int8 MobileNetV2 Pipeline

A int8 MobileNetV2-shaped pipeline: two inverted-residual blocks (one stride-1 with a residual skip, one stride-2 without) wired around a stride-2 stem and a GAP + dense head. Demonstrates the inverted-residual unit and MobileNetV2’s linear-bottleneck rule on the int8 affine grid.

How it works

• Pipeline (NHWC, 16x16x4 input): QPad2D → QConv2DPerChannel 3x3 s=2 stem → qrelu, then IR block 1 (QPointwiseConv2D 8→32 expand → qrelu → QDepthwiseConv2D 3x3 s=1 → qrelu → QPointwiseConv2D 32→8 project → QAdd skip), then IR block 2 (expand → DW 3x3 s=2 → project 32→16, no skip), then QGlobalAvgPool2D → QDense 16→4 (int8 logits).
• Linear bottlenecks: the 1x1 projection convolutions are deliberately not followed by qrelu, matching MobileNetV2’s design rule (expand → DW → project keeps high-rank features in the expanded space and projects back without a nonlinearity). QDepthwiseConv2D is per-channel weight-scaled (TFLite mandate); qrelu and QGlobalAvgPool2D are pass-throughs that reuse the upstream (scale, zero_point).
• Every layer is pure integer at runtime with int32 accumulators; calibration (FLOAT=1 STD=1) is host-only. make run passes within 50% of logits range.

Build and run

cd examples/mobilenetv2_int8
make release
make run
make plot      # needs matplotlib; a venv/pyenv works if it is not already in your Python

Built with -DTINYMIND_ENABLE_QUANTIZATION=1. Extra targets:

• make bench — CSV cycle/byte report to output/mobilenetv2_int8.csv
• make golden — int8 logits for the 4-sample test set to output/mobilenetv2_int8.golden (consumed by unit_test/integration, locked byte-for-byte across SIMD gate combos)

Output

Left panel overlays the float reference against the int8-dequantized logits across the 4 output classes — the two tracks track the same monotone trend with a small gap at the extremes. Right panel shows the per-element absolute error staying under ~0.007, comfortably inside the 50% pass threshold for this two-inverted-residual-block pipeline.

Source on GitHub