Function QuantizeNodes

tensorflow/tools/graph_transforms/quantize_nodes.cc:589–950

Converts any float ops that have eight-bit equivalents into their quantized forms, so that as much calculation as possible is done in the lower-precision format.

// forms, so that as much calculation as possible is done in the lower-precision
// format.
Status QuantizeNodes(const GraphDef& input_graph_def,
                     const TransformFuncContext& context,
                     GraphDef* output_graph_def) {
  // Loop through all of the quantizable op types, and replace any occurrences
  // with equivalent sub-graphs with quantized ops at their core. For example
  // this one-input operation:
  //
  //            Input(float)
  //                |
  //                v
  //            Operation
  //                |
  //                v
  //             (float)
  //
  // Will be turned into it's quantized equivalent:
  //
  //      Input(float)          ReshapeDims
  //         +------v v-------------+
  //         |    Reshape
  //         |      |
  //         |      |          ReductionDims
  //         |      +-----+         |
  //         |      | +---c---------+
  //         |      v v   v v-------+
  //         |      Min   Max
  //         |  +----+      |
  //         v  v  v--------+
  //        Quantize
  //            |
  //            v
  //     QuantizedOperation
  //        |   |   |
  //        v   v   v
  //        Dequantize
  //            |
  //            v
  //         (float)
  //
  // This keeps the inputs and outputs visible to the rest of the graph in
  // float
  // and converts them down to quantized buffers internally for the
  // computation.
  // The result will end up with a lot of redundant dequantize/quantize pairs
  // between adjacent quantized ops, but a later pass removes these where it
  // can.

  std::set<string> ops_to_ignore;
  if (context.params.count("ignore_op") > 0) {
    for (const string& name : context.params.at("ignore_op")) {
      ops_to_ignore.insert(name);
    }
  }

  const std::vector<QuantizedOpInfo>& op_list = GetQuantizedOpList();
  string op_pattern;
  bool is_first = true;
  std::map<string, QuantizedOpInfo> op_map;