Converting Pytorch To Tflite

Converting PyTorch to TFLite

Export a trained checkpoint to a mobile TFLite whose output is numerically identical to the PyTorch reference (within fp16 rounding), and whose input is the raw camera colour format so no per-frame colour conversion is needed on-device. This is a methodology + gotcha skill; you drive the repo's converter (e.g. a TinyNN / onnx2tf / ai-edge-torch based exporter), it does not ship one.

When to use

• Exporting a PyTorch/ONNX checkpoint to .tflite for phone/NPU deployment.
• A converted model looks "close on average" but wrong at edges/thin structures.
• The device delivers YUV or BGR but the model was trained in a different colour space.

The pipeline

1. Load the checkpoint robustly (see loading-model-checkpoints): unwrap nesting, match the prefix by key-overlap, confirm missing=0 unexpected=0, read input channels from the first conv.
2. Set deploy / reparameterize mode BEFORE tracing (gotcha 3).
3. Fold the colour transform into the first conv weight(s) (gotcha 1 & 2) if the device colour space differs from training — so the exported graph eats raw camera data.
4. Convert to TFLite: start with fp16 (safest, ~1e-4 parity); move to int8-hybrid only when latency requires it (see quantizing-on-device-models).
5. Parity-gate the result against PyTorch on real images before trusting it (see below).

Gotcha 1 — fold colour conversion into the first conv (don't convert at runtime)

A linear colour transform C (3×3) commutes with a convolution: W·(C·x) == (W·C)·x. So instead of converting the camera frame to the training colour space every frame, pre-multiply the first conv's weights by C and the exported model consumes the raw format directly.

• Know your matrix's channel-order convention. An RGB↔YUV matrix is often derived for a specific output order (e.g. it reconstructs BGR, not RGB). If your checkpoint was trained in the other order, apply the channel-reorder first, then the colour matrix, so the two cancel correctly.
• For a multi-channel model (e.g. RGB+trimap), fold only the colour channels; leave auxiliary channels untouched.
• Verify the fold algebraically and numerically — a correct fold gives parity ≈ 1e-4 (fp16).

Gotcha 2 — multi-stem architectures have MORE THAN ONE input conv

Some backbones feed the raw input into two places (a stem branch and an encoder stem). If you fold the colour transform into only one, the other still expects the original colour space. The tell-tale symptom: low mean error but high max_abs at edges (globally similar, locally wrong). Fold every input conv. Identify them by a known per-arch map; fallback: fold every 4-D weight whose shape[1] == in_channels.

Gotcha 3 — set deploy mode before tracing (reparameterizable models)

Fused-conv / reparameterizable backbones collapse branches only when switched to deploy mode. Wrappers often nest the real net under an attribute, so set the flag on both the wrapper and the inner model before you trace/export, or the exported graph is the un-fused training graph.

if hasattr(model, "deploy"): model.deploy = True
inner = getattr(model, "model", None)
if inner is not None and hasattr(inner, "deploy"): inner.deploy = True
model.eval()

Parity gate (always run after export)

Run the same preprocessing through PyTorch (training colour) and TFLite (device colour) on a sample of real images; report mae / rmse / max_abs and a task metric (e.g. binary_iou for masks). Gate:

• mae ≈ 1e-4 → pass (fp16 rounding).
• larger mae, or max_abs near 1.0 at sparse pixels → investigate (usually a colour-order or missed-stem bug), do not ship.

See references/export-gotchas.md for the colour-fold algebra, a per-arch input-conv checklist, and fp16-vs-int8 converter-flag notes.

Boundaries

• Methodology + parity discipline only: you drive the actual converter tool; this does not vendor one.
• The static runtime-envelope check is gating-tflite-op-envelopes; the on-device numerical + latency sign-off is validating-on-device-inference; quantization decisions are quantizing-on-device-models. This skill owns the export + numerical-parity step between them.
• All model names/paths/colour matrices are caller inputs — keep zero project values in any eval you run.