Created by: tonyyang-svail

Currently, batch_norm_grad_op is generated by the default grad_op_maker. This is not ideal since MeanOut@GRAD and VarianceOut@GRAD are unused in the programDesc.

Op(batch_norm_grad),
inputs:{
    Bias[batch_norm_0.b_0[64]({})],
    Mean[batch_norm_0.w_1[64]({})],
    MeanOut[batch_norm_0.w_1[64]({})],
    MeanOut@GRAD[batch_norm_0.w_1@GRAD[64]({})],
    SavedMean[batch_norm_0.tmp_0[64]({})],
    SavedMean@GRAD[batch_norm_0.tmp_0@GRAD[64]({})],
    SavedVariance[batch_norm_0.tmp_1[64]({})],
    SavedVariance@GRAD[batch_norm_0.tmp_1@GRAD[64]({})],
    Scale[batch_norm_0.w_0[64]({})],
    Variance[batch_norm_0.w_2[64]({})],
    VarianceOut[batch_norm_0.w_2[64]({})],
    VarianceOut@GRAD[batch_norm_0.w_2@GRAD[64]({})],
    X[conv2d_0.tmp_0[10, 64, 100, 100]({})],
    Y[batch_norm_0.tmp_2[10, 64, 100, 100]({})],
    Y@GRAD[batch_norm_0.tmp_2@GRAD[10, 64, 100, 100]({})]},
outputs:{
    Bias@GRAD[batch_norm_0.b_0@GRAD[64]({})],
    Mean@GRAD[],
    Scale@GRAD[batch_norm_0.w_0@GRAD[64]({})],
    Variance@GRAD[],
    X@GRAD[conv2d_0.tmp_0@GRAD[10, 64, 100, 100]({})]
}