Step 3. Main Transformations

Main transformations are the majority of low precision transformations. Transformations operate with dequantization operations. Main transformations include:

• AddTransformation

• AvgPoolTransformation

• BatchToSpaceTransformation

• ClampTransformation

• ConcatTransformation

• ConvolutionTransformation

• ConvolutionBackpropDataTransformation

• DepthToSpaceTransformation

• FakeQuantizeDecompositionTransformation

• FakeQuantizeTransformation

• InterpolateTransformation

• GroupConvolutionTransformation

• GatherTransformation

• MatMulTransformation

• MaxPoolTransformation

• MultiplyPartialTransformation

• MultiplyTransformation

• MVNTransformation

• NormalizeL2Transformation

• PadTransformation

• PReluTransformation

• ReduceMaxTransformation

• ReduceMeanTransformation

• ReduceMinTransformation

• ReduceSumTransformation

• ReluTransformation

• ReshapeTransformation

• SpaceToBatchTransformation

• SqueezeTransformation

• ShuffleChannelsTransformation

• SplitTransformation

• StridedSliceTransformation

• SubtractTransformation

• TransposeTransformation

• UnsqueezeTransformation

• VariadicSplitTransformation

Let’s explore some main transformations on the example model. Original model:

Result model after main transformations:

Changes in the example model after main transformation:

• All FakeQuantize operations (fakeQuantize1, fakeQuantize2 and fakeQuantize3) were decomposed:

  • original FakeQuantize operations were replaced with new operations with other output intervals and output port precision,

  • dequantization operations.

• Dequantization operations were moved via precision preserved (concat1 and concat2) and quantized (convolution2) operations.

Note

The left branch (branch #1) does not require per-tensor quantization. As a result, the fakeQuantize1 output interval is [0, 255]. But quantized convolution2 requires per-tensor quantization on the right branch (branch #2). Then all connected FakeQuantize interval operations (fakeQuantize1 and fakeQuantize2) are aligned to have per-tensor quantization after the concatenation (concat2) operation.