add meta data link and script to convert lmdb

data/README.md

@@ -55,10 +55,4 @@ You can directly download the feature and test the pre-trained model.
 ### Meta Data
 We also provide the Meta data for ViLBERT, you can download them and run the code more eailiy. 
 
-|Task|                             Link                             |
-| :-------------: | :---------------------------------------------: |
-|       VQA       | |
-|       VCR       | |
-|    RefCOCO+     | |
-| Flickr30k | |
-
+[Google drive](https://drive.google.com/drive/folders/1o7sCLl1_PKCoaGvigCr_uGuBg6koOJm8?usp=sharing)

script/convert_lmdb_flickr30k.py

@@ -0,0 +1,38 @@
+import h5py
+import os
+import pdb
+import numpy as np
+import json
+import sys
+FIELDNAMES = ['image_id', 'image_w','image_h','num_boxes', 'boxes', 'features', 'cls_prob']
+import csv
+import base64
+import pickle
+import lmdb # install lmdb by "pip install lmdb"
+
+csv.field_size_limit(sys.maxsize)
+
+count = 0
+num_file = 1
+name = '/srv/share2/jlu347/bottom-up-attention/feature/flickr30k/flickr30k_test_resnet101_faster_rcnn_genome.tsv.%d'
+infiles = [name % i for i in range(num_file)]
+# infiles.append('/srv/share2/jlu347/bottom-up-attention/feature/coco/coco_val_resnet101_faster_rcnn_genome.tsv.0')
+
+save_path = os.path.join('flickr30k_test_resnet101_faster_rcnn_genome1.lmdb')
+env = lmdb.open(save_path, map_size=1099511627776)
+
+id_list = []
+with env.begin(write=True) as txn:
+    for infile in infiles:
+        with open(infile) as tsv_in_file:
+            reader = csv.DictReader(tsv_in_file, delimiter='\t', fieldnames = FIELDNAMES)
+            for item in reader:
+                img_id = str(item['image_id']).encode()
+                id_list.append(img_id)
+                txn.put(img_id, pickle.dumps(item))
+                if count % 1000 == 0:
+                    print(count)
+                count += 1
+    txn.put('keys'.encode(), pickle.dumps(id_list))
+
+print(count)

script/convert_lmdb_refcoco+.py

@@ -0,0 +1,39 @@
+import h5py
+import os
+import pdb
+import numpy as np
+import json
+import sys
+FIELDNAMES = ['image_id', 'image_w','image_h','num_boxes', 'boxes', 'features', 'cls_prob']
+import csv
+import base64
+csv.field_size_limit(sys.maxsize)
+import sys
+import pickle
+import lmdb # install lmdb by "pip install lmdb"
+
+count = 0
+num_file = 1
+name = '/srv/share2/jlu347/bottom-up-attention/feature/refcoco_unc/refcoco+_unc_resnet101_faster_rcnn_genome.tsv.%d'
+infiles = [name % i for i in range(num_file)]
+
+save_path = os.path.join('refcoco+.lmdb')
+env = lmdb.open(save_path, map_size=1099511627776)
+
+id_list = []
+with env.begin(write=True) as txn:
+    for infile in infiles:
+        with open(infile) as tsv_in_file:
+            reader = csv.DictReader(tsv_in_file, delimiter='\t', fieldnames = FIELDNAMES)
+            for item in reader:
+
+                img_id = str(item['image_id']).encode()
+                id_list.append(img_id)
+                txn.put(img_id, pickle.dumps(item))
+
+                if count % 1000 == 0:
+                    print(count)
+                count += 1
+    txn.put('keys'.encode(), pickle.dumps(id_list))
+
+print(count)

script/convert_lmdb_refcoco+_gt.py

@@ -0,0 +1,39 @@
+import h5py
+import os
+import pdb
+import numpy as np
+import json
+import sys
+FIELDNAMES = ['image_id', 'image_w','image_h','num_boxes', 'boxes', 'features', 'cls_prob']
+import csv
+import base64
+csv.field_size_limit(sys.maxsize)
+import sys
+import pickle
+import lmdb # install lmdb by "pip install lmdb"
+
+count = 0
+num_file = 1
+name = '/srv/share2/jlu347/bottom-up-attention/feature/refcoco_unc/refcoco+_unc_gt_resnet101_faster_rcnn_genome.tsv.%d'
+infiles = [name % i for i in range(num_file)]
+
+save_path = os.path.join('refcoco+_gt.lmdb')
+env = lmdb.open(save_path, map_size=1099511627776)
+
+id_list = []
+with env.begin(write=True) as txn:
+    for infile in infiles:
+        with open(infile) as tsv_in_file:
+            reader = csv.DictReader(tsv_in_file, delimiter='\t', fieldnames = FIELDNAMES)
+            for item in reader:
+
+                img_id = str(item['image_id']).encode()
+                id_list.append(img_id)
+                txn.put(img_id, pickle.dumps(item))
+
+                if count % 1000 == 0:
+                    print(count)
+                count += 1
+    txn.put('keys'.encode(), pickle.dumps(id_list))
+
+print(count)

script/convert_test_lmdb.py

@@ -0,0 +1,35 @@
+import h5py
+import os
+import pdb
+import numpy as np
+import json
+import sys
+FIELDNAMES = ['image_id', 'image_w','image_h','num_boxes', 'boxes', 'features']
+import csv
+import base64
+import pickle
+import lmdb # install lmdb by "pip install lmdb"
+
+csv.field_size_limit(sys.maxsize)
+
+name = '/srv/share2/jlu347/bottom-up-attention/feature/coco/test2015/test2015_resnet101_faster_rcnn_genome.tsv'
+infiles = [name]
+
+save_path = os.path.join('coco_test_resnet101_faster_rcnn_genome.lmdb')
+env = lmdb.open(save_path, map_size=1099511627776)
+
+id_list = []
+count = 0
+with env.begin(write=True) as txn:
+    for infile in infiles:
+        with open(infile) as tsv_in_file:
+            reader = csv.DictReader(tsv_in_file, delimiter='\t', fieldnames = FIELDNAMES)
+            for item in reader:
+                img_id = str(item['image_id']).encode()
+                id_list.append(img_id)
+                txn.put(img_id, pickle.dumps(item))
+                if count % 1000 == 0:
+                    print(count)
+                count += 1
+    txn.put('keys'.encode(), pickle.dumps(id_list))
+

script/convert_trainval_lmdb.py

@@ -0,0 +1,41 @@
+import h5py
+import os
+import pdb
+import numpy as np
+import json
+import sys
+FIELDNAMES = ['image_id', 'image_w','image_h','num_boxes', 'boxes', 'features']
+import csv
+import base64
+import pickle
+import lmdb # install lmdb by "pip install lmdb"
+
+csv.field_size_limit(sys.maxsize)
+
+count = 0
+infiles = []
+
+path = '/srv/share2/jlu347/bottom-up-attention/feature/coco/trainval/'
+infiles.append(path + 'karpathy_train_resnet101_faster_rcnn_genome.tsv.0')
+infiles.append(path + 'karpathy_train_resnet101_faster_rcnn_genome.tsv.1')
+infiles.append(path + 'karpathy_val_resnet101_faster_rcnn_genome.tsv')
+infiles.append(path + 'karpathy_test_resnet101_faster_rcnn_genome.tsv')
+
+save_path = os.path.join('coco_trainval_resnet101_faster_rcnn_genome.lmdb')
+env = lmdb.open(save_path, map_size=1099511627776)
+
+id_list = []
+with env.begin(write=True) as txn:
+    for infile in infiles:
+        with open(infile) as tsv_in_file:
+            reader = csv.DictReader(tsv_in_file, delimiter='\t', fieldnames = FIELDNAMES)
+            for item in reader:
+                img_id = str(item['image_id']).encode()
+                id_list.append(img_id)
+                txn.put(img_id, pickle.dumps(item))
+                if count % 1000 == 0:
+                    print(count)
+                count += 1
+    txn.put('keys'.encode(), pickle.dumps(id_list))
+
+print(count)