markdtw / condensenet-tensorflow Goto Github PK

In model.py, line 131, the condensation criterion:

wis = tf.reduce_sum(wi, axis=-1) # (w_in,)

probably should be:

wis = tf.reduce_sum(tf.abs(wi), axis=-1) # (w_in,)

Beacuse in the original paper authors uses 'the averaged absolute value of weights'.

Hi..The paper says there are concatenations from all dense layers outputs in each block to all other dense layer inputs of all other dense blocks in front of it. I see you have only concatenated dense layers inside each dense block but there is no concatenation from one dense block to another.

The author says this is actually one of the main differences from his previous paper Densenet . Am i missing anything or have i interpreted the paper wrongly?

Hi markdtw,
Thank you for the share!
I have tested your code and found that there was no drop in the number of parameters.
I set the epoch=4, and the parameter numbers are 1.43M after all of the four epochs.
So, it seems that the LGC in the code worded just like standard CONV.
It confused me.

do you have trained model.

Thank you for sharing!
Is Group Lasso regularization implemented in this repo?