def test_regression_normal_log_prob_means_and_stddevs_2d(
self, model_output_type):
tensor_model_outputs = tf.constant([
[[0.3, 0.4, np.log(0.01), np.log(0.02)],
[1.6, 0.6, np.log(2.0), np.log(0.01)]],
[[0.2, 0.2, np.log(0.1), np.log(0.2)],
[0.8, 0.5, np.log(0.5), np.log(0.2)]],
[[0.4, 0.6, np.log(1.0), np.log(1.5)],
[2.4, 0.4, np.log(0.05), np.log(0.1)]],
])
labels = tf.constant([[0.2, 0.4], [1.4, 1.0]])
model_outputs = NewRegressionModelOutputs(
tensor_model_outputs,
model_output_type,
outputs_with_log_stddevs=True)
ens_reg_outputs = stats.RegressionOutputs(
outputs_with_log_stddevs=True)
ens_reg_outputs.update(model_outputs[0])
ens_reg_outputs.update(model_outputs[1])
ens_reg_outputs.update(model_outputs[2])
means, variances = ens_reg_outputs.result()
expected_nll = -tfd.Normal(means, variances**0.5).log_prob(labels)
rnlls = stats.RegressionNormalLogProb(outputs_with_log_stddevs=True)
rnlls.update(model_outputs[0], labels)
rnlls.update(model_outputs[1], labels)
rnlls.update(model_outputs[2], labels)
nlls = rnlls.result()
self.assertAllClose(expected_nll, nlls, atol=TOL)
def test_regression_outputs_only_means_1d(self, model_output_type):
tensor_model_outputs = tf.constant([
[[0.3], [0.6]], # Member 0, Example 0 and 1
[[0.2], [0.5]], # Member 1, Example 0 and 1
[[0.4], [0.4]], # Member 2, Example 0 and 1
])
model_outputs = NewRegressionModelOutputs(tensor_model_outputs,
model_output_type)
ens_reg_outputs = stats.RegressionOutputs()
ens_reg_outputs.update(model_outputs[0])
ens_reg_outputs.update(model_outputs[1])
ens_reg_outputs.update(model_outputs[2])
means, variances = ens_reg_outputs.result()
self.assertAlmostEqual(0.3, float(means[0][0]), delta=TOL)
self.assertAlmostEqual(0.5, float(means[1][0]), delta=TOL)
expected_variance = self._get_mixture_variance(
probs=[1 / 3, 1 / 3, 1 / 3],
means=[0.3, 0.2, 0.4],
stddevs=[1.0, 1.0, 1.0])
self.assertAlmostEqual(float(expected_variance),
float(variances[0][0]),
delta=1e-5)
expected_variance = self._get_mixture_variance(
probs=[1 / 3, 1 / 3, 1 / 3],
means=[0.6, 0.5, 0.4],
stddevs=[1.0, 1.0, 1.0])
self.assertAlmostEqual(float(expected_variance),
float(variances[1][0]),
delta=1e-5)
def test_regression_outputs_only_means_2d_diff_stddev(
self, model_output_type):
tensor_model_outputs = tf.constant([
[[0.3, 0.4], [1.6, 0.6]], # Member 0, Example 0 and 1
[[0.2, 0.2], [0.8, 0.5]], # Member 1, Example 0 and 1
[[0.4, 0.6], [2.4, 0.4]], # Member 2, Example 0 and 1
])
model_outputs = NewRegressionModelOutputs(tensor_model_outputs,
model_output_type,
stddev=0.1)
ens_reg_outputs = stats.RegressionOutputs(stddev=0.1)
ens_reg_outputs.update(model_outputs[0])
ens_reg_outputs.update(model_outputs[1])
ens_reg_outputs.update(model_outputs[2])
means, variances = ens_reg_outputs.result()
self.assertAlmostEqual(0.3, float(means[0][0]), delta=TOL)
self.assertAlmostEqual(0.4, float(means[0][1]), delta=TOL)
self.assertAlmostEqual(1.6, float(means[1][0]), delta=TOL)
self.assertAlmostEqual(0.5, float(means[1][1]), delta=TOL)
# Expected mixture, does not have to use normal distributions
expected_variance = self._get_mixture_variance(
probs=[1 / 3, 1 / 3, 1 / 3],
means=[0.3, 0.2, 0.4],
stddevs=[0.1, 0.1, 0.1])
self.assertAlmostEqual(float(expected_variance),
float(variances[0][0]),
delta=1e-5)
expected_variance = self._get_mixture_variance(
probs=[1 / 3, 1 / 3, 1 / 3],
means=[0.4, 0.2, 0.6],
stddevs=[0.1, 0.1, 0.1])
self.assertAlmostEqual(float(expected_variance),
float(variances[0][1]),
delta=1e-5)
expected_variance = self._get_mixture_variance(
probs=[1 / 3, 1 / 3, 1 / 3],
means=[1.6, 0.8, 2.4],
stddevs=[0.1, 0.1, 0.1])
self.assertAlmostEqual(float(expected_variance),
float(variances[1][0]),
delta=1e-5)
expected_variance = self._get_mixture_variance(
probs=[1 / 3, 1 / 3, 1 / 3],
means=[0.6, 0.5, 0.4],
stddevs=[0.1, 0.1, 0.1])
self.assertAlmostEqual(float(expected_variance),
float(variances[1][1]),
delta=1e-5)
def test_regression_outputs_means_and_variances_2d(self,
model_output_type):
tensor_model_outputs = tf.constant([
[ # member 0 tensor_model_outputs
[0.3, 0.4, np.log(0.01), np.log(0.02)], # Example 0
[1.6, 0.6, np.log(2.0), np.log(0.01)], # Example 1
],
[ # member 1 tensor_model_outputs
[0.2, 0.2, np.log(0.1), np.log(0.2)], # Example 0
[0.8, 0.5, np.log(0.5), np.log(0.2)], # Example 1
],
[ # member 2 tensor_model_outputs
[0.4, 0.6, np.log(1.0), np.log(1.5)], # Example 0
[2.4, 0.4, np.log(0.05), np.log(0.1)], # Example 1
]
])
model_outputs = NewRegressionModelOutputs(
tensor_model_outputs,
model_output_type,
outputs_with_log_stddevs=True)
ens_reg_outputs = stats.RegressionOutputs(
outputs_with_log_stddevs=True)
ens_reg_outputs.update(model_outputs[0]) # Member 0 outputs
ens_reg_outputs.update(model_outputs[1]) # Member 1 outputs
ens_reg_outputs.update(model_outputs[2]) # Member 2 outputs
means, variances = ens_reg_outputs.result()
self.assertAlmostEqual(0.3, float(means[0][0]), delta=TOL)
self.assertAlmostEqual(0.4, float(means[0][1]), delta=TOL)
self.assertAlmostEqual(1.6, float(means[1][0]), delta=TOL)
self.assertAlmostEqual(0.5, float(means[1][1]), delta=TOL)
# Expected mixture, does not have to use normal distributions
expected_variance = self._get_mixture_variance(
probs=[1 / 3, 1 / 3, 1 / 3],
means=[0.3, 0.2, 0.4],
stddevs=[0.01, 0.1, 1.0])
self.assertAlmostEqual(float(expected_variance),
float(variances[0][0]),
delta=1e-5)
expected_variance = self._get_mixture_variance(
probs=[1 / 3, 1 / 3, 1 / 3],
means=[0.4, 0.2, 0.6],
stddevs=[0.02, 0.2, 1.5])
self.assertAlmostEqual(float(expected_variance),
float(variances[0][1]),
delta=1e-5)
expected_variance = self._get_mixture_variance(
probs=[1 / 3, 1 / 3, 1 / 3],
means=[1.6, 0.8, 2.4],
stddevs=[2.0, 0.5, 0.05])
self.assertAlmostEqual(float(expected_variance),
float(variances[1][0]),
delta=1e-5)
expected_variance = self._get_mixture_variance(
probs=[1 / 3, 1 / 3, 1 / 3],
means=[0.6, 0.5, 0.4],
stddevs=[0.01, 0.2, 0.1])
self.assertAlmostEqual(float(expected_variance),
float(variances[1][1]),
delta=1e-5)