Activation statistics: fix computation of Channel-wise APoZ

Instead of returning the average-percentage-of-zeros, we returned the
average-percentage-of-non-zeros.
So inverted the results, and also multiplied by 100, because the name
of the statistic has "percentage" not "fraction" (not very important,
but still...)

distiller/utils.py

@@ -395,7 +395,7 @@ def activation_channels_apoz(activation):
         featuremap_apoz_mat = activation.abs().gt(0).sum(dim=1).float() / activation.size(1)  # batch x 1
     else:
         raise ValueError("activation_channels_apoz: Unsupported shape: ".format(activation.shape))
-    return featuremap_apoz_mat.mean(dim=0).cpu()
+    return 100 - featuremap_apoz_mat.mean(dim=0).mul(100).cpu()
 
 
 def log_training_progress(stats_dict, params_dict, epoch, steps_completed, total_steps, log_freq, loggers):

tests/test_basic.py

@@ -66,7 +66,7 @@ def test_activations():
                         [7.,  0.,  8.],
                         [0.,  9.,  0.]]]])
     assert all(distiller.activation_channels_l1(x) == torch.tensor([21/2,  45/2]))
-    assert all(distiller.activation_channels_apoz(x) == torch.tensor([6/18,  9/18]))
+    assert all(distiller.activation_channels_apoz(x) == torch.tensor([100*(6+6)/(9+9),  100*(4+5)/(9+9)]))
     assert all(distiller.activation_channels_means(x) == torch.tensor([21/18,  45/18]))