add links to Sergei's results

Author: levinas

P1B2/README.md

@@ -103,15 +103,29 @@ Best model saved to: model.A=sigmoid.B=64.D=None.E=20.L1=1024.L2=512.L3=256.P=1e
 Evaluation on test data: {'accuracy': 0.5500}
 ```
 
-### Running the XGBoost classifier
+### Preliminary performance
+
+The XGBoost classifier below achieves ~55% average accuracy on
+validation data in the five-fold cross validation experiment. This
+suggests there may be a low ceiling for the MLP results; there may not
+be enough information in this set of SNP data to classify cancer types
+accurately.
 
 ```
 cd P1B2
 python p1b2_xgboost.py
+```
+
+David Anthony Brewster, Jr, Yashasvi Sharma, and Sergei Maslov have
+achived 60.2% accuracy in their experiment. See the following links for details:
+
+* [DOE-NCI Confluence P1B2 scatch pad](https://doe-nci.atlassian.net/wiki/display/PROJECT/P1B2+Networks)
+* [David Brewster's github repo](https://github.com/ljeabmreosn/NeuralNetwork/tree/master/benchmarks/P1B2)
+
+Try the following commands to run a replicate of their network (modified for the P1B2 data loader).
 
 ```
-The XGBoost classifier achieves ~55% average accuracy on validation
-data in the five-fold cross validation experiment. This suggests there
-may be a low ceiling for the MLP results; there may not be enough
-information in this set of SNP data to classify cancer types
-accurately.
+cd P1B2
+python contrib/David_Brewster/p1b2_ljeabmreosn_mlp.py
+```
+

P1B2/contrib/David_Brewster/p1b2_ljeabmreosn_mlp.py

@@ -0,0 +1,110 @@
+from __future__ import print_function
+
+from keras.models import Model, Sequential
+from keras.layers import Dense, Dropout, Input, advanced_activations
+from keras.optimizers import SGD
+from keras.callbacks import EarlyStopping, ModelCheckpoint
+
+from ..datasets import p1b2
+
+import argparse
+
+args = None
+def set_args():
+    global args
+    parser = argparse.ArgumentParser(description='Variate some music.')
+    parser.add_argument('--train', dest='training_file', help='input training data')
+    parser.add_argument('--test', dest='testing_file', help='input testing data')
+    parser.add_argument('--trainvar', dest='training_variable', help='variable you want to train')
+    args = parser.parse_args()
+
+
+BEST_MODEL_PATH = 'benchmarks/P1B2/best.hdf5'
+
+
+(X_train, y_train), (X_test, y_test) = (None, None), (None, None)
+input_dim = None
+output_dim = None
+
+def create_model():
+    # advanced activation not used yet
+    srelu = advanced_activations.SReLU(
+        t_left_init='zero', 
+        a_left_init='glorot_uniform', 
+        t_right_init='glorot_uniform', 
+        a_right_init='one'
+    )
+
+    # create and return model
+    model = Sequential()
+    model.add(Dense(256, input_dim=input_dim, activation='sigmoid'))
+    model.add(Dense(256, activation='sigmoid'))
+    model.add(Dense(output_dim, activation='sigmoid'))
+    return model
+
+def train_model(model):
+    opt = 'rmsprop'
+    model_checkpoint = ModelCheckpoint(
+        filepath=BEST_MODEL_PATH, 
+        monitor='val_acc', 
+        verbose=0, 
+        save_best_only=True,  
+        mode='auto'
+    )
+    overfitting_stopper = EarlyStopping(
+        monitor='val_acc', 
+        min_delta=0, 
+        patience=5, 
+        verbose=1, 
+        mode='auto'
+    )
+    model.compile(loss='categorical_crossentropy', optimizer=opt, metrics=['accuracy'])
+    model_history = model.fit(X_train, y_train,
+	    batch_size       = 64,
+	    nb_epoch         = 50,
+	    #shuffle          = True,
+	    validation_split = 0.2,
+	    #verbose          = 2
+        callbacks        = [overfitting_stopper, model_checkpoint]
+    )
+
+def load_model_from_path(weights_path):
+    model = create_model()
+    model.load_weights(weights_path)
+    return model
+
+def save_weights(model):
+    for i, layer in enumerate(model.layers):
+        weights = layer.get_weights()
+        print('LAYER {}'.format(i))
+        print(weights)
+        print(weights.s)
+        input()
+
+
+
+
+##################### START OF PROGRAM ##############################
+set_args()
+# gather input
+if args != None and args.training_file and args.testing_file and args.training_variable:
+    try:
+        (X_train, y_train), (X_test, y_test) = p1b2.load_data_from_file(
+            train=args.training_file, 
+            test=args.testing_file, 
+            trainvar=args.training_variable)
+    except Exception as e:
+        print(e)
+        print('(Hint: Are you sure the input files are in valid format?)')
+else:
+    (X_train, y_train), (X_test, y_test) = p1b2.load_data_from_url(n_cols=10000)
+
+input_dim = X_train.shape[1]
+output_dim = y_train.shape[1]
+
+# train our model
+train_model(create_model())
+best_model = load_model_from_path(BEST_MODEL_PATH)
+y_pred = best_model.predict(X_test)
+accuracy = p1b2.evaluate(y_test, y_pred)
+