Skip to main content
E0203: backend output shape mismatch

​
What happened

A hardware backend ran inference and returned an output raster, but the raster has a different number of timesteps (got_t ≠ t) or output neurons (got_n ≠ n) than the model’s declared output shape.

​
Why

This is a backend implementation bug. The backend must run exactly T timesteps and return exactly N_out neurons per the artifact’s declared output layer.

​
How to fix

  1. If you are integrating a hardware backend: verify that your backend’s run_batch implementation returns a tensor of shape [batch, T, N_out] where T and N_out come from the artifact.
  2. Check for off-by-one errors in timestep unrolling.
  3. If this occurs with a production THRINDEX release, file a bug with the artifact, target name, and the exact error message.

​
Example

  • E0206 — the same check performed per-sample in the conformance harness