update docs

docs/tutorial/inference.md

@@ -63,7 +63,7 @@ The example in [examples/basics/export-model.py](../examples/basics/export-model
 ### Exporter
 
 In addition to the standard checkpoint format tensorpack saved for you during training,
-you can also save your models into other formats so it may be more friendly for inference.
+you can also save your models into other formats after training, so it may be more friendly for inference.
 
 1. Export to `SavedModel` format for TensorFlow Serving:
 
@@ -133,7 +133,7 @@ with TowerContext('', is_training=False):
 ```
 
 ```eval_rst
-.. note:: **Do not use metagraph for inference!**. 
+.. note:: **Do not use metagraph for inference!**
 
 	Metagraph is the wrong abstraction for a "model". 
 	It stores the entire graph which contains not only the mathematical model, but also all the
@@ -155,7 +155,7 @@ with TowerContext('', is_training=False):
 ### Step 2: load the checkpoint
 
 You can just use `tf.train.Saver` for all the work.
-Alternatively, use tensorpack's `SaverRestore(path).init(tf.get_default_session())`
+Alternatively, use tensorpack's `get_model_loader(path).init(tf.get_default_session())`
 
 Now, you've already built a graph for inference, and the checkpoint is loaded. 
 You may now:

docs/tutorial/training-interface.md

@@ -62,8 +62,10 @@ The function `launch_train_with_config(config, trainer)`
 uses the raw trainer interface and is almost equivalent to the following two lines of code:
 ```python
 trainer.setup_graph(
-    my_model.get_inputs_desc(), my_input_source,  # or QueueInput(my_dataflow)
-    my_model.build_graph, my_model.get_optimizer)
+    my_model.get_inputs_desc(),
+    my_input_source,  # or QueueInput(my_dataflow)
+    my_model.build_graph,
+    my_model.get_optimizer)
 trainer.train_with_defaults(
     callbacks=config.callbacks,
     monitors=config.monitors,
@@ -76,7 +78,8 @@ trainer.train_with_defaults(
 ```
 If you need more control (e.g., if you want to construct the callbacks after
 setting up the graph), you can write the above two lines by yourself instead.
-
+You don't need to construct a `TrainConfig` any more in that case.
+The function `launch_train_with_config` exists mainly for historical reasons.
 
 ### Keras Interface
 

examples/FasterRCNN/NOTES.md

@@ -67,9 +67,34 @@ Speed:
 
 Possible Future Enhancements:
 
-1. Define a better interface to load custom dataset.
+1. Define a better interface to load different datasets.
 
-1. Support batch>1 per GPU.
+1. Support batch>1 per GPU. Batching with inconsistent shapes is
+   non-trivial to implement in TensorFlow.
 
 1. Use dedicated ops to improve speed. (e.g. a TF implementation of ROIAlign op
    can be found in [light-head RCNN](https://github.com/zengarden/light_head_rcnn/tree/master/lib/lib_kernel))
+
+
+### TensorFlow version notes
+
+TensorFlow ≥ 1.6 supports most common features in this R-CNN implementation.
+However, each version of TensorFlow has bugs that I either reported or fixed,
+and this implementation touches many of those bugs.
+Therefore, not every version of TF ≥ 1.6 supports every feature in this implementation.
+
+This implementation contains workaround for some of those TF bugs.
+However, note that the workaround needs to check your TF version by `tf.VERSION`
+and may not detect bugs properly if your TF version is not an official release
+(e.g., if you use a nighly build).
+
+1. TF < 1.6: Nothing works due to lack of support for empty tensors
+   ([PR](https://github.com/tensorflow/tensorflow/pull/15264))
+   and `FrozenBN` training
+   ([PR](https://github.com/tensorflow/tensorflow/pull/12580)).
+1. TF < 1.10: `SyncBN` with NCCL will fail ([PR](https://github.com/tensorflow/tensorflow/pull/20360)).
+1. TF 1.11 & 1.12: multithread inference will fail ([issue](https://github.com/tensorflow/tensorflow/issues/22750)).
+   Latest tensorpack will apply a workaround.
+1. TF > 1.12: MKL inference will fail ([issue](https://github.com/tensorflow/tensorflow/issues/24650)).
+1. TF > 1.12: Horovod training will fail ([issue](https://github.com/tensorflow/tensorflow/issues/25946)).
+   Latest tensorpack will apply a workaround.