major bug fixes

Author: RoadToML

model.py

@@ -51,10 +51,11 @@
 BS = 600 # batch size
 validation_split = 0.01
 batch_per_epoch = np.ceil(len(df['image']) / BS).astype(int)
-epochs = 100
+epochs = 40
 alpha = 0.01
-lr = 0.000001
+lr = 0.00001
 mom = 0.000002
+optimizer = 'SGD'
 
 # ########################################################
 # Create training data labels and target
@@ -67,7 +68,7 @@ def training_gen():
             try:
                 if len(training_data) != BS:
 
-                    image = cv2.imread(f"small_training_images/{img}.png", 1)
+                    image = cv2.imread(f"smaller_training_images/{img}.png", 1)
                     #image = cv2.resize(image, (400,400))
                     image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
 
@@ -77,12 +78,12 @@ def training_gen():
 
                 else:
                     X = [image for image, angle, normal_v in training_data]
-                    other_inp = [image for image, angle, normal_v in training_data]
-                    Y = [image for image, angle, normal_v in training_data]
+                    other_inp = [normal_v for image, angle, normal_v in training_data]
+                    Y = [angle for image, angle, normal_v in training_data]
 
                     training_data = []
 
-                    X = np.array(image).reshape(-1, 400,400,3)
+                    X = (np.array(image)/255.).reshape(-1, 300,250,3)
                     Y = np.array([angle])
                     other_inp = np.array([velocity])
 
@@ -95,7 +96,7 @@ def training_gen():
 # ########################################################
 # Create test data labels and target
 test_data = []
-_iterator_obj = list(zip(df['image'], df['steering_angle'], df['outcome'], test_df['normal_velocity']))
+_iterator_obj = list(zip(df['image'], df['steering_angle'], df['outcome'], df['normal_velocity']))
 random.shuffle(_iterator_obj)
 
 for img, angle, outcome, velocity in _iterator_obj:
@@ -104,7 +105,7 @@ def training_gen():
         if len(test_data) == int(len(df['image']) * validation_split):
             break
 
-        test_image = cv2.imread(f"small_training_images/{img}.png", 1)
+        test_image = cv2.imread(f"smaller_training_images/{img}.png", 1)
         #test_image = cv2.resize(test_image, (400,400))
         test_image = cv2.cvtColor(test_image, cv2.COLOR_BGR2RGB)
 
@@ -120,7 +121,7 @@ def training_gen():
 test_Y = [label for img, label, velocity in test_data]
 test_other_inp = [velocity for img, label, velocity in test_data]
 
-test_X = np.array(test_X).reshape(-1, 400,400,3)
+test_X = np.array(test_X).reshape(-1, 300,250,3)
 test_Y = np.array(test_Y)
 test_other_inp = np.array(test_other_inp)
 print(test_other_inp.shape)
@@ -129,31 +130,30 @@ def training_gen():
 # Create Model
 def create_model():
     # model = Sequential()
-    input1 = Input(shape=(400,400,3))
+    input1 = Input(shape=(300,250,3))
 
     # model.add(Flatten(input_shape = (940,940,3)))
-    conv1 = Conv2D(36, (3, 3), strides=(2, 2))(input1)
-    #conv1 = model.add(Conv2D(20, (5, 5), strides=(2, 2), input_shape = (940,940,3)))
+    conv1 = Conv2D(36, (3, 3), strides=(2, 2), padding='same')(input1)
     activ1 =LeakyReLU(alpha=alpha)(conv1)
-    pool1 =MaxPooling2D(pool_size=(2, 2))(activ1)
+    pool1 =MaxPooling2D(pool_size=(2, 2), padding='valid')(activ1)
 
-    conv2 = Conv2D(25, (3, 3), strides=(2, 2))(pool1)
+    conv2 = Conv2D(25, (3, 3), strides=(2, 2), padding='same')(pool1)
     activ2 =LeakyReLU(alpha=alpha)(conv2)
-    pool2 = MaxPooling2D(pool_size=(2, 2))(activ2)
+    pool2 = MaxPooling2D(pool_size=(2, 2), padding='valid')(activ2)
 
-    conv3 = Conv2D(16, (3, 3), strides=(2, 2))(pool2)
+    conv3 = Conv2D(16, (3, 3), strides=(2, 2), padding='same')(pool2)
     activ3 = LeakyReLU(alpha=alpha)(conv3)
-    pool3 = MaxPooling2D(pool_size=(2, 2))(activ3)
+    pool3 = MaxPooling2D(pool_size=(2, 2), padding='valid')(activ3)
 
-    conv4 = Conv2D(9, (3, 3), strides=1)(pool3)
-    activ4 = LeakyReLU(alpha=alpha)(conv4)
-    pool4 = MaxPooling2D(pool_size=(1, 1))(activ4)
+    # conv4 = Conv2D(9, (3, 3), strides=(1, 1), padding='valid')(pool3)
+    # activ4 = LeakyReLU(alpha=alpha)(conv4)
+    #pool4 = MaxPooling2D(pool_size=(1, 1), padding='valid')(activ4)
 
-    conv5 = Conv2D(4, (3, 3), strides=1)(pool4)
-    activ5 = LeakyReLU(alpha=alpha)(conv5)
-    pool5 = MaxPooling2D(pool_size=(1, 1))(activ5)
+    # conv5 = Conv2D(4, (3, 3), strides=(1, 1), padding='valid')(activ4)
+    # activ5 = LeakyReLU(alpha=alpha)(conv5)
+    #pool5 = MaxPooling2D(pool_size=(1, 1))(activ5)
 
-    flat1 = Flatten()(pool5)
+    flat1 = Flatten()(pool3)
 
     input2 = Input(shape = (1, ))
     concat1 = Concatenate(axis=1)([flat1, input2])
@@ -167,7 +167,7 @@ def create_model():
     model = Model(inputs = [input1, input2], outputs = dense4)
 
     model.compile(loss='mean_squared_error',
-    metrics=['accuracy', 'mae', 'mse'], optimizer='sgd')
+    metrics=['accuracy', 'mae', 'mse'], optimizer=optimizer)
 
     optimizers.SGD(lr=lr, momentum=mom)
     #optimizers.RMSprop(lr=lr)
@@ -205,4 +205,8 @@ def create_model():
 
 
 loss_graph = plt.plot(loss)
+plt.title(f'MSE of {round(float(loss[-1]), 6)} over {epochs} epochs with {optimizer} optimizer \nand 3 hidden conv layers')
 plt.show()
+
+with open(f'results/{optimizer}_3.txt', 'a') as f:
+    print(loss, file = f)