parse MSCOCO and Flickr captions for V-PCFG

xcfg/data/vpcfg/flickr/README.md

@@ -0,0 +1,5 @@
+run `flickr_dl.sh' to prepare data.
+
+2016 test-split in `https://github.com/multi30k/dataset/tree/master/data/task1/image_splits'
+
+image and caption data are from `http://hockenmaier.cs.illinois.edu/DenotationGraph/data/'

xcfg/data/vpcfg/flickr/flickr_dl.sh

@@ -0,0 +1,7 @@
+#wget https://raw.githubusercontent.com/multi30k/dataset/master/data/task1/image_splits/test_2016_flickr.txt -O test_ids.txt
+#wget https://raw.githubusercontent.com/multi30k/dataset/master/data/task1/image_splits/train.txt -O train_ids.txt
+#wget https://raw.githubusercontent.com/multi30k/dataset/master/data/task1/image_splits/val.txt -O val_ids.txt
+
+wget http://hockenmaier.cs.illinois.edu/DenotationGraph/data/flickr30k.tar.gz
+tar -zxvf flickr30k.tar.gz
+rm flickr30k.tar.gz

xcfg/data/vpcfg/mscoco/README.md

@@ -0,0 +1,9 @@
+run `mscoco_dl.sh' to get mscoco annotation data.
+
+`train_ids.txt' is partitioned into 4 parts so that we can parse faster by parsing the parts in parallel.
+
+the splits seem to come from [train,test,val].txt in `coco_precomp' of `wget http://www.cs.toronto.edu/~faghri/vsepp/data.tar' 
+
+the top 1000 image ids in the [test,val].txt are used.
+
+see https://github.com/fartashf/vsepp#download-data 

xcfg/data/vpcfg/mscoco/mscoco_dl.sh

@@ -0,0 +1,5 @@
+wget http://images.cocodataset.org/annotations/annotations_trainval2014.zip
+unzip annotations_trainval2014.zip
+rm annotations_trainval2014.zip
+rm annotations/instances_*
+rm annotations/person_*