Merge pull request #152 from cnellington/dev

Release v0.2.4

Author: cnellington

contextualized/easy/ContextualizedClassifier.py

@@ -6,6 +6,7 @@
 
 from contextualized.functions import LINK_FUNCTIONS
 from contextualized.easy import ContextualizedRegressor
+from contextualized.regression import LOSSES
 
 
 class ContextualizedClassifier(ContextualizedRegressor):
@@ -15,6 +16,7 @@ class ContextualizedClassifier(ContextualizedRegressor):
 
     def __init__(self, **kwargs):
         kwargs["link_fn"] = LINK_FUNCTIONS["logistic"]
+        kwargs["loss_fn"] = LOSSES["bceloss"]
         super().__init__(**kwargs)
 
     def predict(self, C, X, individual_preds=False, **kwargs):

contextualized/easy/ContextualizedNetworks.py

@@ -80,9 +80,9 @@ def measure_mses(self, C, X, individual_preds=False):
         mses = np.zeros((len(betas), len(C)))  # n_bootstraps x n_samples
         for i in range(X.shape[-1]):
             for j in range(X.shape[-1]):
-                residuals = np.tile(X[:, i], len(betas)) - (
-                    betas[:, :, i, j] * np.tile(X[:, j], len(betas)) + mus[:, :, i, j]
-                )
+                tiled_xi = np.array([X[:, i] for _ in range(len(betas))])
+                tiled_xj = np.array([X[:, j] for _ in range(len(betas))])
+                residuals = tiled_xi - betas[:, :, i, j]*tiled_xj + mus[:, :, i ,j]
                 mses += residuals**2 / (X.shape[-1] ** 2)
         if not individual_preds:
             mses = np.mean(mses, axis=0)
@@ -164,7 +164,7 @@ def __init__(self, **kwargs):
             **kwargs,
         )
 
-    def predict_params(self, C, individual_preds=False, **kwargs):
+    def predict_params(self, C, **kwargs):
         """
 
         :param C:
@@ -173,8 +173,12 @@ def predict_params(self, C, individual_preds=False, **kwargs):
         """
         # Returns betas
         # TODO: No mus for NOTMAD at present.
-        return super().predict_params(C, individual_preds,
-            model_includes_mus=False, **kwargs)
+        return super().predict_params(
+            C,
+            individual_preds=kwargs.get("individual_preds", False),
+            model_includes_mus=False,
+            uses_y=False,
+            project_to_dag=kwargs.get("project_to_dag", True))
 
     def predict_networks(self, C, with_offsets=False, project_to_dag=True, **kwargs):
         """

contextualized/regression/__init__.py

@@ -7,7 +7,7 @@
     MultitaskMultivariateDataset,
     MultitaskUnivariateDataset,
 )
-from contextualized.regression.losses import MSE
+from contextualized.regression.losses import MSE, BCELoss
 from contextualized.regression.regularizers import REGULARIZERS
 from contextualized.regression.lightning_modules import (
     NaiveContextualizedRegression,
@@ -25,7 +25,7 @@
     "multitask_multivariate": MultitaskMultivariateDataset,
     "multitask_univariate": MultitaskUnivariateDataset,
 }
-LOSSES = {"mse": MSE}
+LOSSES = {"mse": MSE, "bceloss": BCELoss}
 MODELS = ["multivariate", "univariate"]
 METAMODELS = ["simple", "subtype", "multitask", "tasksplit"]
 TRAINERS = {"regression_trainer": RegressionTrainer}

contextualized/regression/losses.py

@@ -2,6 +2,7 @@
 Losses used in regression.
 """
 
+import torch
 
 def MSE(Y_true, Y_pred):
     """
@@ -20,3 +21,8 @@ def MSE(Y_true, Y_pred):
     """
     residual = Y_true - Y_pred
     return residual.pow(2).mean()
+
+
+def BCELoss(Y_true, Y_pred):
+    loss = -(Y_true * torch.log(Y_pred + 1e-8) + (1 - Y_true) * torch.log(1 - Y_pred + 1e-8))
+    return loss.mean()

setup.py

@@ -5,7 +5,7 @@
 from setuptools import find_packages, setup
 
 DESCRIPTION = "An ML toolbox for estimating context-specific parameters."
-VERSION = '0.2.3'
+VERSION = '0.2.4'
 
 setup(
     name='contextualized',