update docs

docs/tutorial/trainer.md

@@ -58,7 +58,7 @@ implement different distribution strategies.
 They take care of device placement, gradient averaging and synchronoization
 in the efficient way and all reach the same performance as the
 [official TF benchmarks](https://www.tensorflow.org/performance/benchmarks).
-It takes only one line of code change to use them, i.e. `trainer=SyncMultiGPUTrainerReplicated()`.
+It takes only one line of code change to use them, e.g. `trainer=SyncMultiGPUTrainerReplicated(...)`.
 
 Note some __common problems__ when using these trainers:
 

docs/tutorial/training-interface.md

@@ -7,8 +7,7 @@ when you don't want to customize too much.
 
 ### With ModelDesc and TrainConfig
 
-This is an interface that's most familiar to old tensorpack users,
-and is used for single-cost tasks only.
+This interface is enough for most types of single-cost tasks.
 A lot of examples are written in this interface.
 
 [SingleCost trainers](../modules/train.html#tensorpack.train.SingleCostTrainer)
@@ -20,24 +19,22 @@ class MyModel(ModelDesc):
   def inputs(self):
     return [tf.placeholder(dtype, shape, name), tf.placeholder(dtype, shape, name), ... ]
 
-  def _build_graph(self, inputs):
-    tensorA, tensorB = inputs
+  def build_graph(self, tensorA, tensorB, ...):  # inputs
     # build the graph
-    self.cost = xxx   # define the cost tensor
+    return cost   # returns the cost tensor
 
-  def _get_optimizer(self):
+  def optimizer(self):
     return tf.train.GradientDescentOptimizer(0.1)
 ```
 
-`_get_inputs` should define the metainfo of all the inputs your graph will take to build.
+`inputs` should define the metainfo of all the inputs your graph will take to build.
 
-`_build_graph` takes a list of `inputs` tensors which will match `_get_inputs`.
+`build_graph` takes inputs tensors that matches what you've defined in `inputs()`.
 
-You can use any symbolic functions in `_build_graph`, including TensorFlow core library
+You can use any symbolic functions in `build_graph`, including TensorFlow core library
 functions and other symbolic libraries.
-`_build_graph` will be the tower function,
-so you need to follow [some rules](trainer.md#tower-trainer).
-You also need to set `self.cost` in this function.
+`build_graph` will be the tower function, so you need to follow [some rules](trainer.md#tower-trainer).
+Because this interface is for single-cost training, you need to return the cost tensor.
 
 After defining such a model, use it with `TrainConfig` and `launch_train_with_config`:
 
@@ -61,18 +58,23 @@ and
 [launch_train_with_config](../modules/train.html#tensorpack.train.launch_train_with_config)
 for detailed functionalities.
 
+### Keras Interface
+
+Some wrappers were made on top of tensorpack trainers, to create a Keras-like
+interface. See [Tensorpack+Keras examples](../examples/keras) for details.
+
 ### Raw Trainer Interface
 
-To get a lower-level control, you can also access methods of trainer directly:
+To get a lower-level control, you can also access trainer methods directly:
 
-__Build__ the graph: For general trainer, build the graph by yourself.
-For single-cost trainer, build the graph by
+__Build the graph__:
+For single-cost trainers, build the graph by calling
 [SingleCostTrainer.setup_graph](../modules/train.html#tensorpack.train.SingleCostTrainer.setup_graph).
 
-__Run__ the iterations: Call
+__Start training__: Call
 [Trainer.train()](../modules/train.html#tensorpack.train.Trainer.train),
 or
 [Trainer.train_with_defaults()](../modules/train.html#tensorpack.train.Trainer.train_with_defaults)
-which applies some defaults options for normal use cases.
+which applies some defaults options for common use cases.
 
-Read the API documentation for detail usage.
+Read their API documentation for detail usage.

examples/FasterRCNN/model_fpn.py

@@ -139,15 +139,15 @@ def multilevel_rpn_losses(
     assert len(multilevel_box_logits) == num_lvl
 
     losses = []
+    with tf.name_scope('rpn_losses'):
         for lvl in range(num_lvl):
-        with tf.name_scope('RPNLoss_Lvl{}'.format(lvl + 2)):
+            with tf.name_scope('Level{}'.format(lvl + 2)):
                 anchors = multilevel_anchors[lvl]
                 label_loss, box_loss = rpn_losses(
                     anchors.gt_labels, anchors.encoded_gt_boxes(),
                     multilevel_label_logits[lvl], multilevel_box_logits[lvl])
                 losses.extend([label_loss, box_loss])
 
-    with tf.name_scope('rpn_losses'):
         total_label_loss = tf.add_n(losses[::2], name='label_loss')
         total_box_loss = tf.add_n(losses[1::2], name='box_loss')
         add_moving_summary(total_label_loss, total_box_loss)

examples/FasterRCNN/model_rpn.py

@@ -37,7 +37,6 @@ def rpn_head(featuremap, channel, num_anchors):
     return label_logits, box_logits
 
 
-@under_name_scope()
 def rpn_losses(anchor_labels, anchor_boxes, label_logits, box_logits):
     """
     Args:

examples/FasterRCNN/train.py

@@ -200,6 +200,7 @@ class ResNetC4Model(DetectionModel):
 
         if is_training:
             # rpn loss
+            with tf.name_scope('rpn_losses'):
                 rpn_label_loss, rpn_box_loss = rpn_losses(
                     anchors.gt_labels, anchors.encoded_gt_boxes(), rpn_label_logits, rpn_box_logits)
 

examples/PennTreebank/PTB-LSTM.py

@@ -169,11 +169,11 @@ def get_config():
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.')
+    parser.add_argument('--gpu', type=int, help='the GPU to use')
     parser.add_argument('--load', help='load model')
     args = parser.parse_args()
     if args.gpu:
-        os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
+        os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu)
 
     config = get_config()
     if args.load:

tensorpack/callbacks/hooks.py

@@ -29,7 +29,7 @@ class CallbackToHook(tf.train.SessionRunHook):
 class HookToCallback(Callback):
     """
     Make a ``tf.train.SessionRunHook`` into a callback.
-    Note that the `coord` argument in `after_create_session` will be None.
+    Note that when `SessionRunHook.after_create_session` is called, the `coord` argument will be None.
     """
 
     _chief_only = False

tensorpack/callbacks/trigger.py

@@ -9,15 +9,13 @@ __all__ = ['PeriodicTrigger', 'PeriodicCallback', 'EnableCallbackIf']
 
 class PeriodicTrigger(ProxyCallback):
     """
-    Schedule to trigger a callback every k global steps or every k epochs by its :meth:`trigger()` method.
+    Trigger a callback every k global steps or every k epochs by its :meth:`trigger()` method.
     Most existing callbacks which do something every epoch are implemented
     with :meth:`trigger()` method.
 
-    All other methods (``before/after_run``, ``trigger_step``, etc) are unaffected.
+    All other methods (``before/after_run``, ``trigger_step``, etc) of the input callback are unaffected.
     """
 
-    _chief_only = False
-
     def __init__(self, triggerable, every_k_steps=None, every_k_epochs=None):
         """
         Args:
@@ -59,8 +57,6 @@ class PeriodicRunHooks(ProxyCallback):
     All other methods are untouched.
     """
 
-    _chief_only = False
-
     def __init__(self, callback, every_k_steps):
         """
         Args:
@@ -88,7 +84,7 @@ class PeriodicRunHooks(ProxyCallback):
 
 class EnableCallbackIf(ProxyCallback):
     """
-    Enable ``{before,after}_epoch``, ``{before,after}_run``,
+    Enable the ``{before,after}_epoch``, ``{before,after}_run``,
     ``trigger_{epoch,step}``
     methods of a callback, only when some condition satisfies.
     The other methods are unaffected.
@@ -98,8 +94,6 @@ class EnableCallbackIf(ProxyCallback):
         ``pred`` will be evaluated only in ``before_run``.
     """
 
-    _chief_only = False
-
     def __init__(self, callback, pred):
         """
         Args:
@@ -142,19 +136,15 @@ class EnableCallbackIf(ProxyCallback):
 
 class PeriodicCallback(EnableCallbackIf):
     """
-    Make the calls to the following methods of a callback **less** frequent:
-    ``{before,after}_epoch``, ``{before,after}_run``, ``trigger_{epoch,step}``.
-
-    These methods will be enabled only when ``global_step % every_k_steps == 0`
+    The ``{before,after}_epoch``, ``{before,after}_run``, ``trigger_{epoch,step}``
+    methods of the given callback will be enabled only when ``global_step % every_k_steps == 0`
     or ``epoch_num % every_k_epochs == 0``. The other methods are unaffected.
 
-    Note that this can only makes a callback **less** frequent than before.
-    :class:`PeriodicTrigger` can
-    make a callback which supports :meth:`trigger()` method more frequent than before.
+    Note that this can only makes a callback **less** frequent than itself.
+    If you have a callback that by default runs every epoch by its :meth:`trigger()` method,
+    use :class:`PeriodicTrigger` to schedule it more frequent than itself.
     """
 
-    _chief_only = False
-
     def __init__(self, callback, every_k_steps=None, every_k_epochs=None):
         """
         Args: