lut & update vqa

scripts/dump_train_config.py

@@ -49,6 +49,7 @@ with tqdm.tqdm(total=NR_DP_TEST, leave=True, unit='data points') as pbar:
         if idx > NR_DP_TEST:
             break
         pbar.update()
+    from IPython import embed; embed()
 
 
 

tensorpack/dataflow/common.py

@@ -16,7 +16,7 @@ class BatchData(ProxyDataFlow):
         """
         Group data in `ds` into batches.
 
-        :param ds: a DataFlow instance
+        :param ds: a DataFlow instance. Its component must be either a scalar or a numpy array
         :param remainder: whether to return the remaining data smaller than a batch_size.
             If set True, will possibly return a data point of a smaller 1st dimension.
             Otherwise, all generated data are guranteed to have the same size.

tensorpack/dataflow/dataset/visualqa.py

@@ -4,6 +4,7 @@
 # Author: Yuxin Wu <ppwwyyxxc@gmail.com>
 
 from ..base import DataFlow
+from ...utils import *
 from six.moves import zip, map
 from collections import Counter
 import json
@@ -17,16 +18,17 @@ class VisualQA(DataFlow):
     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)
+        with timed_operation('Reading VQA JSON 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()
+            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:
@@ -42,15 +44,17 @@ class VisualQA(DataFlow):
             yield [q, a]
 
     def get_common_answer(self, n):
-        """ Get the n most common answers (could be phrases) """
+        """ Get the n most common answers (could be phrases)
+            n=3000 ~= thresh 4
+        """
         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
+        """ Get the n most common words in questions
+            n=4600 ~= thresh 6
         """
         from nltk.tokenize import word_tokenize # will need to download 'punckt'
         cnt = Counter()
@@ -64,7 +68,8 @@ 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)
+        print json.dumps(k)
         break
-    vqa.get_common_question_words(100)
+#    vqa.get_common_question_words(100)
+    vqa.get_common_answer(100)
     #from IPython import embed; embed()

tensorpack/dataflow/prefetch.py

@@ -7,6 +7,7 @@ import multiprocessing
 from six.moves import range
 from .base import ProxyDataFlow
 from ..utils.concurrency import ensure_procs_terminate
+from ..utils import logger
 
 __all__ = ['PrefetchData']
 

tensorpack/utils/lut.py

+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+# File: lut.py
+# Author: Yuxin Wu <ppwwyyxxc@gmail.com>
+
+import six
+
+__all__ = ['LookUpTable']
+
+class LookUpTable(object):
+    def __init__(self, objlist):
+        self.idx2obj = dict(enumerate(objlist))
+        self.obj2idx = {v : k for k, v in six.iteritems(self.idx2obj)}

tensorpack/utils/utils.py

@@ -11,8 +11,8 @@ import numpy as np
 
 from . import logger
 
-__all__ = ['timed_operation', 'change_env', 'get_rng', 'memoized',
-           'get_nr_gpu']
+__all__ = ['timed_operation', 'change_env',
+        'get_rng', 'memoized', 'get_nr_gpu']
 
 #def expand_dim_if_necessary(var, dp):
 #    """