Add parametrizability for how the training data is presented

1 files changed, 8 insertions, 4 deletions

diff --git a/model/mlp.py b/model/mlp.py
index 05898a5..fc86b7b 100644
--- a/model/mlp.py
+++ b/model/mlp.py
@@ -51,14 +51,18 @@ class Stream(object):
         self.config = config
 
     def train(self, req_vars):
-        stream = TaxiDataset('train')
-        stream = DataStream(stream, iteration_scheme=TaxiTimeCutScheme())
-
         valid = TaxiDataset(self.config.valid_set, 'valid.hdf5', sources=('trip_id',))
         valid_trips_ids = valid.get_data(None, slice(0, valid.num_examples))[0]
 
+        stream = TaxiDataset('train')
+
+        if hasattr(self.config, 'use_cuts_for_trainig') and self.config.use_cuts_for_training:
+            stream = DataStream(stream, iteration_scheme=TaxiTimeCutScheme())
+        else:
+            stream = DataStream(stream, iteration_scheme=ShuffledExampleScheme(stream.num_examples))
+
         stream = transformers.TaxiExcludeTrips(valid_trips_ids, stream)
-        stream = transformers.TaxiGenerateSplits(stream, max_splits=1)
+        stream = transformers.TaxiGenerateSplits(stream, max_splits=self.config.max_splits)
 
         stream = transformers.TaxiAddDateTime(stream)
         stream = transformers.TaxiAddFirstLastLen(self.config.n_begin_end_pts, stream)