playing around

hw1/behavior_clone.py

@@ -7,6 +7,7 @@ from hw1 import tf_util
 import gym
 
 envname = 'RoboschoolAnt-v1'
+envname = 'RoboschoolHumanoid-v1'
 
 with open(os.path.join('expert_data', envname + '.pkl'), 'rb') as f:
     expert_data = pickle.load(f)
@@ -14,38 +15,38 @@ with open(os.path.join('expert_data', envname + '.pkl'), 'rb') as f:
 x = tf.placeholder(tf.float32, shape=[None, expert_data['observations'].shape[1]])
 y_true = tf.placeholder(tf.float32, shape=[None, expert_data['actions'].shape[1]])
 
-hidden = tf.layers.Dense(units=64, activation=tf.nn.relu)
-model = tf.layers.Dense(units=expert_data['actions'].shape[1], use_bias=False)(hidden)
-y_pred = model(x)
+hidden = tf.layers.Dense(units=256, activation=tf.nn.relu)
+model = tf.layers.Dense(units=expert_data['actions'].shape[1], use_bias=False)
+y_pred = model(hidden(x))
 
 loss = tf.losses.mean_squared_error(labels=y_true, predictions=y_pred)
 
 optimizer = tf.train.AdamOptimizer(0.001)
 train = optimizer.minimize(loss)
 '''
-(x_train, y_train),(x_test, y_test) = mnist.load_data()
-x_train, x_test = x_train / 255.0, x_test / 255.0
 
 model = tf.keras.models.Sequential([
-  tf.keras.layers.Flatten(input_shape=(28, 28)),
-  tf.keras.layers.Dense(512, activation=tf.nn.relu),
-  tf.keras.layers.Dropout(0.2),
-  tf.keras.layers.Dense(10, activation=tf.nn.softmax)
+    tf.keras.layers.Dense(64, activation=tf.nn.relu, input_shape=[expert_data['observations'].shape[1]]),
+    tf.keras.layers.Dense(expert_data['actions'].shape[1])
 ])
 model.compile(optimizer='adam',
-              loss='sparse_categorical_crossentropy',
-              metrics=['accuracy'])
+              loss='mse',
+              metrics=['mae', 'mse'])
+model.summary()
+model.fit(expert_data['observations'], expert_data['actions'], epochs=5, batch_size=256)
 
-model.fit(x_train, y_train, epochs=5)
-model.evaluate(x_test, y_test)
 '''
+epochs = 300
+batch_size = 256
 with tf.Session() as sess:
+    '''
+    '''
     tf_util.initialize()
-
-    for i in range(1000):
+    for e in range(epochs):
+        for i in range(0, expert_data['observations'].shape[0], batch_size):
             _, loss_value = sess.run(
                 (train, loss),
-            feed_dict={x: expert_data['observations'], y_true: expert_data['actions']})
+                feed_dict={x: expert_data['observations'][i:i + batch_size, :], y_true: expert_data['actions'][i:i + batch_size, :]})
         print(loss_value)
 
     # Play
@@ -57,15 +58,17 @@ with tf.Session() as sess:
     steps = 0
     while not done:
         action = sess.run(y_pred, feed_dict={x: obs[np.newaxis, :]})
+        # action = model.predict(obs[np.newaxis, :])
         # observations.append(obs)
         # actions.append(action)
         obs, r, done, _ = env.step(np.squeeze(action))
         totalr += r
         steps += 1
         env.render()
-        if steps % 100 == 0: print("%i/%i" % (steps, max_steps))
+        if steps % 100 == 0:
+            print("%i/%i" % (steps, max_steps))
         if steps >= max_steps:
             break
-    # print(sess.run(y_pred))
 
-np.mean((expert_data['actions'] - np.mean(expert_data['actions'], axis=0)) ** 2)
+# print(sess.run(y_pred))
+# np.mean((expert_data['actions'] - np.mean(expert_data['actions'], axis=0)) ** 2)