Sorry if my question is at the wrong place, I'm new in this community.

So, I have dataset with total of 1 million images (augmented) that separated in 28 classes. I followed this tutorial https://www.tensorflow.org/hub/tutorials/image_retraining to do transfer learning using Inception V3 in Tensorflow to create my own model. But I have no strong background in ML or DL so I'm not sure how to tune the parameter correctly for training step.

This is the training source code that I'm using : https://github.com/tensorflow/hub/raw/master/examples/image_retraining/retrain.py
Using the default setting I was able to get 80~84% accuracy with 16.000 steps. I've tried to change the training, validation, test ratio, training_steps, batch. But, still the accuracy is below 90%. So, Im seeking for advice which parameter that should I change to achieve good accuracy.


  • training_steps
  • learning_rate
  • testing_percentage
  • validation_percentage
  • eval_step_interval
  • train_batch_size
  • test_batch_size
  • validation_batch_size

Thank you

  • 1
    $\begingroup$ Tweak the learning rate. Usually a low value gives better performance, but not too low because it'll have a slower convergence. $\endgroup$ – Nightmerker Jan 24 '19 at 21:38
  • $\begingroup$ my current learning rate is 0.001 and the training takes more than 1 day. Do I need to reduce it again? Or reduce the learning rate but also reduce the training_steps to reduce training time? $\endgroup$ – gameon67 Jan 25 '19 at 0:52
  • $\begingroup$ This is up to fine tuning the parameters, you got a pretty decent accuracy percent for being the first time, don't be discouraged for it. I don't think that reducing the training steps will help, try using a 0.0005 learning rate and a 80% training/validation and 20% test $\endgroup$ – Nightmerker Jan 25 '19 at 3:58

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.