From 667997a2c948dbc72a1ef67984a7b46e6bc0b95e Mon Sep 17 00:00:00 2001
From: Joerg Martin <joerg.martin@ptb.de>
Date: Wed, 5 Jan 2022 13:16:47 +0000
Subject: [PATCH] Fixed true_res
---
Experiments/evaluate_metrics.py | 17 ++++++++++-------
1 file changed, 10 insertions(+), 7 deletions(-)
diff --git a/Experiments/evaluate_metrics.py b/Experiments/evaluate_metrics.py
index e72dc11..2f8f6c8 100644
--- a/Experiments/evaluate_metrics.py
+++ b/Experiments/evaluate_metrics.py
@@ -111,7 +111,8 @@ def collect_metrics(x_y_pairs, seed=0,
# RMSE
training_state = net.training
net.train()
- not_averaged_predictions = net.predict(x, number_of_draws=noneiv_number_of_draws,
+ not_averaged_predictions = net.predict(x,\
+ number_of_draws=noneiv_number_of_draws,
take_average_of_prediction=False)
noneiv_mean = torch.mean(not_averaged_predictions[0], dim=1)
if len(y.shape) <= 1:
@@ -127,9 +128,11 @@ def collect_metrics(x_y_pairs, seed=0,
noneiv_metrics['rmse'] = np.sqrt(np.mean(scaled_res**2))
noneiv_metrics['bias'] = np.mean(scaled_res)
noneiv_metrics['coverage_numerical'], noneiv_metrics['coverage_theory'] =\
- epistemic_coverage(not_averaged_predictions, y, normalize_errors=False)
+ epistemic_coverage(not_averaged_predictions, y,\
+ normalize_errors=False)
noneiv_metrics['coverage_normalized'],_ =\
- epistemic_coverage(not_averaged_predictions, y, normalize_errors=True)
+ epistemic_coverage(not_averaged_predictions, y,\
+ normalize_errors=True)
noneiv_metrics['res_std'] = normalized_std(not_averaged_predictions, y)
# metrics that need a ground truth
@@ -138,8 +141,8 @@ def collect_metrics(x_y_pairs, seed=0,
noneiv_metrics['true_coverage_theory'] =\
epistemic_coverage(not_averaged_predictions, true_y,
normalize_errors=False, noisy_y=False)
- true_res = true_y - noneiv_mean
- noneiv_metrics['true_rmse'] = np.sqrt(np.mean(scaled_res**2))
+ true_res = (true_y - noneiv_mean).detach().cpu().numpy().flatten()
+ noneiv_metrics['true_rmse'] = np.sqrt(np.mean(true_res**2))
# NLL
@@ -208,8 +211,8 @@ def collect_metrics(x_y_pairs, seed=0,
epistemic_coverage(not_averaged_predictions, true_y,
normalize_errors=False, noisy_y=False)
- true_res = true_y - eiv_mean
- eiv_metrics['true_rmse'] = np.sqrt(np.mean(scaled_res**2))
+ true_res = (true_y - eiv_mean).detach().cpu().numpy().flatten()
+ eiv_metrics['true_rmse'] = np.sqrt(np.mean(true_res**2))
# NLL
if scale_outputs:
--
GitLab