fix: tf.mul -> tf.multiply, feat: allow NaN's in y_train

concise/concise.py

@@ -285,7 +285,6 @@ def _get_var_initialization(self, graph, X_feat_train, y_train):
 
         return var
 
-    # DONE
     def _build_graph(self, graph, var):
         with graph.as_default():
             tf_X_seq = tf.placeholder(tf.float32, shape=[None, 1, None, self._num_channels])
@@ -348,7 +347,16 @@ def model(data, tf_X_feat, var):
             y_pred = model(tf_X_seq, tf_X_feat, var)
             # TODO: enable other loss functions like softmax, huber loss etc
             #   - how to include loss-paramteres like k in the case of huber_loss
-            loss = tf.reduce_mean(tf.square(y_pred - tf_y))
+
+            # allow for NA's in Y
+            # - use tf.is_nan(tf_y) - https://www.tensorflow.org/api_docs/python/tf/is_nan
+            # - sum only those values that are none NA
+            #    - https://www.tensorflow.org/api_docs/python/tf/where
+            y_diff = tf.where(tf.is_nan(tf_y),
+                              x=tf.zeros_like(y_pred),
+                              y=tf_y - y_pred,
+                              name="y_diff")
+            loss = tf.reduce_mean(tf.square(y_diff))
 
             # add regularization
             # regularization = motif_lamb * tf.nn.l2_loss(motif_base_weights) +
@@ -394,7 +402,7 @@ def model(data, tf_X_feat, var):
             #
             # http://www.subsubroutine.com/sub-subroutine/2016/11/12/painting-like-van-gogh-with-convolutional-neural-networks
 
-            init = tf.global_variables_initializer() #tf.initialize_all_variables()
+            init = tf.global_variables_initializer()  # tf.initialize_all_variables()
 
             other_var = {
                 "tf_X_feat": tf_X_feat,
@@ -727,7 +735,6 @@ def _train_lbfgs(self, X_feat_train, X_seq_train, y_train,
                           (best_performance_epoch, best_performance))
                     break
 
-
             # get the test accuracies
             train_accuracy_final = self._accuracy_in_session(sess, other_var,
                                                              X_feat_train, X_seq_train, y_train)
@@ -1083,11 +1090,10 @@ def from_dict(cls, obj_dict):
 
         weights = obj_dict["output"]["weights"]
 
-
         if weights is not None:
             # fix the dimentionality of X_feat in case it was 0 dimentional
             if weights["feature_weights"].shape == (0,):
-                weights["feature_weights"].shape = (0 , obj_dict["param"]["num_tasks"])
+                weights["feature_weights"].shape = (0, obj_dict["param"]["num_tasks"])
             dc._set_var_res(weights)
 
         return dc
@@ -1252,7 +1258,7 @@ def train(self, X_feat, X_seq, y, id_vec=None, n_folds=10, use_stored_folds=None
             # overwrite n_epochs with the best average number of best epochs
             dc._param["n_epochs"] = int(np.array(best_val_acc_epoch_l).mean())
             print("tranining global model with n_epochs = " + str(dc._param["n_epochs"]))
-            
+
             dc.train(X_feat, X_seq, y,
                      n_cores=n_cores
                      )

concise/math_helper.py

@@ -2,8 +2,11 @@
 # mean-squared error
 import numpy as np
 
+# ignore NA-values
 def mse(x, y):
-    return ((x - y) ** 2).mean(axis=None)
+    # equivalent to TF representation
+    y_diff = np.where(np.isnan(y) | np.isnan(x), 0, x - y)
+    return ((y_diff) ** 2).mean(axis=None)
 
 # exponentiated root-mean-squared error
 def ermse(x, y):

concise/tf_helper.py

@@ -18,7 +18,7 @@ def l1_loss(tensor, weight=1.0, scope=None):
         weight = tf.convert_to_tensor(weight,
                                       dtype=tensor.dtype.base_dtype,
                                       name='loss_weight')
-        loss = tf.mul(weight, tf.reduce_sum(tf.abs(tensor)), name='value')
+        loss = tf.multiply(weight, tf.reduce_sum(tf.abs(tensor)), name='value')
         return loss
 
 def huber_loss(tensor, k=1, scope=None):

tests/test_na_y.py

@@ -0,0 +1,36 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+"""
+test when some values are non-NA
+----------------------------------
+"""
+import pytest
+import os
+import numpy as np
+
+from concise import concise
+from concise import math_helper
+from concise import helper
+from tests.setup_concise_load_data import load_example_data
+
+# replace some Y values with NA
+def data_w_na():
+    data = load_example_data(num_tasks=3)
+    param, X_feat, X_seq, y, id_vec = data
+    y[0:51, 0] = np.NaN
+    y[51:101, 1] = np.NaN
+    y[102:300, 1] = np.NaN
+    data = (param, X_feat, X_seq, y, id_vec)
+    return data
+
+def test_not_na():
+    param, X_feat, X_seq, y, id_vec = data_w_na()
+
+    c = concise.Concise(**param)
+    c.train(X_feat, X_seq, y,
+            X_feat_valid=X_feat, X_seq_valid=X_seq, y_valid=y)
+
+    # Not a single one can be Nan
+    assert not np.any(np.isnan(c.get_accuracy()["loss_history"]))
+    assert not np.any(np.isnan(c.get_accuracy()["train_acc_history"]))
+    assert not np.any(np.isnan(c.get_accuracy()["val_acc_history"]))