Pre-trained conv layers weight about yolo_tensorflow

[hizhangp]

You can add an expression self.restorer = tf.saver(self.net.collection[:-6]) before self.saver, and then change self.saver.restore(self.sess, self.weights_file) into self.restorer.restore(self.sess, self.weights_file) to avoid restoring weights from fc layers.

[hizhangp]

Just change expression self.restorer = tf.saver(self.net.collection[:-6]) into self.restorer = tf.saver(self.variable_to_restore[:-6]) in the new version of training code.

[BeSlower]

Really thanks for your help.

[CompilerBian]

I met the same problem. But adding these two expressions don't work for the recently updated version of the training code, how to restore the weight file without fc layers now? Thanks!

[CompilerBian]

Thanks for your help. I have successfully trained the model with my own dataset.

[feitiandemiaomi]

@hizhangp I understand , it is answer I need , and , if I just change the some layer ,eg 8th conv layer ,what I should modify for train from the scratch

[ss199302]

@hizhangp Hello, why hasn't my loss changed from the beginning to the end of training? Do you know what the reason is? Thank you very much!

[ss199302]

@hizhangp Hello, why hasn't my loss changed from the beginning to the end of training? Do you know what the reason is? Thank you very much!