suitable prune rule

[PiseyYou]

When training, there are something different between your code and @Lam1360 code, such as:

in Lam1360 code, the training is:
1, python train.py --model_def config/yolov3-hand.cfg -sr --s 0.01

but in yours, missing the back value.

2，for step2, in Lam1360 code, he didn't show the suitable prune rule, but in your code, it is

python test_prune.py

no suitable prune rule following.

I was wonder whether you have already test and the result is fine, or you missing some key point.

[PiseyYou]

if the auther has already test the result, can you upload the result, and there are more information can discuss which can lead to the results we want.

[PiseyYou]

in @talebolano code, the prune rule is:

python prune.py --cfg yolov3.cfg --weights checkpoints/yolov3_sparsity_100.weights --percent 0.3

the --percent 0.3 is following, I wonder whether this it the suitable value, or more test result.

[violet17]

It is the same...... This parameter is " --s 0.01" is the default parameter in the code.

The prune rule is (1)避免剪掉所有channel的最高阈值(每个BN层的gamma的最大值的最小值即为阈值上限)
(2)剪枝的时候会先对 bias 做处理，用其后的卷积层或者 BN 层来吸收这个参数，然后再将权重赋值给 compact model，
这样做的话就能保证 pruned_model （只对 weight 置 0 的模型）和 Compact_model 对相同的输入有相同的输出

The prune ratio mainly depend on the application.