Florian FHermisch

SiLiKhon / activation_monitoring_callback.py

Last active May 2, 2021 14:26

Ugly code to monitor layer activations when training tf keras models with `model.fit` or similar.

	# Disclaimer: this is just an ugly solution to monitoring layer activations
	# when training a model in tf keras.

	# this code is executed right after the model is created.
	class Callback(tf.keras.callbacks.Callback):
	def __init__(self, logging_data):
	super().__init__()

	# `logging_data` is a namedtuple of the follwoing type:
	# LoggingData = namedtuple("LoggingData", ['get_val_dataset_fn', 'get_writer_fn', 'freq'])