OpsNeeded.md

These notes are for aiding a TF.Net translation to DiffSharp

Base Operations needed for BERT Inference

This is a list of the base TensorFlow ops needed to run BERT. This list excludes higher level ops which are compositions of the base ops. The list also excludes trivial, or framework specific ops.

• Add
• BatchMatMul
• BiasAdd
• Cast
• ExpandDims
• GatherV2
• MatMul
• Mean
• Merge
• Mul
• OneHot
• Pow
• Reshape
• Rsqrt
• Slice
• Softmax
• SquaredDifference
• Squeeze
• StopGradient
• StridedSlice
• Sub
• Tanh
• Transpose

Additional Base Operations needed for BERT training

• AddN
• ArgMax
• BatchMatMul
• BiasAdd
• BiasAddGrad
• BroadcastGradientArgs
• ConcatV2
• DynamicStitch
• FloorDiv
• FloorMod
• Greater
• GreaterEqual
• InvertPermutation
• IsFinite
• L2Loss
• Log
• Maximum
• Minimum
• Neg
• Pack
• Pad
• Prod
• Range
• RealDiv
• Reciprocal
• RsqrtGrad
• Sqrt
• Square
• StridedSlice
• StridedSliceGrad
• Sum
• Tanh
• TanhGrad
• Tile
• UnsortedSegmentSum

Additional Base Operations needed for Asertions

• All
• Assert
• Equal
• Switch