homework7

homework8/table_tennis.py

@@ -8,6 +8,7 @@ from skimage.transform import resize
 from collections import deque
 from collections import deque
 import os.path
 import os.path
 import os, time
 import os, time
+import cv2
 
 
 # Define Hyperparameters
 # Define Hyperparameters
 FLAGS = tf.flags.FLAGS
 FLAGS = tf.flags.FLAGS
@@ -51,15 +52,15 @@ def atari_model():
     # Normalise the inputs from [0,255] to [0,1] - to make processing easier
     # Normalise the inputs from [0,255] to [0,1] - to make processing easier
     normalised = keras.layers.Lambda(lambda x: x/255.0, name='normalised')(frames_input)
     normalised = keras.layers.Lambda(lambda x: x/255.0, name='normalised')(frames_input)
     # Conv1 is 16 8x8 filters with a stride of 4 and a ReLU
     # Conv1 is 16 8x8 filters with a stride of 4 and a ReLU
-    conv1 = '?'
+    conv1 = keras.layers.Conv2D(filters=32, kernel_size=8, strides=4, activation="relu")(normalised)
     # Conv2 is 32 4x4 filters with a stride of 2 and a ReLU
     # Conv2 is 32 4x4 filters with a stride of 2 and a ReLU
-    conv2 = '?'
+    conv2 = keras.layers.Conv2D(filters=32, kernel_size=4, strides=2, activation="relu")(conv1)
     # Flatten the output from Conv2
     # Flatten the output from Conv2
     conv2_flatten = keras.layers.Flatten()(conv2)
     conv2_flatten = keras.layers.Flatten()(conv2)
     # Then a fully connected layer with 128 ReLU units
     # Then a fully connected layer with 128 ReLU units
-    dense1 = '?'
+    dense1 = keras.layers.Dense(units=128, activation="relu")(conv2_flatten)
     # Then a fully connected layer with a unit to map to each of the actions and no activation
     # Then a fully connected layer with a unit to map to each of the actions and no activation
-    output =  '?'
+    output = keras.layers.Dense(units=Action_size)(dense1)
     # Then we multiply the output by the action mask
     # Then we multiply the output by the action mask
     # When trying to find the value of all the actions this will be a mask full of 1s
     # When trying to find the value of all the actions this will be a mask full of 1s
     # When trying to find the value of a specific action, the mask will only be 1 for a single action
     # When trying to find the value of a specific action, the mask will only be 1 for a single action
@@ -73,7 +74,7 @@ def atari_model():
     # Define optimiser
     # Define optimiser
     optimiser = tf.train.AdamOptimizer()
     optimiser = tf.train.AdamOptimizer()
     # Compile model
     # Compile model
-    loss = '?'  # Task-2: to choose the loss function
+    loss = "mse"  # Task-2: to choose the loss function
     model.compile(optimizer=optimiser, loss=loss)
     model.compile(optimizer=optimiser, loss=loss)
     # Return the model
     # Return the model
     return model
     return model
@@ -81,7 +82,7 @@ def atari_model():
 # Create a model to use as a target
 # Create a model to use as a target
 def atari_model_target():
 def atari_model_target():
     # Task-3: to implement the code for atari_model_target
     # Task-3: to implement the code for atari_model_target
-    model = '?'
+    model = atari_model()
     return model
     return model
 
 
 # get action from model using epsilon-greedy policy
 # get action from model using epsilon-greedy policy
@@ -116,7 +117,7 @@ def train_memory_batch(memory, model):
     # Fill up our arrays with our minibatch
     # Fill up our arrays with our minibatch
     for id, val in enumerate(mini_batch):
     for id, val in enumerate(mini_batch):
         state[id] = val[0]
         state[id] = val[0]
-        print(val[0].shape)
+        # print(val[0].shape)
         next_state[id] = val[3]
         next_state[id] = val[3]
         action.append(val[1])
         action.append(val[1])
         reward.append(val[2])
         reward.append(val[2])
@@ -209,7 +210,7 @@ def train():
 
 
             # Select an action based on our current model
             # Select an action based on our current model
             # Task-4: select_model = model_target or select_model = model
             # Task-4: select_model = model_target or select_model = model
-            select_model = '?'
+            select_model = model
             action = get_action(state_history, epsilon, global_step, select_model)
             action = get_action(state_history, epsilon, global_step, select_model)
 
 
             # Convert action from array numbers to real numbers
             # Convert action from array numbers to real numbers
@@ -221,6 +222,8 @@ def train():
 
 
             # Record output from the environment
             # Record output from the environment
             observation, reward, done, info = env.step(real_action)
             observation, reward, done, info = env.step(real_action)
+            cv2.imshow("", observation)
+            cv2.waitKey(1)
 
 
             # Process the observation
             # Process the observation
             next_state = pre_processing(observation)
             next_state = pre_processing(observation)