Update app.py

app.py

@@ -6,12 +6,88 @@ import os
 import numpy as np
 import tensorflow as tf
 
-emotion_labels = {0: 'neutral', 1: 'calm', 2: 'happy', 3: 'sad', 4: 'angry'
-, 5: 'fearful'}
+emotion_labels = {0: 'neutral', 1: 'calm', 2: 'happy', 3: 'sad', 4: 'angry', 5: 'fearful'}
 
 def trained_model(model_path):
+
+    input_visual = tf.keras.Input((120, 120, 3, 10), name="input_visual")  # 90 - 120
+    input_audio_cnn = tf.keras.Input((150, 512, 1), name="input_audio_cnn")
+    input_audio_wave = tf.keras.Input((20, 13077), name="input_audio_wave")
+
+    # Visual branch
+    x_v = tf.keras.layers.Conv3D(10, (3, 3, 3), strides=(2, 2, 1), padding='same')(input_visual)
+    x_v = tf.keras.layers.BatchNormalization()(x_v)
+    x_v = tf.keras.layers.ReLU()(x_v)
+    x_v = tf.keras.layers.MaxPooling3D((3, 3, 1))(x_v)
+
+    x_v = tf.keras.layers.Conv3D(40, (3, 3, 3), strides=(2, 2, 1), padding='same')(x_v)
+    x_v = tf.keras.layers.BatchNormalization()(x_v)
+    x_v = tf.keras.layers.ReLU()(x_v)
+    x_v = tf.keras.layers.MaxPooling3D((3, 3, 1))(x_v)
+
+    x_v = tf.keras.layers.Flatten()(x_v)
+
+    x_v = tf.keras.layers.Dropout(0.2)(x_v)
+    x_v = tf.keras.layers.Dense(500)(x_v)
+    x_v = tf.keras.layers.BatchNormalization()(x_v)
+    x_v = tf.keras.layers.ReLU()(x_v)
+
+    # Audio cnn branch
+    x_c = tf.keras.layers.Conv2D(5, (3, 3), strides=(2, 2), padding='same')(input_audio_cnn)
+    x_c = tf.keras.layers.BatchNormalization()(x_c)
+    x_c = tf.keras.layers.ReLU()(x_c)
+    x_c = tf.keras.layers.MaxPooling2D((3, 3))(x_c)
+
+    x_c = tf.keras.layers.Conv2D(30, (3, 3), strides=(2, 2), padding='same')(x_c)
+    x_c = tf.keras.layers.BatchNormalization()(x_c)
+    x_c = tf.keras.layers.ReLU()(x_c)
+    x_c = tf.keras.layers.MaxPooling2D((2, 2))(x_c)
+
+    x_c = tf.keras.layers.Conv2D(100, (3, 3), strides=(1, 1), padding='same')(x_c)
+    x_c = tf.keras.layers.BatchNormalization()(x_c)
+    x_c = tf.keras.layers.ReLU()(x_c)
+    x_c = tf.keras.layers.Conv2D(200, (3, 3), strides=(1, 1), padding='same')(x_c)
+    x_c = tf.keras.layers.BatchNormalization()(x_c)
+    x_c = tf.keras.layers.ReLU()(x_c)
+    x_c = tf.keras.layers.MaxPooling2D((2, 2))(x_c)
+
+    x_c = tf.keras.layers.Flatten()(x_c)
+
+    x_c = tf.keras.layers.Dropout(0.2)(x_c)
+    x_c = tf.keras.layers.Dense(500)(x_c)
+    x_c = tf.keras.layers.BatchNormalization()(x_c)
+    x_c = tf.keras.layers.ReLU()(x_c)
+
+    # Audio wave branch
+    x_w = tf.keras.layers.LSTM(500)(input_audio_wave)
+    x_w = tf.keras.layers.RepeatVector(20)(x_w)
+    x_w = tf.keras.layers.LSTM(500)(x_w)
+
+    x_w = tf.keras.layers.Flatten()(x_w)
+
+    x_w = tf.keras.layers.Dropout(0.2)(x_w)
+    x_w = tf.keras.layers.Dense(500)(x_w)
+    x_w = tf.keras.layers.BatchNormalization()(x_w)
+    x_w = tf.keras.layers.ReLU()(x_w)
+
+    # Audio fusion
+    x_a = x_c + x_w
+    x_a = tf.keras.layers.Dense(500)(x_a)
+    x_a = tf.keras.layers.BatchNormalization()(x_a)
+    x_a = tf.keras.layers.ReLU()(x_a)
+
+    # Fusion
+    x = x_a + x_v
+    x = tf.keras.layers.Dense(500)(x)
+    x = tf.keras.layers.BatchNormalization()(x)
+    x = tf.keras.layers.ReLU()(x)
+
+    # Output
+    x = tf.keras.layers.Dropout(0.1)(x) 
+    x = tf.keras.layers.Dense(6, activation='softmax', name='output_classification')(x)  # 8 - 6
+
     
-    model = tf.load_model(model_path)
+    model = model.load(model_path)
 
     return model