Recency weights issue

        with numpyro.plate(
            "data_plate", size=y.shape[0]
            with handlers.scale(scale=sample_weights):
                    numpyro.sample("y", dist.Normal(mu, var), obs=y)

I have this setup where i have sample weights defined to scale the log likelihood. right now the weights are simply a numpy array which have mean of 1. However, these weights do not seem to be doing anything… when i change the weights, the model is unchanged. am i doing something wrong? thanks so much

@fehiepsi @martinjankowiak any assistance here would be much appreciated. It seems sample weights only hurt the model even though weights definitely should help (and do for a simple lin reg)

handlers.scale does indeed “work”. you haven’t provided any evidence to the contrary so not clear how to help. maybe your expectations about the consequence of scaling the log likelihood are incorrect?