trainer.build_train_tower(). and some doc fix.

docs/conf.py

@@ -23,19 +23,17 @@ os.environ['TENSORPACK_DOC_BUILDING'] = '1'
 
 
 MOCK_MODULES = ['scipy',
-                'tensorflow', 'tensorflow.contrib',
-                'tensorflow.python.ops',
-                'tensorflow.contrib.framework',
-                'tensorflow.models',
-                'tensorflow.models.rnn',
-                'tensorflow.models.rnn.ptb',
-                'tensorflow.python',
-                'tensorflow.python.training',
-                'sklearn.datasets',
+                #'tensorflow', 'tensorflow.contrib',
+                #'tensorflow.python.ops',
+                #'tensorflow.contrib.framework',
+                #'tensorflow.python',
+                #'tensorflow.python.training',
+                'sklearn.datasets', 'sklearn',
                 'scipy.misc', 'h5py', 'nltk',
-                'cv2', 'scipy.io', 'dill', 'zmq', 'subprocess32', 'lmdb', 'tornado.concurrent',
-                'tornado', 'msgpack', 'msgpack_numpy', 'ale_python_interface',
-                'sklearn', 'functools32']
+                'cv2', 'scipy.io', 'dill', 'zmq', 'subprocess32', 'lmdb',
+                'tornado.concurrent', 'tornado',
+                'msgpack', 'msgpack_numpy',
+                'gym', 'functools32']
 for mod_name in MOCK_MODULES:
     sys.modules[mod_name] = mock.Mock(name=mod_name)
 

docs/requirements.txt

@@ -2,5 +2,6 @@ termcolor
 numpy
 tqdm
 decorator
+tensorflow
 Sphinx==1.5.1
 recommonmark==0.4.0

examples/GAN/GAN.py

@@ -73,9 +73,7 @@ class GANTrainer(FeedfreeTrainerBase):
 
     def _setup(self):
         super(GANTrainer, self)._setup()
-        with TowerContext(''):
-            actual_inputs = self._get_input_tensors()
-            self.model.build_graph(actual_inputs)
+        self.build_train_tower()
         opt = self.model.get_optimizer()
 
         self.g_min = opt.minimize(self.model.g_loss, var_list=self.model.g_vars, name='g_op')

tensorpack/tfutils/optimizer.py

@@ -11,6 +11,9 @@ __all__ = ['apply_grad_processors', 'ProxyOptimizer',
 
 
 class ProxyOptimizer(tf.train.Optimizer):
+    """
+    A transparent proxy which delegates all methods of :class:`tf.train.Optimizer`
+    """
     def __init__(self, opt):
         self._opt = opt
 
@@ -54,8 +57,8 @@ def apply_grad_processors(opt, gradprocs):
 
 class PostProcessVariablesOptimizer(ProxyOptimizer):
     """
-    An optimizer which applies an operation to variables (e.g. clipping,
-    quantization) after updating the gradient.
+    An optimizer which applies an operation to variables
+    (e.g. clipping, quantization) after updating the gradient.
     """
     def __init__(self, opt, func, colocate=True):
         """

tensorpack/tfutils/tower.py

@@ -102,6 +102,10 @@ class TowerContext(object):
             self._scope.__exit__(exc_type, exc_val, exc_tb)
         return False
 
+    def __str__(self):
+        return "TowerContext(name={}, is_training={})".format(
+            self._name, self._is_training)
+
 
 def get_current_tower_context():
     global _CurrentTowerContext

tensorpack/train/feedfree.py

@@ -6,7 +6,7 @@
 import tensorflow as tf
 
 from ..utils import log_deprecated
-from ..tfutils.tower import TowerContext
+from ..tfutils.tower import TowerContext, get_current_tower_context
 from .input_data import QueueInput, FeedfreeInput
 
 from .base import Trainer
@@ -27,8 +27,20 @@ class FeedfreeTrainerBase(Trainer):
             summary_str = self.summary_op.eval()
             self.add_summary(summary_str)
 
-    def _get_input_tensors(self):
-        return self._input_method.get_input_tensors()
+    def build_train_tower(self):
+        """
+        Get input tensors from `self.input_method` and build the graph.
+        """
+        def f():
+            inputs = self._input_method.get_input_tensors()
+            self.model.build_graph(inputs)
+        ctx = get_current_tower_context()
+        if ctx is None:
+            with TowerContext(''):
+                f()
+        else:
+            assert ctx.is_training, ctx
+            f()
 
     def _setup(self):
         assert isinstance(self._input_method, FeedfreeInput), type(self._input_method)
@@ -39,16 +51,15 @@ class SingleCostFeedfreeTrainer(FeedfreeTrainerBase):
     """ A feedfree Trainer which assumes a single cost. """
     def _get_cost_and_grad(self):
         """ get the cost and gradient"""
-        actual_inputs = self._get_input_tensors()
-        self.model.build_graph(actual_inputs)
-        cost_var = self.model.get_cost()
+        self.build_train_tower()
+        cost = self.model.get_cost()
         opt = self.config.optimizer
         # GATE_NONE faster?
         grads = opt.compute_gradients(
-            cost_var,
+            cost,
             gate_gradients=tf.train.Optimizer.GATE_NONE,
             colocate_gradients_with_ops=True)
-        return cost_var, grads
+        return cost, grads
 
     def run_step(self):
         """ Simply run ``self.train_op``, which minimizes the cost."""

tensorpack/train/multigpu.py

@@ -141,8 +141,7 @@ class SyncMultiGPUTrainer(MultiGPUTrainer,
             # grads = grad_list[0]
         else:
             def get_cost():
-                actual_inputs = self._get_input_tensors()
-                self.model.build_graph(actual_inputs)
+                self.build_train_tower()
                 return self.model.get_cost()
 
             cost_list = MultiGPUTrainer._multi_tower_costs(