reload previous model in a new training

[jinhangw]

Hello,

I have a model trained with 500 images (Kitti data and ~200 images of my own). Now I want to add new images to improve this model.
My question is, instead of re-train with all 520 images, can I reload this model and add 20 new images to re-train and improve this model? Maybe something like saver = tf.train.import_meta_graph('my_test_model-1000.meta') in tensorflow. How can I achieve it with minimal modification on train.py or tensorvision code?

Thank you very much!

[MarvinTeichmann]

Go to your logdir, increase max_iterations in hyoes and use 'tv-continue' to continue training.

[jinhangw]

Thanks!
Do you mean hypes[solver][max_steps] and the continue_training(logdir) function in tensorvision.train? I'm still a bit confused about 'tv-continue'.

[StuvX]

I've successfully done what you're asking - in your RUNS directory there should be the saved model, go to model_files/hypes.json and increase max_iterations. Make sure that you have the new images referenced in your train.txt and val.txt files and that hypes.json points to these. Then to continue training use tv-continue --logdir RUNS/your_model/

[StuvX]

Oh, if you haven't already make sure you follow Marvin's instructions for installing tensorvision that are toward the bottom of the intro page.

[jinhangw]

Thank you very much for your help!

[jk7g14]

thanks guys I could do what I wanted to do after seeing these answers.