feat: report the feature vector the model was actually fed - #39
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Segunda ronda de instrumentación. El panel ahora resume, por parte del cuerpo, el vector que el modelo recibió de verdad en la última ventana, junto a la misma estadística medida sobre las 598 grabaciones de test de SWL-LSE.
Motivo: en el navegador el modelo puntúa cerca del ruido (5-16%) mientras mide 0.741 offline, y la paridad TS/Python del vector y del GRU ya está verificada — así que lo que difiere es la entrada. Ninguna ablación aislada (ventana corta, aspecto, manos intercambiadas, sin pose) reproduce esos números, así que el siguiente paso es mirar el vector en vez de seguir conjeturando.
Incluye un test que fija algo que no era obvio: las coordenadas de las manos son relativas al torso, así que si falta la pose el bloque de manos no se degrada, se queda a cero.
No cambia el comportamiento de reconocimiento.