added tests for construction of neural nets

mhcflurry/feedforward.py

@@ -27,28 +27,6 @@ import theano
 theano.config.exception_verbosity = 'high'
 
 
-def compile_forward_predictor(model, theano_mode=None):
-    """
-    In cases where we want to get predictions from a model that hasn't
-    been compiled (to avoid overhead of compiling training code),
-    use this helper to only compile the subset of Theano needed for
-    forward-propagation/predictions.
-    """
-
-    model.X_test = model.get_input(train=False)
-    model.y_test = model.get_output(train=False)
-    if type(model.X_test) == list:
-        predict_ins = model.X_test
-    else:
-        predict_ins = [model.X_test]
-
-    model._predict = theano.function(
-        predict_ins,
-        model.y_test,
-        allow_input_downcast=True,
-        mode=theano_mode)
-
-
 def make_network(
         input_size,
         embedding_input_dim=None,
@@ -61,8 +39,7 @@ def make_network(
         dropout_probability=0.0,
         model=None,
         optimizer=None,
-        learning_rate=0.001,
-        compile_for_training=True):
+        learning_rate=0.001):
 
     if model is None:
         model = Sequential()
@@ -112,16 +89,13 @@ def make_network(
         output_dim=1,
         init=init))
     model.add(Activation(output_activation))
-    if compile_for_training:
-        model.compile(loss=loss, optimizer=optimizer)
-    else:
-        compile_forward_predictor(model)
+    model.compile(loss=loss, optimizer=optimizer)
     return model
 
 
 def make_hotshot_network(
         peptide_length=9,
-        layer_sizes=[500],
+        layer_sizes=[100],
         activation="relu",
         init="lecun_uniform",
         loss="mse",
@@ -146,7 +120,7 @@ def make_embedding_network(
         peptide_length=9,
         embedding_input_dim=20,
         embedding_output_dim=20,
-        layer_sizes=[500],
+        layer_sizes=[100],
         activation="relu",
         init="lecun_uniform",
         loss="mse",

mhcflurry/imputation.py

@@ -25,7 +25,6 @@ from .data import (
     create_allele_data_from_peptide_to_ic50_dict,
 )
 
-
 def prune_dense_matrix_and_labels(
         X,
         peptide_list,

test/test_neural_nets.py

+from mhcflurry.feedforward import (
+    make_embedding_network,
+    make_hotshot_network,
+)
+import numpy as np
+
+def test_make_embedding_network():
+    nn = make_embedding_network(
+        peptide_length=3,
+        layer_sizes=[3],
+        activation="tanh",
+        embedding_input_dim=3,
+        embedding_output_dim=20,
+        learning_rate=0.1)
+
+    X_negative = np.array([
+        [0] * 3,
+        [1] * 3,
+        [1, 0, 0],
+        [1, 1, 0],
+        [0, 1, 0],
+        [0, 0, 1],
+        [1, 0, 1],
+    ])
+    X_positive = np.array([
+        [0, 2, 0],
+        [1, 2, 0],
+        [1, 2, 1],
+        [0, 2, 1],
+        [2, 2, 0],
+        [2, 2, 1],
+        [2, 2, 2],
+    ])
+    X_index = np.vstack([X_negative, X_positive])
+    Y = np.array([0.0] * len(X_negative) + [1.0] * len(X_positive))
+    nn.fit(X_index, Y, nb_epoch=10)
+    Y_pred = nn.predict(X_index)
+    print(Y)
+    print(Y_pred)
+    for (Y_i, Y_pred_i) in zip(Y, Y_pred):
+        assert abs(Y_i - Y_pred_i) <= 0.25, (Y_i, Y_pred_i)
+
+def test_make_hotshot_network():
+    nn = make_hotshot_network(
+        peptide_length=3,
+        activation="relu",
+        layer_sizes=[4],
+        n_amino_acids=2,
+        learning_rate=0.1)
+    X_binary = np.array([
+        [True, False, True, False, True, False],
+        [True, False, True, False, False, True],
+        [True, False, False, True, True, False],
+        [True, False, False, True, False, True],
+        [False, True, True, False, True, False],
+        [False, True, True, False, False, True],
+    ], dtype=bool)
+    Y = np.array([0.0, 0.0, 0.0, 0.0, 1.0, 1.0])
+    nn.fit(X_binary, Y, nb_epoch=10)
+    Y_pred = nn.predict(X_binary)
+    print(Y)
+    print(Y_pred)
+    for (Y_i, Y_pred_i) in zip(Y, Y_pred):
+        assert abs(Y_i - Y_pred_i) <= 0.25, (Y_i, Y_pred_i)