Minor fixes

pywatts/board.py

@@ -2,7 +2,7 @@ import tensorflow as tf
 import subprocess
 
 writer = tf.summary.FileWriter("tensorboard")
-checkpoint = tf.train.get_checkpoint_state('tf_pywatts_model')
+checkpoint = tf.train.get_checkpoint_state('tf_pywatts_model_best')
 with tf.Session() as sess:
     saver = tf.train.import_meta_graph(checkpoint.model_checkpoint_path + '.meta')
     saver.restore(sess, checkpoint.model_checkpoint_path)

pywatts/kcross.py

@@ -16,7 +16,7 @@ def split(data, k):
     data_list = data['dc'].tolist()
 
     # Each sample has 337 elements
-    samples = [data_list[i:i+337] for i in range(0, len(data_list) - 337, 20)]
+    samples = [data_list[i:i+337] for i in range(0, len(data_list) - 337, 30)]
     # Randomly shuffle samples
     random.shuffle(samples)
 
@@ -42,13 +42,13 @@ def split(data, k):
     return X_train, y_train, X_eval, y_eval
 
 
-def train(nn, X_train, y_train, X_eval, y_eval, steps=10):
+def train(nn, X_train, y_train, X_eval, y_eval, steps=100):
     """Trains the Network nn using k-cross-validation"""
     evaluation = []
     for count, train_data in enumerate(X_train):
         for i in range(steps):
-            nn.train(train_data, y_train[count], batch_size=30, steps=100) #batch_size=int(len(train_data['dc'])/336), steps=1)
-            evaluation.append(nn.evaluate(X_eval[count], y_eval[count], batch_size=int(len(X_eval[count]['dc'])/336)))
+            nn.train(train_data, y_train[count], batch_size=1000, steps=30) #batch_size=int(len(train_data['dc'])/336), steps=1)
+            evaluation.append(nn.evaluate(X_eval[count], y_eval[count]))
             print("Training %s: %s/%s" % (count, (i+1), steps))
 
     return evaluation

pywatts/neural.py

@@ -19,6 +19,9 @@ def pywatts_input_fn(X, y=None, num_epochs=None, shuffle=True, batch_size=1):
     else:
         dataset = tf.data.Dataset.from_tensor_slices((dict(features), labels))
 
+    if num_epochs is not None:
+        return dataset.batch(len(features['0']))
+
     if shuffle:
         return dataset.shuffle(len(features['0']*len(features)*4)).repeat().batch(batch_size)
     else:
@@ -31,7 +34,7 @@ class Net:
 
     def __init__(self, feature_cols=__feature_cols):
         self.__regressor = tf.estimator.DNNRegressor(feature_columns=feature_cols,
-                                                     hidden_units=[64, 128, 64],
+                                                     hidden_units=[128, 512, 128],
                                                      model_dir='tf_pywatts_model')
 
     def train(self, training_data, training_results, batch_size, steps):

pywatts/routines.py

@@ -66,10 +66,12 @@ def train(nn, X_train, y_train, X_val, y_val, steps=100):
 
 def plot_training(evaluation):
     loss = []
+    steps = []
     for e in evaluation:
-        loss.append(e['average_loss'])
+        loss.append(e['loss'])
+        steps.append(e['global_step'])
 
-    pp.plot(loss)
+    pp.plot(steps, loss)
     # Needed for execution in PyCharm
     pp.show()
 

pywatts/test_kcross_train.py

@@ -8,7 +8,7 @@ K = 10
 NUM_EVAL_STATIONS = 40
 TRAIN = True
 PLOT = True
-TRAIN_STEPS = 20
+TRAIN_STEPS = 10
 
 
 df = pywatts.db.rows_to_df(list(range(1, NUM_STATIONS_FROM_DB)))

pywatts/test_predict24.py

@@ -6,7 +6,7 @@ import matplotlib.pyplot as pp
 
 PREDICT_QUERY = "query-sample_24hour.json"
 PREDICT_RESULT = PREDICT_QUERY.replace("query", "result")
-QUERY_ID = 4
+QUERY_ID = 0
 
 
 pred_query = input_query("../sample_data/" + PREDICT_QUERY, QUERY_ID)