Analysis of Results

[scottfernandes]

Hi, I would like to analyze the results after imputation, is there anyway I could see the probabilities as tabular models use conditional probability under the hood, or any other way for analysis

[jacobf18]

@scottfernandes That functionality isn't there at the moment but it wouldn't be hard to add. You would just need to pull out the preds variable in the following code (if you are using TabImpute v2 interface):

def _get_imputation_single(
        self, X_normalized: np.ndarray, num_repeats: int = 1
    ) -> np.ndarray:
        if num_repeats > 1:
            raise ValueError("`num_repeats > 1` is not supported in TabImputeV2.")

        X_normalized_tensor = torch.from_numpy(X_normalized).to(self.device)
        with torch.no_grad():
            X_normalized_tensor = X_normalized_tensor.unsqueeze(0).to(torch.bfloat16)
            preds = self.model(X_normalized_tensor)
            medians = self.bar_distribution.median(logits=preds).squeeze(0)

        X_imputed = medians.cpu().detach().numpy()
        missing_mask = np.isnan(X_normalized)
        X_normalized[missing_mask] = X_imputed[missing_mask]
        return X_normalized, X_imputed

The preds variable is the logits and will be shaped (1, num_rows, num_cols, 5000) where the 5000 is over the binned space. You would then use the self.borders FulllBarDistribution (check out the tabimpute.model.bar_distribution file) which would let you get any relevant quantities you would want (quantiles, variance, mean, etc.). However, before using those values directly, note that the matrix is normalized per column by subtracting mean and dividing by standard deviation, so the values used in the self.borders file would need to be "denormalized" back to the original space.