Assume that I am going to do more training with a similar data set in the future, is there any benefit to me using a fine tune checkpoint from a model that I created from my own training as opposed to the original SSD_Mobilenet_V1 version (for example 5000 images and 50000 steps). Does it improve any future training or am I just better off using the original one every time I train? I'm probably searching for the wrong thing, but I cannot find anything that suggests improvement. In my reading, I thought that the fine tune only helped with the final layer which leads me to believe its a pointless exercise no matter how good the resulting model was. Can anyone confirm this point and any useful reading for reference?
$\begingroup$
$\endgroup$
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
2
Yes, you can. After the below lines everything else will be same as what you have done previously to save the checkpoint.
with tf.Session() as sess:
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint('./')) # checkpoint file path
Please refer to the below links:
https://nathanbrixius.wordpress.com/2016/05/24/checkpointing-and-reusing-tensorflow-models/
https://github.com/tensorflow/nmt/issues/51
https://blog.metaflow.fr/tensorflow-saving-restoring-and-mixing-multiple-models-c4c94d5d7125
-
$\begingroup$ Thanks for your response. 2 more questions. Should i need to continue with the original data set and secondly, is there a risk of overfitting if i train in this fashion and train for too long? $\endgroup$ Commented Dec 30, 2018 at 22:51
-
$\begingroup$ You can change the data set as long as the there is no changes in the features. It is difficult to predict the overfitting before the actual training happens. But, yes there are chances that model might overfit if you train for too long. $\endgroup$ Commented Mar 18, 2019 at 5:44