Tools to analyse models

accosmin / nano

C++ library [machine learning & numerical optimization] - superseeded by libnano

MIT License

1 stars 0 forks source link

Closed accosmin closed 6 years ago

accosmin commented 9 years ago

activation histograms [-1,+1] per node or per unit (during training or evaluation): should not observe highly skewed distributions if training succeeds
weight/bias statistics (e.g. L2) per node or per unit
save the convolution filters per node (using image_grid_t)