search

rishizek / tensorflow-deeplab-v3-plus

DeepLabv3+ built in TensorFlow
MIT License
833 stars 307 forks source link

Why is training like this (use my own data set)? #43

Closed DahuoJ closed 5 years ago

DahuoJ commented 5 years ago

Hello!! @rishizek ,I can run it successfully! but I have some doubts. 4 y ifj z_2j z 2 emq z vz ef jq45l 43 v5q w

It's too slow, and it's not reasonable to save weights every 4 times of training. I'm not sure where I need to make code changes. I wish it cannot save checkpoint for 4 into ./model/model.ckpt and start evaluation. What should I do to make it display only the following two messages? w 4 g35m q cfhp ks8 3v

I guess it may be because my own data set is too small? I am very confused, How can I make it smooth? Any help will be appreciated!!

rishizek commented 5 years ago

Hi @DahuoJ , thanks for your interest in the repo. You may increase the number of training epochs between evaluations given by --epochs_per_eval, whose default value is 1.

DahuoJ commented 5 years ago

Thanks, this has an effect. Does this mean that --train_epochs also needs to be adjusted?

rishizek commented 5 years ago

No. --train_epochs indicates the number of iteration you train the training dataset. Whereas the --epochs_per_eval means how often you evaluate dateset during training. Though, of course, if the number of training is not sufficient, you need to increase --train_epochs.

DahuoJ commented 5 years ago

Thanks for your advice, I have started normal training and it takes about 14 hours to get results. 👍 👍