PyTorch migration: Remove tensorflow components, add FATE estimators

.travis.yml

@@ -1,6 +1,6 @@
 language: python
 python:
-    - 3.7.9 # Pinned since tensorflow 1.x is not available for python > 3.7
+    - 3.8
 
 cache:
   directories:
@@ -23,18 +23,16 @@ stages:
 jobs:
   fast_finish: true
   include:
-    - python: 3.7.9
+    - python: 3.8
       env: TOXENV=test1
-    - python: 3.7.9
+    - python: 3.8
       env: TOXENV=test2
-    - python: 3.7.9
+    - python: 3.8
       env: TOXENV=test3
-    - python: 3.7.9
-      env: TOXENV=test4
-    - python: 3.7.9
+    - python: 3.8
       env: TOXENV=lint
     - stage: docs
-      python: 3.7.9
+      python: 3.8
       env: TOXENV=docs
 
 before_deploy:
@@ -47,6 +45,6 @@ deploy:
     script: poetry publish -v --build
     on:
       tags: true
-      python: 3.7.9
+      python: 3.8
       repo: kiudee/cs-ranking
       branch: master

HISTORY.rst

@@ -2,7 +2,14 @@
 History
 =======
 
-Unreleased
+2.0.0 (Unreleased)
+------------------
+
+* The library has been migrated to pytorch. This is a breaking change. You will
+  likely need to adapt to this new version if you have been using estimators
+  from version 1.x.
+
+1.3.0 (Unreleased)
 ------------------
 
 * We no longer override any of the defaults of our default optimizer (SGD). In

README.rst

@@ -1,19 +1,29 @@
 |Build Status| |Coverage| |Binder|
 
+****
+NOTE
+****
+
+This repository is currently in the process of a migration from tensorflow to
+PyTorch. You should use the latest released version if you are not interested
+in the partial PyTorch implementation.
+
 *******
 CS-Rank
 *******
+
 CS-Rank is a Python package for context-sensitive ranking and choice
 algorithms.
 
 We implement the following new object ranking/choice architectures:
 
 * FATE (First aggregate then evaluate)
-* FETA (First evaluate then aggregate)
+* FETA (First evaluate then aggregate) (currently not available due to the
+  PyTorch migration)
 
 In addition, we also implement these algorithms for choice functions:
 
-* RankNetChoiceFunction
+* RankNetChoiceFunction (currently not available due to the PyTorch migration)
 * GeneralizedLinearModel
 * PairwiseSVMChoiceFunction
 
@@ -24,12 +34,10 @@ setting:
 * MixedLogitModel
 * NestedLogitModel
 * PairedCombinatorialLogit
-* RankNetDiscreteChoiceFunction
+* RankNetDiscreteChoiceFunction (currently not available due to the PyTorch
+  migration)
 * PairwiseSVMDiscreteChoiceFunction
 
-Check out our `interactive notebooks`_ to quickly find out what our package can
-do.
-
 
 Getting started
 ===============
@@ -73,7 +81,7 @@ Another option is to clone the repository and install CS-Rank using::
 
 Dependencies
 ------------
-CS-Rank depends on Tensorflow, Keras, NumPy, SciPy, matplotlib, scikit-learn,
+CS-Rank depends on PyTorch, skorch, NumPy, SciPy, matplotlib, scikit-learn,
 joblib and tqdm. For data processing and generation you will
 also need PyGMO, H5Py and pandas.
 

csrank/choicefunction/__init__.py

@@ -1,21 +1,11 @@
 from .baseline import AllPositive
-from .cmpnet_choice import CmpNetChoiceFunction
 from .fate_choice import FATEChoiceFunction
-from .fatelinear_choice import FATELinearChoiceFunction
-from .feta_choice import FETAChoiceFunction
-from .fetalinear_choice import FETALinearChoiceFunction
 from .generalized_linear_model import GeneralizedLinearModel
 from .pairwise_choice import PairwiseSVMChoiceFunction
-from .ranknet_choice import RankNetChoiceFunction
 
 __all__ = [
     "AllPositive",
-    "CmpNetChoiceFunction",
     "FATEChoiceFunction",
-    "FATELinearChoiceFunction",
-    "FETAChoiceFunction",
-    "FETALinearChoiceFunction",
     "GeneralizedLinearModel",
     "PairwiseSVMChoiceFunction",
-    "RankNetChoiceFunction",
 ]

csrank/choicefunction/choice_functions.py

@@ -2,8 +2,11 @@
 import logging
 
 import numpy as np
+import skorch
+import torch.nn as nn
 
 from csrank.constants import CHOICE_FUNCTION
+from csrank.learner import SkorchInstanceEstimator
 from csrank.metrics_np import f1_measure
 from csrank.util import progress_bar
 
@@ -69,3 +72,49 @@ def _tune_threshold(self, X_val, Y_val, thin_thresholds=1, verbose=0):
             " a micro F1-measure of {:.2f}".format(threshold, best)
         )
         return threshold
+
+
+class SkorchChoiceFunction(ChoiceFunctions, SkorchInstanceEstimator):
+    """A variable choice estimator based on some scoring module.
+
+    This estimator takes a scoring module and combines it with a sigmoid
+    activation to predict scores between 0 and 1. The choice is then made based
+    on a fixed threshold value. This makes it very simple to derive new
+    estimators with any given scoring function. Refer to skorch's documentation
+    for supported parameters. For example the optimizer or the optimizer's
+    learning rate could be overridden.
+
+    Parameters
+    ----------
+    module : torch module (class)
+        This is the scoring module. It should be an uninstantiated
+        ``torch.nn.Module`` class that expects the number of features per
+        object as its only parameter on initialization.
+
+    criterion : torch criterion (class)
+        The criterion that is used to evaluate and optimize the module.
+
+    threshold : float
+        The threshold value that is used to convert scores to a choice. Must be
+        between 0 and 1. Defaults to 0.5.
+
+    **kwargs : skorch NeuralNet arguments
+        All keyword arguments are passed to the constructor of
+        ``skorch.NeuralNet``. See the documentation of that class for more
+        details.
+    """
+
+    def __init__(self, module, criterion=nn.BCELoss, threshold=0.5, **kwargs):
+        super().__init__(module=module, criterion=criterion, **kwargs)
+        # The scoring is trained to predict something close to "0" for
+        # non-chosen values, something close to "1" for chosen values. So 0.5
+        # is a natural threshold. It would be possible to additionally tune
+        # that threshold.
+        self.threshold_ = threshold
+
+    def initialize_module(self, *args, **kwargs):
+        params = self.get_params_for("module")
+        # Add a Sigmoid activation since the resulting "scores" should be
+        # between 0 and 1.
+        self.module_ = nn.Sequential(self.module(**params), nn.Sigmoid())
+        self.module_ = skorch.utils.to_device(self.module_, self.device)