Optimise BPTT sequence batching inside NeuralNetwork training loop and add integration tests

Author: FFMG

include/neuralnetwork/neuralnetwork.cpp

@@ -763,6 +763,9 @@ void NeuralNetwork::train(const std::vector<std::vector<double>>& training_input
 
   std::vector<std::vector<double>> bptt_in;
   std::vector<std::vector<double>> bptt_out;
+  // Initial creation of BPTT batches before the epoch loop
+  create_bptt_batches(batch_training_inputs, batch_training_outputs, bptt_in, bptt_out);
+
   for (auto epoch = 0; epoch < number_of_epoch; ++epoch)
   {
     // Learning rate
@@ -773,8 +776,11 @@ void NeuralNetwork::train(const std::vector<std::vector<double>>& training_input
 
     _learning_rate = base_helper->learning_rate();
 
-    // (re) create the bptt batches (now returns flattened sequences)
-    create_bptt_batches(batch_training_inputs, batch_training_outputs, bptt_in, bptt_out);
+    // Only recreate/shuffle BPTT batches if shuffle is enabled
+    if (_options.shuffle_bptt_batches())
+    {
+      create_bptt_batches(batch_training_inputs, batch_training_outputs, bptt_in, bptt_out);
+    }
 
     const size_t total_sequences = bptt_in.size();
     for (size_t i = 0; i < total_sequences; i += batch_size)

tests/network_integration_tests.cpp

@@ -427,3 +427,38 @@ TEST(NetworkIntegrationTest, LogTrainingInfoOptionAndSerialization)
   std::remove(test_path.c_str());
 }
 
+TEST(NetworkIntegrationTest, ShuffleBpttBatchesBehavior)
+{
+  auto options_no_shuffle = NeuralNetworkOptions::create({ 2, 2, 1 })
+    .with_learning_rate(0.01)
+    .with_number_of_epoch(5)
+    .with_shuffle_bptt_batches(false)
+    .with_enable_bptt(true)
+    .with_bptt_max_ticks(3)
+    .build();
+
+  auto options_shuffle = NeuralNetworkOptions::create({ 2, 2, 1 })
+    .with_learning_rate(0.01)
+    .with_number_of_epoch(5)
+    .with_shuffle_bptt_batches(true)
+    .with_enable_bptt(true)
+    .with_bptt_max_ticks(3)
+    .build();
+
+  std::vector<std::vector<double>> inputs = {
+    {0.1, 0.2}, {0.3, 0.4}, {0.5, 0.6}, {0.7, 0.8}, {0.9, 1.0}
+  };
+  std::vector<std::vector<double>> outputs = {
+    {0.3}, {0.7}, {1.1}, {1.5}, {1.9}
+  };
+
+  NeuralNetwork nn_no_shuffle(options_no_shuffle);
+  nn_no_shuffle.train(inputs, outputs);
+
+  NeuralNetwork nn_shuffle(options_shuffle);
+  nn_shuffle.train(inputs, outputs);
+
+  SUCCEED();
+}
+
+