Make validation optional

rafid-dev · rafid-dev · commit 9b6fb1682ab7 · 2024-03-29T23:34:38.000+06:00
diff --git a/src/main.cu b/src/main.cu
@@ -11,7 +11,7 @@ int main(int argc, char* argv[]) {
     argparse::ArgumentParser program("Grapheus");
 
     program.add_argument("data").required().help("Directory containing training files");
-    program.add_argument("--val-data").required().help("Directory containing validation files");
+    program.add_argument("--val-data").help("Directory containing validation files");
     program.add_argument("--output").required().help("Output directory for network files");
     program.add_argument("--resume").help("Weights file to resume from");
     program.add_argument("--epochs")
@@ -66,9 +66,6 @@ int main(int argc, char* argv[]) {
     // Fetch training dataset paths
     std::vector<std::string> train_files = dataset::fetch_dataset_paths(program.get("data"));
 
-    // Fetch validation dataset paths
-    std::vector<std::string> val_files = dataset::fetch_dataset_paths(program.get("--val-data"));
-
     // Print training dataset file list if files are found
     if (!train_files.empty()) {
         std::cout << "Training Dataset Files:" << std::endl;
@@ -84,6 +81,13 @@ int main(int argc, char* argv[]) {
         exit(0);
     }
 
+    // Fetch validation dataset paths
+    std::vector<std::string> val_files;
+
+    if (program.present("--val-data")) {
+        val_files = dataset::fetch_dataset_paths(program.get("--val-data"));
+    }
+
     // Print validation dataset file list if files are found
     if (!val_files.empty()) {
         std::cout << "Validation Dataset Files:" << std::endl;
@@ -93,9 +97,6 @@ int main(int argc, char* argv[]) {
         std::cout << "Total validation files: " << val_files.size() << std::endl;
         std::cout << "Total validation positions: " << dataset::count_total_positions(val_files)
                   << std::endl;
-    } else {
-        std::cout << "No validation files found in " << program.get("--val-data") << std::endl;
-        exit(0);
     }
 
     const int   total_epochs  = program.get<int>("--epochs");
@@ -121,10 +122,13 @@ int main(int argc, char* argv[]) {
     using BatchLoader = dataset::BatchLoader<chess::Position>;
 
     BatchLoader train_loader {train_files, batch_size};
-    BatchLoader val_loader {val_files, batch_size};
-
     train_loader.start();
-    val_loader.start();
+
+    std::optional<BatchLoader> val_loader;
+    if (val_files.size() > 0) {
+        val_loader.emplace(val_files, batch_size);
+        val_loader->start();
+    }
 
     model::BerserkModel model {static_cast<size_t>(ft_size), lambda, static_cast<size_t>(save_rate)};
     model.set_loss(MPE {2.5, true});
@@ -145,7 +149,7 @@ int main(int argc, char* argv[]) {
     model.train(train_loader, val_loader, total_epochs, epoch_size);
 
     train_loader.kill();
-    val_loader.kill();
+    val_loader->kill();
 
     close();
     return 0;
diff --git a/src/models/chessmodel.h b/src/models/chessmodel.h
@@ -11,6 +11,8 @@
 #include "../operations/operations.h"
 
 #include <algorithm>
+#include <optional>
+
 namespace model {
 
 using namespace nn;
@@ -45,14 +47,14 @@ struct ChessModel : Model {
      * of epochs.
      *
      * @param train_loader The batch loader for training data.
-     * @param val_loader The batch loader for validation data.
+     * @param val_loader The batch loader for validation data (optional).
      * @param epochs Number of training epochs (default: 1500).
      * @param epoch_size Number of batches per epoch (default: 1e8).
      */
-    void train(BatchLoader& train_loader,
-               BatchLoader& val_loader,
-               int          epochs     = 1500,
-               int          epoch_size = 1e8) {
+    void train(BatchLoader&                train_loader,
+               std::optional<BatchLoader>& val_loader,
+               int                         epochs     = 1500,
+               int                         epoch_size = 1e8) {
 
         this->compile(train_loader.batch_size);
 
@@ -90,19 +92,25 @@ struct ChessModel : Model {
                 }
             }
 
-            // Validation phase
-            for (int b = 1; b <= epoch_size / val_loader.batch_size; b++) {
-                auto* ds = val_loader.next();
-                setup_inputs_and_outputs(ds);
+            int val_epoch_size = epoch_size / 10;
 
-                float val_batch_loss = loss();
-                total_val_loss += val_batch_loss;
+            // Validation phase (if validation loader is provided)
+            if (val_loader.has_value()) {
+                for (int b = 1; b <= val_epoch_size / val_loader->batch_size; b++) {
+                    auto* ds = val_loader->next();
+                    setup_inputs_and_outputs(ds);
+
+                    float val_batch_loss = loss();
+                    total_val_loss += val_batch_loss;
+                }
             }
 
-            float epoch_loss = total_epoch_loss / (epoch_size / train_loader.batch_size);
-            float val_loss   = total_val_loss / (epoch_size / val_loader.batch_size);
+            float epoch_loss = total_epoch_loss / (val_epoch_size / train_loader.batch_size);
+            float val_loss   = (val_loader.has_value())
+                                   ? total_val_loss / (val_epoch_size / val_loader->batch_size)
+                                   : 0;
 
-            printf(", val_loss = [%1.8f]", val_loss);
+            printf(", val_loss = [%1.8f] ", val_loss);
             next_epoch(epoch_loss, val_loss);
             std::cout << std::endl;
         }
