tested conservatism

.github/workflows/push.yml

@@ -39,3 +39,6 @@ jobs:
         poetry run pylint pyrfd
         poetry run flake8 pyrfd
         poetry run black --check pyrfd
+      
+    - name: Run tests
+      run: poetry run pytest tests

pyrfd/covariance.py

@@ -10,7 +10,6 @@
 import pandas as pd
 import numpy as np
 import matplotlib.pyplot as plt
-import torch
 from scipy import stats
 from tqdm import tqdm
 
@@ -46,22 +45,21 @@ def __init__(
         gradient_var: Tuple[float, float] | None = None,
         dims=None,
     ) -> None:
+
         self.mean = mean
+        self.dims = dims
 
-        if variance is None:
-            self.var_reg = None
-        else:
+        self.var_reg = None
+        self.g_var_reg = None
+
+        if variance is not None:
             self.var_reg = ScalarRegression(*variance)
             assert self.var_reg.is_plausible_variance_regression
 
-        if gradient_var is None:
-            self.g_var_reg = None
-        else:
+        if gradient_var is not None:
             self.g_var_reg = ScalarRegression(*gradient_var)
             assert self.g_var_reg.is_plausible_variance_regression
 
-        self.dims = dims
-
         self._fitted = False
         if self._is_fitted():
             self._fitted = True
@@ -173,7 +171,7 @@ def learning_rate(self, loss, grad_norm, *, b_size_inv=0, conservatism=0):
             def diff_cost(stepsize):
                 cond_var = self.cond_variance(stepsize, b_size_inv=b_size_inv)
                 d_cond_var = self.diff_cond_variance(stepsize, b_size_inv=b_size_inv)
-                reg = regularization * 0.5 * d_cond_var / torch.sqrt(cond_var)
+                reg = regularization * 0.5 * d_cond_var / np.sqrt(cond_var)
 
                 dce = self.diff_cond_expectation(stepsize, loss, grad_norm, b_size_inv)
                 return dce + reg
@@ -407,7 +405,7 @@ def scale(self):
         )
 
     def kernel(self, neg_sq_half):
-        return self.variance * torch.exp(neg_sq_half / (self.scale**2))
+        return self.variance * np.exp(neg_sq_half / (self.scale**2))
 
     def diff_kernel(self, neg_sq_half):
         return self.kernel(neg_sq_half) / (self.scale**2)
@@ -434,7 +432,7 @@ def learning_rate(self, loss, grad_norm, *, b_size_inv=0, conservatism=0):
         return (
             var_g_adjust
             * (self.scale**2)
-            / (torch.sqrt(tmp**2 + (self.scale * grad_norm * var_g_adjust) ** 2) + tmp)
+            / (np.sqrt(tmp**2 + (self.scale * grad_norm * var_g_adjust) ** 2) + tmp)
         )
 
 

pyrfd/regression.py

@@ -26,7 +26,7 @@ def __call__(self, x):
     @property
     def is_plausible_variance_regression(self):
         """ variances are positive, bool if slope and intercept are postivie """
-        return (self.slope > 0) and (self.intercept > 0)
+        return (self.slope >= 0) and (self.intercept > 0)
 
     @property
     def slope(self):

tests/covariance_test.py

@@ -1,7 +1,11 @@
 import pytest
 from pyrfd.covariance import SquaredExponential
 
-@pytest.mark.parametrize("mean,var,scale", [(1,1,2)])
+@pytest.mark.parametrize("mean,var,scale", [
+    (1,1,2),
+    (0.1,3,10),
+    (10,3,0.5),
+])
 def test_squared_exponential(mean, var, scale):
     cov = SquaredExponential(
         mean=mean,