def quantize(self,tensor):
# TF alternative (rounding mantissa):
# TF alternative is not working with slim, because it produces 2 additional layers at once(float16 and float32). sparsity estimation tries to take both, but fails.
#o_type=tensor.dtype
#result=tf.cast(tf.cast(tensor,tf.float16),tf.float32)
#print(result)
#return result
# kernel alternative (cutting off mantissa):
return Wrapped.quant_halffp(tensor)
def quantize(self,tensor):
return Wrapped.quant_sparse(tensor, self.threshold)
def quantize(self,tensor):
return Wrapped.quant_log(tensor)