diff --git a/.gitignore b/.gitignore
index 6522d78b5..4c40f1924 100644
--- a/.gitignore
+++ b/.gitignore
@@ -59,3 +59,4 @@ docs/_build/
 
 # PyBuilder
 target/
+*.dat
diff --git a/examples/ResNet/README.md b/examples/ResNet/README.md
index 86629478b..3b5324af1 100644
--- a/examples/ResNet/README.md
+++ b/examples/ResNet/README.md
@@ -4,4 +4,7 @@
 Implements the paper "Deep Residual Learning for Image Recognition", [http://arxiv.org/abs/1512.03385](http://arxiv.org/abs/1512.03385)
 with the variants proposed in "Identity Mappings in Deep Residual Networks", [https://arxiv.org/abs/1603.05027](https://arxiv.org/abs/1603.05027).
 
+The train error shown here is a moving average of the error rate of each batch in training.
+The validation error here is computed on test set.
+
 ![cifar10](https://github.com/ppwwyyxx/tensorpack/raw/master/examples/ResNet/cifar10-resnet.png)
diff --git a/requirements.txt b/requirements.txt
index 8a054c562..7986dfc9f 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -3,3 +3,4 @@ pillow
 scipy
 tqdm
 h5py
+nltk
diff --git a/tensorpack/dataflow/dataset/visualqa.py b/tensorpack/dataflow/dataset/visualqa.py
new file mode 100644
index 000000000..98574606c
--- /dev/null
+++ b/tensorpack/dataflow/dataset/visualqa.py
@@ -0,0 +1,70 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+# File: visualqa.py
+# Author: Yuxin Wu <ppwwyyxxc@gmail.com>
+
+from ..base import DataFlow
+from six.moves import zip, map
+from collections import Counter
+import json
+
+__all__ = ['VisualQA']
+
+# TODO shuffle
+class VisualQA(DataFlow):
+    """
+    Visual QA dataset. See http://visualqa.org/
+    Simply read q/a json file and produce q/a pairs in their original format.
+    """
+    def __init__(self, question_file, annotation_file):
+        qobj = json.load(open(question_file))
+        self.task_type = qobj['task_type']
+        self.questions = qobj['questions']
+        self._size = len(self.questions)
+
+        aobj = json.load(open(annotation_file))
+        self.anno = aobj['annotations']
+        assert len(self.anno) == len(self.questions), \
+            "{}!={}".format(len(self.anno), len(self.questions))
+        self._clean()
+
+    def _clean(self):
+        for a in self.anno:
+            for aa in a['answers']:
+                del aa['answer_id']
+
+    def size(self):
+        return self._size
+
+    def get_data(self):
+        for q, a in zip(self.questions, self.anno):
+            assert q['question_id'] == a['question_id']
+            yield [q, a]
+
+    def get_common_answer(self, n):
+        """ Get the n most common answers (could be phrases) """
+        cnt = Counter()
+        for anno in self.anno:
+            cnt[anno['multiple_choice_answer']] += 1
+        return [k[0] for k in cnt.most_common(n)]
+
+    def get_common_question_words(self, n):
+        """
+        Get the n most common words in questions
+        """
+        from nltk.tokenize import word_tokenize # will need to download 'punckt'
+        cnt = Counter()
+        for q in self.questions:
+            cnt.update(word_tokenize(q['question'].lower()))
+        del cnt['?']    # probably don't need this
+        ret = cnt.most_common(n)
+        return [k[0] for k in ret]
+
+if __name__ == '__main__':
+    vqa = VisualQA('/home/wyx/data/VQA/MultipleChoice_mscoco_train2014_questions.json',
+            '/home/wyx/data/VQA/mscoco_train2014_annotations.json')
+    for k in vqa.get_data():
+        #print json.dumps(k)
+        break
+    vqa.get_common_question_words(100)
+    #from IPython import embed; embed()