search

Hahn-Schickard / Automatic-Structured-Pruning

We have implemented a framework that supports developers to structured prune neural networks of Tensorflow Models
Apache License 2.0
27 stars 4 forks source link

Copy paste error: l1 vs. l2 metric #7

Closed vse-motec closed 2 years ago

vse-motec commented 2 years ago

When inspecting the code, I found a copy-paste error. The implementation for pruning dense layers has the L1 norm avg_neuron_w.append(np.average(np.abs(new_layer_param[0][:, i]))) and the L2 norm avg_neuron_w.append(np.linalg.norm(new_layer_param[0][:, i])) implemented correctly.

However, when inspecting the implementation for conv layers, the code currently is the following: L1 norm: avg_filter_w.append(np.average(np.abs(filters[0][:, :, :, i]))) L2 norm: avg_filter_w.append(np.average(np.abs(filters[0][:, :, :, i])))

This looks like a copy-paste error. According to the dense layer implementation, the L2 norm for conv layers should be avg_filter_w.append(np.linalg.norm(filters[0][:, :, :, i]))

I provide a pull request with a fix.