+7% accuracy (Philipps idea)

src/DEBS2018/training/train_tf.py

@@ -40,12 +40,13 @@ def build_model_columns():
   #numeric column because of float value in certain range
   #COURSE is the key, necessary to map attribute to feature column
   course = tf.feature_column.numeric_column('COURSE')
+  shiptype = tf.feature_column.numeric_column('SHIPTYPE')
   #see above
   lon = tf.feature_column.numeric_column('LON')
   lat = tf.feature_column.numeric_column('LAT')
   
   #combine to single variable
-  base_columns = [course, lon, lat]
+  base_columns = [lon, lat, course, shiptype]
   return base_columns
 
 #construct estimator for the model (with directory specification)
@@ -79,6 +80,10 @@ def input_fn(data_file, num_epochs, batch_size):
   #reading raw file
   dataset = tf.data.TextLineDataset(data_file)
   
+  #shuffle for each epoch run
+  #HINT: does not improve accuracy; since there are timestamps shuffling maybe not useful
+  #dataset = dataset.shuffle(buffer_size=32000)
+  
   #map and parse input strings into features
   dataset = dataset.map(parse_csv, num_parallel_calls=5)
   
@@ -94,8 +99,12 @@ def input_fn(data_file, num_epochs, batch_size):
 #main
 def main(unused_argv):
   #parametrization
+  #for a good explanation look at https://towardsdatascience.com/epoch-vs-iterations-vs-batch-size-4dfb9c7ce9c9
+  #num_epochs is the number of full runs through the training data file (5 = run five times)
   num_epochs = 5
+  #batch_size derives the amount of steps in each epoch (1 means there is one step with all 400.000 input tuples at once)
   batch_size = 10
+  #directories
   model_dir = "./model/"
   train_data = "./train_data/training_data.csv"
   test_data = "./train_data/testing_data.csv"