Simplify model

Signed-off-by: Yaacov Zamir <[email protected]>

model.py

@@ -55,7 +55,7 @@ class DriverModel(nn.Module):
 
     Architecture:
     - Input layer: Size determined by
-      3 (width) x 4 (height) x 7 (possible obstacles) + 3 (car current lane)= 87 neurons.
+      6 (width) x 4 (height) x 7 (possible obstacles) + 6 (car current lane)= 174 neurons.
     - Hidden Layer 1: 512 neurons, followed by batch normalization and 50% dropout.
     - Hidden Layer 2: 256 neurons, followed by batch normalization and 50% dropout.
     - Hidden Layer 3: 128 neurons, followed by batch normalization and 50% dropout.
@@ -71,13 +71,13 @@ class DriverModel(nn.Module):
     - Dropout with a rate of 50% is applied after each hidden layer to prevent overfitting.
 
     Note:
-    - The model expects a flattened version of the 3x4x7 input tensor, which should be reshaped to (batch_size, 84) before being passed to the model.
+    - The model expects a flattened version of the 6x4x7+6 input tensor, which should be reshaped to (batch_size, 174) before being passed to the model.
     """
 
     def __init__(self):
         super(DriverModel, self).__init__()
 
-        self.fc1 = nn.Linear(3 * 4 * 7 + 3, 512)
+        self.fc1 = nn.Linear(6 * 4 * 7 + 6, 512)
         self.bn1 = nn.BatchNorm1d(512)
         self.dropout1 = nn.Dropout(0.5)
 
@@ -128,7 +128,7 @@ def view_to_inputs(array, car_lane):
 
     Args:
         array (list[list[str]]): 2D array representation of the world with obstacles as strings.
-        car_lane (int): current lane of the car, can be 0, 1 or 2.
+        car_lane (int): current lane of the car, can be 0..5 inclusive.
 
     Returns:
         torch.Tensor: A tensor of shape (1, height * width * num_obstacle_types) suitable for model input.
@@ -150,7 +150,7 @@ def view_to_inputs(array, car_lane):
             tensor[i, j, OBSTACLE_TO_INDEX[obstacle]] = 1
 
     world_tensor = tensor.view(-1)
-    car_lane_tensor = torch.tensor([0, 0, 0])
+    car_lane_tensor = torch.zeros(6)
     car_lane_tensor[car_lane] = 1
 
     return torch.cat((world_tensor, car_lane_tensor))

mydriver.py

@@ -48,15 +48,13 @@
 # ----------------------------------------------------------------------------------
 
 
-def build_lane_view(world, height, lane):
+def build_lane_view(world):
     """
     Build a 3xN 2D array representation of the world based on the car's lane and x position.
 
     Args:
         world (World): An instance of the World class providing read-only
                        access to the current game state.
-        height (int): The height of the returned 2D array.
-        lane (int) : The car's current lane, 0 or 1. This determines the starting x-coordinate for the 2D array.
 
     Returns:
         list[list[str]]: 3xN array representation of the world view from the car, where N is the specified height.
@@ -68,11 +66,9 @@ def build_lane_view(world, height, lane):
         The starting x-coordinate is determined by the car's lane. If the lane is 0, the starting x is 0. If the lane is 1, the starting x is 3.
         The function also provides a wrapper around world.get to handle negative y values, returning an empty string for such cases.
     """
+    height = 4
     car_y = world.car.y
 
-    # Determine the starting x-coordinate for the 2D array based on the car's x position
-    start_x = 0 if lane == 0 else 3
-
     # Calculate the starting y-coordinate based on the car's y position and the desired height
     start_y = car_y - height
 
@@ -84,7 +80,7 @@ def get_value(j, i):
 
     # Generate the 2D array from start_y up to world.car.y
     array = [
-        [get_value(j + start_x, i) for j in [0, 1, 2]] for i in range(start_y, car_y)
+        [get_value(j, i) for j in range(6)] for i in range(start_y, car_y)
     ]
 
     return array
@@ -94,30 +90,15 @@ def drive(world):
     """
     Determine the appropriate driving action based on the current state of the world.
 
-    The function first constructs a 3xN 2D view of the world based on the car's position.
-    This view is then converted to an input tensor format suitable for the model.
-    Depending on the car's x position within its lane, the function uses one of two models (`model_x1` or `model_x0`)
-    to predict the best action. If the world view was flipped (because the car is on the rightmost side of its lane),
-    the action might be flipped back (e.g., from LEFT to RIGHT or vice versa).
-
     Args:
         world (World): An instance of the World class providing read-only access to the current game state.
 
     Returns:
         str: The determined action for the car to take. Possible actions include those defined in the `actions` class.
-
-    Notes:
-        The function uses two models (`model_x1` and `model_x0`) to predict actions based on the car's x position within its lane.
-        The `flip_world` flag determines if the world view was flipped horizontally, which affects the final action decision.
     """
-    view_height = 4
-
-    lane = 0 if world.car.x < 3 else 1
-    x_in_lane = world.car.x % 3
-
     # Convert real world input, into a tensor
-    view = build_lane_view(world, view_height, lane)
-    input_tensor = view_to_inputs(view, x_in_lane).unsqueeze(0)
+    view = build_lane_view(world)
+    input_tensor = view_to_inputs(view, world.car.x).unsqueeze(0)
 
     # Use neural network model to get the outputs tensor
     output = model(input_tensor)

train.py

@@ -56,16 +56,16 @@
 
 def generate_obstacle_array():
     """
-    Generates a 4x3 2D array with random obstacles.
+    Generates a 4x6 2D array with random obstacles.
 
     Returns:
-        list[list[str]]: 4x3 2D array with random obstacles.
+        list[list[str]]: 4x6 2D array with random obstacles.
     """
-    array = [["" for _ in range(3)] for _ in range(4)]
+    array = [["" for _ in range(6)] for _ in range(4)]
 
     for i in range(4):
         obstacle = random.choice(list(OBSTACLE_TO_INDEX.keys()))
-        position = random.randint(0, 2)
+        position = random.randint(0, 5)
         array[i][position] = obstacle
 
     return array
@@ -115,7 +115,7 @@ def generate_batch(batch_size):
     inputs = []
     targets = []
     for _ in range(batch_size):
-        car_x = random.choice([0, 1, 2])
+        car_x = random.choice([0, 1, 2, 3, 4, 5])
         array = generate_obstacle_array()
         correct_output = driver_simulator(array, car_x)
 
@@ -172,7 +172,7 @@ def main():
         "--checkpoint-out", default="", help="Path to the output checkpoint file."
     )
     parser.add_argument(
-        "--num-epochs", type=int, default=25, help="Number of epochs for training."
+        "--num-epochs", type=int, default=80, help="Number of epochs for training."
     )
     parser.add_argument(
         "--batch-size", type=int, default=250, help="Batch size for training."