Skip to content

Handle no-bias fusing of Dense and BN

Javier Duarte requested to merge github/fork/thesps/use_bias_fix into master

Created by: thesps

@vloncar fixed a bug related to Keras layers with use_bias=False and the Dense-BatchNorm layer fusion. The 'bias' values from the BatchNorm end up in the Dense layer, but the Dense layer's bias precision is overruled since the Keras layer had use_bias=False. Wrong results follow...

Now, in that scenario, the Dense layer (which is now the fused Dense-BatchNorm) picks up the bias precision of the BatchNorm.

Merge request reports