Create readme

DrivingAI/model.py

@@ -14,21 +14,21 @@ def __init__(self, save=False, summary=False, global_step=None):
         os.makedirs(self.log_directory, exist_ok=True)
         os.makedirs(self.network_directory, exist_ok=True)
         self._save = save
-        self.saver = tf.train.Saver()
+        self.saver = tf.train.Saver(tf.trainable_variables())
         self.global_step = global_step
         self.session = tf.Session()
+        self.session.run(tf.global_variables_initializer())
+        self.session.run(tf.local_variables_initializer())
         self.coord = tf.train.Coordinator()
         tf.train.start_queue_runners(self.session, self.coord)
         try:
             ckpt = tf.train.get_checkpoint_state(self.network_directory)
             if ckpt is None:
-                self.session.run(tf.global_variables_initializer())
                 print("\nCreated a new network\n")
             else:
                 self.saver.restore(self.session, ckpt.model_checkpoint_path)
                 print("\nLoaded an existing network\n")
         except Exception as e:
-            self.session.run(tf.global_variables_initializer())
             print("\nCreated a new network (%s)\n"%repr(e))
         if summary:
             self.summary_ops = tf.summary.merge_all()
@@ -43,7 +43,7 @@ def save_summary(self, global_step, fd=None):
 
     def save_network(self):
         if self._save:
-            self.saver.save(self.session, self.model_file_name, self.global_step)
+            self.saver.save(self.session, self.model_file_name, self.global_step, write_meta_graph=False)
 
     def __enter__(self):
         return self

readme.md

@@ -0,0 +1,29 @@
+# Selfdriving Car AI and Simulator
+This project contains a neural network for driving a car in a simulator. The simulator is also part of the project. The goal with not using a existing game/simulator is to allow more control over the data being fed to the AI, which made some experimentation possible.
+
+## Simulator
+The simulator is made with unity. It generates random terrains in order to create varied learning situations. The simulator communicates with the AI through a local socket, this means that often both the simulator and the AI have to be started.
+
+## AI
+The AI receives the following input:  
+1. A color image.  
+2. A grayscale image created from the depth and normal buffers (a LADAR scanner would be the real life equivalent).  
+3. The current speed of the car.  
+
+The output is acceleration and turning values.
+
+## Download
+A windows version of the simulator can be downloaded [here](https://github.com/Aggrathon/CarAiSimulator/releases).
+The trained network is unfortunately too big to distribute here (maybe the fully connected layers coud be smaller).
+
+## Usage
+Here is the normal flow for using the AI:
+1. Use the simulator and the `record.py` script to to create examples of how humans drive.  
+2. Train the AI on the recorded examples using the `learn.py` script.  
+3. Improve the AI with reinforcement learning, using the simulator and the `train.py` script.  
+4. Let the AI drive in the simulator with the `drive.py` script.
+
+## Dependencies
+- Python 3
+- Tensorflow
+- Unity (2017.2)