Merge pull request #45 from jfkirk/wmrb-fix

WMRB on Serial Predictions

tensorrec/loss_graphs.py

@@ -14,23 +14,36 @@ class AbstractLossGraph(object):
     is_sampled_with_replacement = False
 
     @abc.abstractmethod
-    def connect_loss_graph(self, tf_prediction_serial, tf_interactions_serial, tf_prediction, tf_interactions,
-                           tf_rankings, tf_sample_predictions):
+    def connect_loss_graph(self, tf_prediction_serial, tf_interactions_serial, tf_interactions, tf_n_users, tf_n_items,
+                           tf_prediction, tf_rankings, tf_sample_predictions, tf_n_sampled_items):
         """
         This method is responsible for consuming a number of possible nodes from the graph and calculating loss from
         those nodes.
+
+        The following parameters are always passed in
         :param tf_prediction_serial: tf.Tensor
         The recommendation scores as a Tensor of shape [n_samples, 1]
         :param tf_interactions_serial: tf.Tensor
         The sample interactions corresponding to tf_prediction_serial as a Tensor of shape [n_samples, 1]
-        :param tf_prediction: tf.Tensor
-        The recommendation scores as a Tensor of shape [n_users, n_items]
         :param tf_interactions: tf.SparseTensor
         The sample interactions as a SparseTensor of shape [n_users, n_items]
+        :param tf_n_users: tf.placeholder
+        The number of users in tf_interactions
+        :param tf_n_items: tf.placeholder
+        The number of items in tf_interactions
+
+        The following parameters are passed in if is_dense is True
+        :param tf_prediction: tf.Tensor
+        The recommendation scores as a Tensor of shape [n_users, n_items]
         :param tf_rankings: tf.Tensor
         The item ranks as a Tensor of shape [n_users, n_items]
+
+        The following parameters are passed in if is_sample_based is True
         :param tf_sample_predictions: tf.Tensor
         The recommendation scores of a sample of items of shape [n_users, n_sampled_items]
+        :param tf_n_sampled_items: tf.placeholder
+        The number of items per user in tf_sample_predictions
+
         :return: tf.Tensor
         The loss value.
         """
@@ -124,30 +137,34 @@ class WMRBLossGraph(AbstractLossGraph):
     Approximation of http://ceur-ws.org/Vol-1905/recsys2017_poster3.pdf
     Interactions can be any positive values, but magnitude is ignored. Negative interactions are also ignored.
     """
-    is_dense = True
     is_sample_based = True
 
-    def connect_loss_graph(self, tf_prediction, tf_interactions, tf_sample_predictions, **kwargs):
+    def connect_loss_graph(self, tf_prediction_serial, tf_interactions, tf_sample_predictions, tf_n_items,
+                           tf_n_sampled_items, **kwargs):
 
         # WMRB expects [-1, 1] bounded predictions
-        bounded_prediction = tf.nn.tanh(tf_prediction)
+        bounded_prediction = tf.nn.tanh(tf_prediction_serial)
         bounded_sample_prediction = tf.nn.tanh(tf_sample_predictions)
 
-        return self.weighted_margin_rank_batch(tf_prediction=bounded_prediction,
+        return self.weighted_margin_rank_batch(tf_prediction_serial=bounded_prediction,
                                                tf_interactions=tf_interactions,
-                                               tf_sample_predictions=bounded_sample_prediction)
+                                               tf_sample_predictions=bounded_sample_prediction,
+                                               tf_n_items=tf_n_items,
+                                               tf_n_sampled_items=tf_n_sampled_items)
 
     @classmethod
-    def weighted_margin_rank_batch(cls, tf_prediction, tf_interactions, tf_sample_predictions):
+    def weighted_margin_rank_batch(cls, tf_prediction_serial, tf_interactions, tf_sample_predictions, tf_n_items,
+                                   tf_n_sampled_items):
         positive_interaction_mask = tf.greater(tf_interactions.values, 0.0)
         positive_interaction_indices = tf.boolean_mask(tf_interactions.indices,
                                                        positive_interaction_mask)
 
         # [ n_positive_interactions ]
-        positive_predictions = tf.gather_nd(tf_prediction, indices=positive_interaction_indices)
+        positive_predictions = tf.boolean_mask(tf_prediction_serial,
+                                               positive_interaction_mask)
 
-        n_items = tf.cast(tf.shape(tf_prediction)[1], dtype=tf.float32)
-        n_sampled_items = tf.cast(tf.shape(tf_sample_predictions)[1], dtype=tf.float32)
+        n_items = tf.cast(tf_n_items, dtype=tf.float32)
+        n_sampled_items = tf.cast(tf_n_sampled_items, dtype=tf.float32)
 
         # [ n_positive_interactions, n_sampled_items ]
         mapped_predictions_sample_per_interaction = tf.gather(params=tf_sample_predictions,

tensorrec/recommendation_graphs.py

@@ -57,22 +57,17 @@ def bias_prediction_serial(tf_prediction_serial, tf_projected_user_biases, tf_pr
     return tf_prediction_serial + gathered_user_biases + gathered_item_biases
 
 
-def gather_sampled_item_predictions(tf_prediction, tf_sampled_item_indices):
+def densify_sampled_item_predictions(tf_sample_predictions_serial, tf_n_sampled_items, tf_n_users):
     """
-    Gathers the predictions for the given sampled items.
-    :param tf_prediction:
-    :param tf_sampled_item_indices:
+    Turns the serial predictions of the sample items in to a dense matrix of shape [ n_users, n_sampled_items ]
+    :param tf_sample_predictions_serial:
+    :param tf_n_sampled_items:
+    :param tf_n_users:
     :return:
     """
-    prediction_shape = tf.shape(tf_prediction)
-    flattened_prediction = tf.reshape(tf_prediction, shape=[prediction_shape[0] * prediction_shape[1]])
-
-    indices_shape = tf.shape(tf_sampled_item_indices)
-    flattened_indices = tf.reshape(tf_sampled_item_indices, shape=[indices_shape[0] * indices_shape[1]])
-
-    gathered_predictions = tf.gather(params=flattened_prediction, indices=flattened_indices)
-    reshaped_gathered_predictions = tf.reshape(gathered_predictions, shape=indices_shape)
-    return reshaped_gathered_predictions
+    densified_shape = tf.stack([tf_n_users, tf_n_sampled_items])
+    densified_predictions = tf.reshape(tf_sample_predictions_serial, shape=densified_shape)
+    return densified_predictions
 
 
 def rank_predictions(tf_prediction):

tensorrec/tensorrec.py

@@ -11,7 +11,7 @@
     EuclidianSimilarityPredictionGraph
 )
 from .recommendation_graphs import (project_biases, split_sparse_tensor_indices, bias_prediction_dense,
-                                    bias_prediction_serial, rank_predictions, gather_sampled_item_predictions)
+                                    bias_prediction_serial, rank_predictions, densify_sampled_item_predictions)
 from .representation_graphs import AbstractRepresentationGraph, LinearRepresentationGraph
 from .session_management import get_session
 from .util import sample_items, calculate_batched_alpha
@@ -85,7 +85,7 @@ def __init__(self, n_components=100,
             # Feed placeholders
             'tf_n_users', 'tf_n_items', 'tf_user_feature_indices', 'tf_user_feature_values', 'tf_item_feature_indices',
             'tf_item_feature_values', 'tf_interaction_indices', 'tf_interaction_values', 'tf_learning_rate', 'tf_alpha',
-            'tf_sampled_item_indices'
+            'tf_sample_indices', 'tf_n_sampled_items'
         ]
         self.graph_operation_hook_attr_names = [
 
@@ -245,15 +245,19 @@ def _build_tf_graph(self, n_user_features, n_item_features):
         # Initialize placeholder values for inputs
         self.tf_n_users = tf.placeholder('int64')
         self.tf_n_items = tf.placeholder('int64')
+        self.tf_n_sampled_items = tf.placeholder('int64')
+
+        # SparseTensor placeholders
         self.tf_user_feature_indices = tf.placeholder('int64', [None, 2])
         self.tf_user_feature_values = tf.placeholder('float', [None])
         self.tf_item_feature_indices = tf.placeholder('int64', [None, 2])
         self.tf_item_feature_values = tf.placeholder('float', [None])
         self.tf_interaction_indices = tf.placeholder('int64', [None, 2])
         self.tf_interaction_values = tf.placeholder('float', [None])
+
+        self.tf_sample_indices = tf.placeholder('int64', [None, None])
         self.tf_learning_rate = tf.placeholder('float', None)
         self.tf_alpha = tf.placeholder('float', None)
-        self.tf_sampled_item_indices = tf.placeholder('int64', [None, None])
 
         # Construct the features and interactions as sparse matrices
         tf_user_features = tf.SparseTensor(self.tf_user_feature_indices, self.tf_user_feature_values,
@@ -294,6 +298,16 @@ def _build_tf_graph(self, n_user_features, n_item_features):
             tf_x_item=tf_x_item,
         )
 
+        tf_transposed_sample_indices = tf.transpose(self.tf_sample_indices)
+        tf_x_user_sample = tf_transposed_sample_indices[0]
+        tf_x_item_sample = tf_transposed_sample_indices[1]
+        tf_sample_predictions_serial = self.prediction_graph_factory.connect_serial_prediction_graph(
+            tf_user_representation=self.tf_user_representation,
+            tf_item_representation=self.tf_item_representation,
+            tf_x_user=tf_x_user_sample,
+            tf_x_item=tf_x_item_sample,
+        )
+
         # Add biases, if this is a biased estimator
         if self.biased:
             tf_user_feature_biases, tf_projected_user_biases = project_biases(
@@ -316,6 +330,12 @@ def _build_tf_graph(self, n_user_features, n_item_features):
                                                                tf_x_user=tf_x_user,
                                                                tf_x_item=tf_x_item)
 
+            tf_sample_predictions_serial = bias_prediction_serial(tf_prediction_serial=tf_sample_predictions_serial,
+                                                                  tf_projected_user_biases=tf_projected_user_biases,
+                                                                  tf_projected_item_biases=tf_projected_item_biases,
+                                                                  tf_x_user=tf_x_user_sample,
+                                                                  tf_x_item=tf_x_item_sample)
+
         tf_interactions_serial = tf_interactions.values
 
         # Construct API nodes
@@ -339,18 +359,23 @@ def _build_tf_graph(self, n_user_features, n_item_features):
         loss_graph_kwargs = {
             'tf_prediction_serial': self.tf_prediction_serial,
             'tf_interactions_serial': tf_interactions_serial,
+            'tf_interactions': tf_interactions,
+            'tf_n_users': self.tf_n_users,
+            'tf_n_items': self.tf_n_items,
         }
         if self.loss_graph_factory.is_dense:
             loss_graph_kwargs.update({
                 'tf_prediction': self.tf_prediction,
-                'tf_interactions': tf_interactions,
                 'tf_rankings': self.tf_rankings,
             })
         if self.loss_graph_factory.is_sample_based:
-            tf_sample_predictions = gather_sampled_item_predictions(
-                tf_prediction=self.tf_prediction, tf_sampled_item_indices=self.tf_sampled_item_indices
+            tf_sample_predictions = densify_sampled_item_predictions(
+                tf_sample_predictions_serial=tf_sample_predictions_serial,
+                tf_n_sampled_items=self.tf_n_sampled_items,
+                tf_n_users=self.tf_n_users,
             )
-            loss_graph_kwargs.update({'tf_sample_predictions': tf_sample_predictions})
+            loss_graph_kwargs.update({'tf_sample_predictions': tf_sample_predictions,
+                                      'tf_n_sampled_items': self.tf_n_sampled_items})
 
         # Build loss graph
         self.tf_basic_loss = self.loss_graph_factory.connect_loss_graph(**loss_graph_kwargs)
@@ -475,11 +500,12 @@ def fit_partial(self, interactions, user_features, item_features, epochs=1, lear
 
                 # Handle random item sampling, if applicable
                 if self.loss_graph_factory.is_sample_based:
-                    sampled_item_indices = sample_items(n_users=feed_dict[self.tf_n_users],
-                                                        n_items=feed_dict[self.tf_n_items],
-                                                        n_sampled_items=n_sampled_items,
-                                                        replace=self.loss_graph_factory.is_sampled_with_replacement)
-                    feed_dict[self.tf_sampled_item_indices] = sampled_item_indices
+                    sample_indices = sample_items(n_users=feed_dict[self.tf_n_users],
+                                                  n_items=feed_dict[self.tf_n_items],
+                                                  n_sampled_items=n_sampled_items,
+                                                  replace=self.loss_graph_factory.is_sampled_with_replacement)
+                    feed_dict[self.tf_sample_indices] = sample_indices
+                    feed_dict[self.tf_n_sampled_items] = n_sampled_items
 
                 # TODO find something more elegant than these cascaded ifs
                 if not verbose:

tensorrec/util.py

@@ -5,10 +5,14 @@
 
 
 def sample_items(n_items, n_users, n_sampled_items, replace):
-    return np.array([
-        np.random.choice(a=n_items, size=n_sampled_items, replace=replace) + (user_count * n_users)
-        for user_count in range(n_users)
-    ])
+    items_per_user = [np.random.choice(a=n_items, size=n_sampled_items, replace=replace) for _ in range(n_users)]
+
+    sample_indices = []
+    for user, users_items in enumerate(items_per_user):
+        for item in users_items:
+            sample_indices.append((user, item))
+
+    return sample_indices
 
 
 def calculate_batched_alpha(num_batches, alpha):

test/test_recommendation_graphs.py

@@ -5,7 +5,7 @@
 
 from tensorrec.recommendation_graphs import (
     project_biases, split_sparse_tensor_indices, bias_prediction_dense, bias_prediction_serial,
-    gather_sampled_item_predictions, rank_predictions
+    densify_sampled_item_predictions, rank_predictions
 )
 from tensorrec.session_management import get_session
 
@@ -105,25 +105,18 @@ def test_bias_prediction_serial(self):
 
         self.assertTrue((biased_predictions == expected_biased_predictions).all())
 
-    def test_gather_sampled_item_predictions(self):
-        input_data = np.array([
-            [1, 2, 3, 4],
-            [5, 6, 7, 8],
-            [9, 10, 11, 12]
-        ])
-        sample_indices = np.array([
-            [0, 3],  # Corresponds to [1, 4]
-            [5, 6],  # Corresponds to [6, 7]
-            [8, 8],  # Corresponds to [9, 9]
-        ])
-        result = gather_sampled_item_predictions(
-            tf_prediction=tf.identity(input_data), tf_sampled_item_indices=tf.identity(sample_indices)
+    def test_densify_sampled_item_predictions(self):
+        input_data = np.array([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12])
+        result = densify_sampled_item_predictions(
+            tf_sample_predictions_serial=input_data,
+            tf_n_sampled_items=4,
+            tf_n_users=3,
         ).eval(session=self.session)
 
         expected_result = np.array([
-            [1, 4],
-            [6, 7],
-            [9, 9],
+            [1, 2, 3, 4],
+            [5, 6, 7, 8],
+            [9, 10, 11, 12]
         ])
         self.assertTrue((result == expected_result).all())