Do not set tensor by `self.xxx=` in models. (#318)

examples/A3C-Gym/train-atari.py

@@ -102,27 +102,26 @@ class Model(ModelDesc):
 
     def _build_graph(self, inputs):
         state, action, futurereward, action_prob = inputs
-        logits, self.value = self._get_NN_prediction(state)
-        self.value = tf.squeeze(self.value, [1], name='pred_value')  # (B,)
-        self.policy = tf.nn.softmax(logits, name='policy')
+        logits, value = self._get_NN_prediction(state)
+        value = tf.squeeze(value, [1], name='pred_value')  # (B,)
+        policy = tf.nn.softmax(logits, name='policy')
         is_training = get_current_tower_context().is_training
         if not is_training:
             return
-        log_probs = tf.log(self.policy + 1e-6)
+        log_probs = tf.log(policy + 1e-6)
 
         log_pi_a_given_s = tf.reduce_sum(
             log_probs * tf.one_hot(action, NUM_ACTIONS), 1)
-        advantage = tf.subtract(tf.stop_gradient(self.value), futurereward, name='advantage')
+        advantage = tf.subtract(tf.stop_gradient(value), futurereward, name='advantage')
 
-        pi_a_given_s = tf.reduce_sum(self.policy * tf.one_hot(action, NUM_ACTIONS), 1)  # (B,)
+        pi_a_given_s = tf.reduce_sum(policy * tf.one_hot(action, NUM_ACTIONS), 1)  # (B,)
         importance = tf.stop_gradient(tf.clip_by_value(pi_a_given_s / (action_prob + 1e-8), 0, 10))
 
         policy_loss = tf.reduce_sum(log_pi_a_given_s * advantage * importance, name='policy_loss')
-        xentropy_loss = tf.reduce_sum(
-            self.policy * log_probs, name='xentropy_loss')
-        value_loss = tf.nn.l2_loss(self.value - futurereward, name='value_loss')
+        xentropy_loss = tf.reduce_sum(policy * log_probs, name='xentropy_loss')
+        value_loss = tf.nn.l2_loss(value - futurereward, name='value_loss')
 
-        pred_reward = tf.reduce_mean(self.value, name='predict_reward')
+        pred_reward = tf.reduce_mean(value, name='predict_reward')
         advantage = symbf.rms(advantage, name='rms_advantage')
         entropy_beta = tf.get_variable('entropy_beta', shape=[],
                                        initializer=tf.constant_initializer(0.01), trainable=False)

examples/Char-RNN/char-rnn.py

@@ -81,8 +81,7 @@ class Model(ModelDesc):
                                   initializer=tf.constant_initializer())
             ret = symbolic_functions.shapeless_placeholder(ret, 0, name=n)
             return ret
-        self.initial = initial = \
-            (rnn.LSTMStateTuple(get_v('c0'), get_v('h0')),
+        initial = (rnn.LSTMStateTuple(get_v('c0'), get_v('h0')),
                    rnn.LSTMStateTuple(get_v('c1'), get_v('h1')))
 
         embeddingW = tf.get_variable('embedding', [param.vocab_size, param.rnn_size])
@@ -91,12 +90,12 @@ class Model(ModelDesc):
         input_list = tf.unstack(input_feature, axis=1)  # seqlen x (Bxrnnsize)
 
         outputs, last_state = rnn.static_rnn(cell, input_list, initial, scope='rnnlm')
-        self.last_state = tf.identity(last_state, 'last_state')
+        last_state = tf.identity(last_state, 'last_state')
 
         # seqlen x (Bxrnnsize)
         output = tf.reshape(tf.concat(outputs, 1), [-1, param.rnn_size])  # (Bxseqlen) x rnnsize
         logits = FullyConnected('fc', output, param.vocab_size, nl=tf.identity)
-        self.prob = tf.nn.softmax(logits / param.softmax_temprature, name='prob')
+        prob = tf.nn.softmax(logits / param.softmax_temprature, name='prob')
 
         xent_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
             logits=logits, labels=tf.reshape(nextinput, [-1]))

examples/GAN/GAN.py

@@ -66,6 +66,10 @@ class GANModelDesc(ModelDescBase):
 
 class GANTrainer(Trainer):
     def __init__(self, config):
+        """
+        GANTrainer expects a ModelDesc in config which sets the following attribute
+        after :meth:`_build_graph`: g_loss, d_loss, g_vars, d_vars.
+        """
         input = QueueInput(config.dataflow)
         model = config.model
 

examples/cifar-convnet.py

@@ -15,7 +15,7 @@ from tensorpack.dataflow import dataset
 """
 A small convnet model for Cifar10 or Cifar100 dataset.
 
-Cifar10:
+Cifar10 trained on 1 GPU:
     91% accuracy after 50k step.
     41 step/s on TitanX