* Update BERT example README

* Update CHANGELOG

Author: ZhitingHu

CHANGELOG.md

@@ -4,6 +4,7 @@
 ### New features
 
 * Add support for GPT-2 345M model in [examples/gpt-2](https://github.com/asyml/texar/tree/master/examples/gpt-2). ([#156](https://github.com/asyml/texar/pull/156)) 
+* Add BERT modules, including `texar.modules.BERTEncoder` ([doc](https://texar.readthedocs.io/en/latest/code/modules.html#texar.modules.BertEncoder)) and `texar.modules.BERTClassifier` ([doc](https://texar.readthedocs.io/en/latest/code/modules.html#bertclassifierv)). ([#167](https://github.com/asyml/texar/pull/167))
 
 ### Feature improvements
 
@@ -16,7 +17,8 @@
 
 * Fix `texar.losses.reduce_batch_time` when `sequence` has dtype other than `tf.float32`. ([#143](https://github.com/asyml/texar/issues/143))
 * Fix `texar.losses.reduce_dimensions` when `average_axes` or `sum_axes` is `int`. ([#141](https://github.com/asyml/texar/pull/141))
-* Fix GPT-2 tokenization loading path. ([165](https://github.com/asyml/texar/pull/165))
+* Fix [GPT-2](https://github.com/asyml/texar/tree/master/examples/gpt-2) tokenization loading path. ([165](https://github.com/asyml/texar/pull/165))
+* Fix [examples/vae_text](https://github.com/asyml/texar/tree/master/examples/vae_text) EOS bug. ([168](https://github.com/asyml/texar/pull/168)) 
 
 ## [v0.2.0](https://github.com/asyml/texar/releases/tag/v0.2.0) (2019-04-09)
 

examples/bert/README.md

@@ -2,7 +2,10 @@
 
 This is a Texar implementation of Google's BERT model, which allows to load pre-trained model parameters downloaded from the [official release](https://github.com/google-research/bert) and build/fine-tune arbitrary downstream applications with **distributed training** (This example showcases BERT for sentence classification).
 
-With Texar, building the BERT model is as simple as creating a [`TransformerEncoder`](https://texar.readthedocs.io/en/latest/code/modules.html#transformerencoder) instance. We can initialize the parameters of the TransformerEncoder using a pre-trained BERT checkpoint by calling `init_bert_checkpoint(path_to_bert_checkpoint)`. 
+This example shows two ways of building a BERT classifier, at different abstraction levels:
+
+  * Use `texar.modules.BERTClassifier` ([doc](https://texar.readthedocs.io/en/latest/code/modules.html#texar.modules.BertClassifier)) directly. The module supports both sequence classification (one label per sequence) and sequence labeling (one label per token). --- See `bert_classifier_main_v2.py` for implementation.
+  * Use lower-level modules by creating a `TransformerEncoder` ([doc](https://texar.readthedocs.io/en/latest/code/modules.html#transformerencoder)) instance and adding additional layers. Initialization with a pre-trained BERT checkpoint is done by calling `init_bert_checkpoint(path_to_bert_checkpoint)`. --- See `bert_classifier_main.py` for implementation. 
 
 In sum, this example showcases:
 
@@ -77,6 +80,8 @@ Here:
 - `config_data`: The data configuration. See the default [`config_data.py`](./config_data.py) for example. Make sure to specify `num_classes`, `num_train_data`, `max_seq_length`, and `tfrecord_data_dir` as used or output in the above [data preparation](#prepare-data) step.
 - `output_dir`: The output path where checkpoints and TensorBoard summaries are saved.
 
+*[NOTE: you can also use `bert_classifier_main_v2.py` in the above]*
+
 For **Multi-GPU training** on one or multiple machines, you may first install the prerequisite OpenMPI and Hovorod packages, as detailed in the [distributed_gpu](https://github.com/asyml/texar/tree/master/examples/distributed_gpu) example. 
 
 Then run the following cmd for training and evaluation. The cmd trains the model on local with 2 GPUs. Evaluation is performed with the single rank-0 GPU.