Patch f1 (https://github.com/ebonyclock/vizdoom_cig2017) to work under latest TensorFlow + latest tensorpack

f1.diff

diff --git i/f1/F1_track1/agent.py w/f1/F1_track1/agent.py
index a717842..4ae8405 100644
--- i/f1/F1_track1/agent.py
+++ w/f1/F1_track1/agent.py
@@ -140,15 +140,15 @@ class FinalEnv(RLEnvironment):
 
 
 class Model(ModelDesc):
-    def _get_input_vars(self):
-        return [InputVar(tf.float32, (None,) + IMAGE_SHAPE3, 'image'),
-                InputVar(tf.float32, (None, 5), 'vars')]
+    def _get_inputs(self):
+        return [InputDesc(tf.float32, (None,) + IMAGE_SHAPE3, 'image'),
+                InputDesc(tf.float32, (None, 5), 'vars')]
 
-    def _get_NN_prediction(self, state, is_training):
+    def _get_NN_prediction(self, state):
         """ image: [0,255]"""
         image, vars = state
         image = image / 255.0
-        with argscope(Conv2D, nl=PReLU.f):
+        with argscope(Conv2D, nl=PReLU.symbolic_function):
             l = (LinearWrap(image)
                 .Conv2D('conv0', out_channel=32, kernel_shape=7, stride=2)
             # 60
@@ -160,16 +160,16 @@ class Model(ModelDesc):
             # 7
                 .Conv2D('conv4', out_channel=192, kernel_shape=3, padding='VALID')
             # 5
-                .FullyConnected('fcimage', 1024, nl=PReLU.f)())
+                .FullyConnected('fcimage', 1024, nl=PReLU.symbolic_function)())
 
         vars = tf.tile(vars, [1, 10], name='tiled_vars')
-        feat = tf.concat(1, [l, vars])
+        feat = tf.concat([l, vars], 1)
 
         policy = FullyConnected('fc-pi-m', feat, out_dim=NUM_ACTIONS, nl=tf.identity) * 0.1
         return policy
 
-    def _build_graph(self, inputs, is_training):
-        policy = self._get_NN_prediction(inputs, is_training)
+    def _build_graph(self, inputs):
+        policy = self._get_NN_prediction(inputs)
         self.logits = tf.nn.softmax(policy, name='logits')
 
 class Model2(ModelDesc):
@@ -213,9 +213,9 @@ class Runner(object):
         cfg = PredictConfig(
                 model=Model(),
                 session_init=SaverRestore('model.tfmodel'),
-                input_var_names=['image', 'vars'],
-                output_var_names=['logits'])
-        self._pred_func = get_predict_func(cfg)
+                input_names=['image', 'vars'],
+                output_names=['logits'])
+        self._pred_func = OfflinePredictor(cfg)
 
     def action_func(self, inputs):
         f = self._pred_func
diff --git i/f1/F1_track1/history.py w/f1/F1_track1/history.py
index 5049412..511cef6 100644
--- i/f1/F1_track1/history.py
+++ w/f1/F1_track1/history.py
@@ -11,12 +11,12 @@ __all__ = ['HistoryPlayerWithVar']
 class HistoryPlayerWithVar(HistoryFramePlayer):
     def current_state(self):
         assert len(self.history) != 0
-        assert len(self.history[0]) == 2, "state needs to be like [img, vars]"
-        diff_len = self.history.maxlen - len(self.history)
-        zeros = [np.zeros_like(self.history[0][0]) for k in range(diff_len)]
-        for k in self.history:
+        assert len(self.history.buf[0]) == 2, "state needs to be a 2-list like [img, vars]"
+        diff_len = self.history.buf.maxlen - len(self.history)
+        zeros = [np.zeros_like(self.history.buf[0][0]) for k in range(diff_len)]
+        for k in self.history.buf:
             zeros.append(k[0])
         img = np.concatenate(zeros, axis=2)
-        gvar = self.history[-1][1]
+        gvar = self.history.buf[-1][1]
         return img, gvar
 

Hi, I have tried the this patch under latest TF and Tensorpack and found the 'history.py' should be modified to make these scripts work.
The indexing is not supported for 'self.history' now and should be changed to 'self.history.buf' I think.

Hope this will help, thanks for sharing your code!

Thanks. The diff is updated.