Mikhail Nikitin

@tungdop2 hello ! Xavier_normal is one of default choices to initialize conv and dense layers. Also, in authors' implementation weights of classifier are actually updating: https://github.com/ydwen/opensphere/blob/main/runner.py#L98 So, it looks that...

You should export LargeMarginInnerProduct weights into new InnerProduct layer without bias term.

@gxstudy Your deploy fc layer should look like: ``` layer { name: "fc9" type: "InnerProduct" bottom: "fc8" top: "fc9" inner_product_param { num_output: 2 bias_term: false } } ```