Work on fine-tuning and Important Feature selections

WebServer/AIPython/ai-trainer.py

@@ -8,6 +8,7 @@ import os
 os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
 from sklearn.model_selection import train_test_split
 from keras import Sequential, layers, optimizers, losses
+from keras.callbacks import ReduceLROnPlateau
 
 def TrainAI(include_pull):
 
@@ -35,36 +36,43 @@ def TrainAI(include_pull):
     # Split out the test and train
     train_features, test_features, train_labels, test_labels = train_test_split(X, Y, test_size=0.2)
 
-    # Create a normalizer to nomralize the data
-    normalizer = layers.Normalization(axis=-1)
-    normalizer.adapt(np.array(train_features))
-
     # Create the DNN
     dnn_model = Sequential([
         layers.Input(shape=(train_features.shape[1],)), # Load the feature count dynamically
-        normalizer,
-        layers.Dense(64, activation='elu'),
-        layers.Dense(64, activation='elu'),
-        layers.Dense(1)
+        layers.Dense(256, activation='elu'),
+        layers.Dropout(0.1),                    # Small dropout to prevent it from memorizing noise
+        layers.Dense(128, activation='elu'),
+        layers.Dropout(0.1),                    # Small dropout to prevent it from memorizing noise
+        layers.Dense(24, activation='elu'),
+        layers.Dense(1, activation='linear')
     ])
 
+    # Allow negative numbers
+    dnn_model.add(layers.LeakyReLU(alpha=0.01))
+
     # Configure the model
     dnn_model.compile(
-        optimizer=optimizers.Adam(learning_rate=0.00001, clipvalue=1.0),
-        loss=losses.Huber()
+        optimizer=optimizers.Adam(learning_rate=0.0001, clipvalue=1.0),
+        loss="mse",
+        metrics=['mae']             # See it train while it runs
     )
 
     # Show the summary before training the model
     dnn_model.summary()
 
+    # Learning rate reducer
+    reduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.5, patience=5, min_lr=0.0000001)
+
     # Train the model
     Training_Data = dnn_model.fit( 
         train_features,
         train_labels,
         batch_size=64,
-        epochs=39,                  # Tuned to the point before overfitting
+        epochs=50,                  # Tuned to the point before overfitting
         verbose=1,                  # Show progress
-        validation_split = 0.2      # Calculate validation results on 20% of the training data.
+        validation_split = 0.2,     # Calculate validation results on 20% of the training data.
+        shuffle=True,
+        callbacks=[reduce_lr]       # Reduce the learning_rate every run
     )
 
     # Predict
@@ -79,7 +87,7 @@ def TrainAI(include_pull):
     _ = plt.plot(lims, lims)
 
 
-    # Current Test Results: 1.221876300405711e-05
+    # Current Test Results: 0.3382711112499237
     test_results = dnn_model.evaluate(
         test_features, test_labels, verbose=0
     )
@@ -89,4 +97,6 @@ def TrainAI(include_pull):
     dnn_model.save(os.path.join(DATA_DIR, "model.keras"))
 
 if __name__ == "__main__":
-    TrainAI(False)
+    TrainAI(True)
+
+# Last train Predicted Target_Close: [0.0022113274317234755, 0.0021446370519697666, 0.0022628342267125845, 0.002175702480599284, 0.0021452796645462513, 0.0020838389173150063, 0.0017336219316348433, 0.002210840117186308, 0.0021144403144717216, 0.0021278387866914272, 0.0021266420371830463, 0.002261851681396365, 0.002108299173414707, 0.002121902070939541, 0.0022294146474450827]