From fa8ee05e088d8131261d7eca72fd3c0e59c080f6 Mon Sep 17 00:00:00 2001
From: Alexey <iamfullofspam@gmail.com>
Date: Thu, 30 Apr 2015 15:45:32 +0300
Subject: [PATCH] minor changes in comments, added TODOs

---
 cmsuml/nnet.py     | 17 ++++++++++-------
 tests/test_nnet.py | 11 ++++++-----
 2 files changed, 16 insertions(+), 12 deletions(-)

diff --git a/cmsuml/nnet.py b/cmsuml/nnet.py
index de03013..c8885a0 100755
--- a/cmsuml/nnet.py
+++ b/cmsuml/nnet.py
@@ -204,6 +204,9 @@ def adadelta_trainer(x, y, w, parameters, derivatives, loss, stages=1000, decay_
 
 
 # TODO think of dropper and noises
+# TODO add scaler
+# TODO yield system
+
 
 def normalize_weight(y, sample_weight):
     """ guarantees that sample weight is numpy.array of shape [n_samples],
@@ -214,7 +217,7 @@ def normalize_weight(y, sample_weight):
 
 
 class AbstractNeuralNetworkClassifier(BaseEstimator, ClassifierMixin):
-    def __init__(self, layers=None, loss=log_loss, trainer='irprop-', trainer_parameters=None, random_state=None):
+    def __init__(self, layers=None, loss='log_loss', trainer='irprop-', trainer_parameters=None, random_state=None):
         """
         Constructs the neural network based on Theano (for classification purposes).
         Supports only binary classification, supports weights, which makes it usable in boosting.
@@ -222,9 +225,9 @@ def __init__(self, layers=None, loss=log_loss, trainer='irprop-', trainer_parame
         Works in sklearn fit-predict way: X is [n_samples, n_features], y is [n_samples], sample_weight is [n_samples].
         Works as usual sklearn classifier, can be used in boosting, for instance, pickled, etc.
         :param layers: list of int, e.g [9, 7] - the number of units in each *hidden* layer
-        :param loss: loss function used (log_loss by default)
-        :param trainer: string, describes the method
-        :param trainer_parameters: parameters passed to trainer function (learning_rate, etc., trainer-specific).
+        :param loss: loss function used (log_loss by default), str ot function(y, pred, w) -> float
+        :param trainer: string, name of optimization method used
+        :param dict trainer_parameters: parameters passed to trainer function (learning_rate, etc., trainer-specific).
         """
         self.layers = layers
         self.loss = loss
@@ -362,7 +365,7 @@ def prepare(self):
         n1, n2, n3 = self.layers_
         W1 = self._create_shared_matrix('W1', n1, n2)
         W2 = self._create_shared_matrix('W2', n2, n3)
-        # this parameter is responsible for scaling, it is computed as well
+        # this parameter is responsible for scaling, it is optimised too
         G = theano.shared(value=0.1, name='G')
         self.parameters['G'] = G
 
@@ -406,7 +409,7 @@ def activation(input):
 
 
 class PairwiseSoftplusNeuralNetwork(AbstractNeuralNetworkClassifier):
-    """The result is computed as h = sigmoid(Ax), output = sum_{ij} B_ij h_i (1 - h_j) """
+    """The result is computed as h = softplus(Ax), output = sum_{ij} B_ij h_i (1 - h_j) """
 
     def prepare(self):
         n1, n2, n3 = self.layers_
@@ -426,7 +429,7 @@ class ObliviousNeuralNetwork(AbstractNeuralNetworkClassifier):
     """ Uses idea of oblivious trees,
      but not strict cuts on features and not rectangular cuts, but linear conditions
      """
-    # TODO: needs pretraining, like oblivious tree first
+    # TODO: needs pretraining (i.e. by oblivious tree) first
     def prepare(self):
         n1, n2, n3 = self.layers_
         W1 = theano.shared(value=self.random_state.normal(size=[n1, n2]).astype(floatX), name='W1')
diff --git a/tests/test_nnet.py b/tests/test_nnet.py
index e6a3c45..afc40d3 100644
--- a/tests/test_nnet.py
+++ b/tests/test_nnet.py
@@ -38,16 +38,17 @@ def test_nnet(n_samples=200, n_features=5, distance=0.5):
     assert 0 == 1
 
 
-test_nnet()
-
 
 def test_oblivious(n_samples=200, n_features=5, distance=0.5):
     # Since oblivious NN is very slow, it is tested separately
-    nn = nnet.ObliviousNeuralNetwork(layers=[n_features, 3, 1], trainer='irprop-')
-    X, y = make_blobs(n_samples=n_samples, n_features=5, centers=[numpy.ones(n_features) * distance,
-                                                                  - numpy.ones(n_features) * distance])
+    nn = nnet.ObliviousNeuralNetwork(layers=[3], trainer='irprop-')
+    X, y = make_blobs(n_samples=n_samples, n_features=5,
+                      centers=[numpy.ones(n_features) * distance, - numpy.ones(n_features) * distance])
     nn.fit(X, y, batch=100, verbose=10)
     print(roc_auc_score(y, nn.predict_proba(X)[:, 1]), nn)
 
 
+test_oblivious()
+test_nnet()
+