the value you would use as scalar_str.

    # op defining some metric you want to track
    accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
    tf.summary.scalar('accuracy', accuracy)

Given paths to event files, we can use built in Tensorflow functionality plus some list-to-array conversion and array manipulation to retrieve the values of the desired summary scalar. The following function returns a numpy array with one column corresponding to the global step and the other to the metric you have retrieved (lc stands for learning curve in this context).

import numpy as np
from tensorflow.python.summary import event_accumulator
def get_lc(event_file, scalar_str):
    ea = event_accumulator.EventAccumulator(event_file)
    ea.Reload()
    lc = np.stack(
      [np.asarray([scalar.step, scalar.value])
      for scalar in ea.Scalars(scalar_str)])
    return(lc)