I am trying to use the batch normalization layer from tensorflow, but the
mvNCCompile (NCSDK version 2.08.01) would not complete because it says:
[Error 5] Toolkit Error: Stage Details Not Supported: Neg.
I did look into why this happens and the batch normalize layer is actually split into the following nodes:
node.name= batchnorm/add/y node.type= Const
node.name= batchnorm/add node.type= Add
node.name= batchnorm/Rsqrt node.type= Rsqrt
node.name= batchnorm/mul node.type= Mul
node.name= batchnorm/mul_1 node.type= Mul
node.name= batchnorm/Neg node.type= Neg
Where the last one is indeed a Neg operation which, as far as I know, is simply a sign inversion of each value.
My concern is that it is said on the release page that the batch normalization layer should be supported.
Here is how I use the batch normalization layer, where
conv is the previous layer, and
scales are the weights of the batch normalization layer, of type tf.Variable:
conv = tf.nn.batch_normalization( x=conv, mean=means, variance=variances, offset=None, scale=scales, variance_epsilon=1e-5 )
Thank you very much!