resnet_block_int8

resnet_block_int8

Phase 10 demonstration: int8 ResNet-style residual block.

What it shows

A single residual block stitched together from the Phase-10 composition ops and per-channel QConv2DPerChannel:

input  [4][4][2]
   │
   ├── QPad2D(1,1) ── QConv2DPerChannel(3x3, per-channel) ── qrelu
   │                                                          │
   │                                                          QPad2D(1,1)
   │                                                          │
   │                                                          QConv2DPerChannel(3x3, per-channel)
   │                                                          │
   └─────────────────────────────────────► QAdd ◄─────────────┘
                                            │
                                          qrelu
                                            │
                                       output [4][4][2]

QPad2D carries SAME padding around the VALID QConv2DPerChannel to preserve spatial dimensions through both convolutions; pad value is set to each tensor's input zero point so the padded cells decode back to real-zero in the affine domain. QAdd rescales the conv2 branch and the identity skip onto a shared output grid using TFLite ADD semantics (buildQAddParams). The closing qrelu clamps at the sum-tensor zero point.

How it runs

The driver hand-builds small but non-trivial weights, runs the block in float over an 8-sample synthetic dataset to collect activation ranges, calibrates per-tensor activation params (computeAffineParamsAsymmetric) and per-channel symmetric weight scales (computePerChannelSymmetricScales), quantizes weights and biases, builds per-channel Requantizers, then runs the block end-to-end on int8. Output is dequantized and compared against the float reference.

Numbers

After make run:

• conv1 single-layer max-abs error: ~0.5 % of conv1 range
• block end-to-end max-abs error: ~15 % of output range (8 samples)

Pass threshold is 25 % of the dynamic range — generous because a 4-stage int8 chain without QAT or cross-layer equalization will naturally pick up double-digit relative error. Phase 15 (calibration upgrades: percentile, KL, conv+bn fold, CLE) tightens this.

Files

• resnet_block_int8.cpp — driver. Self-contained; no external weight file. Uses QConv2DPerChannel, QPad2D, QAdd, qrelu, plus calibration helpers from cpp/include/qcalibration.hpp.
• Makefile — builds with TINYMIND_ENABLE_FLOAT=1 STD=1 QUANTIZATION=1.

Build and run

make clean && make && make run