update to tf 1.4

Author: Smellly

data-prepro/CUB200_preprocess/prepro_cub_annotation.py

@@ -40,8 +40,8 @@
             caps.append(tmp)
 
         test_data[i] = caps
-# json.dump(test_data, open('cub_dataset/K_test_annotation.json', 'w'))
-pk.dump(test_data, open('cub_dataset/K_test_annotation.pkl', 'w'))
+json.dump(test_data, open('cub_data/K_test_annotation.json', 'w'))
+pk.dump(test_data, open('cub_data/K_test_annotation.pkl', 'w'))
 
 print("Processing train_data")
 for i in dataset['train_id']:
@@ -66,8 +66,8 @@
         train_data[i] = caps
 print 'number of skip train data: ' + str(skip_num)
 [u'info', u'images', u'licenses', u'type', u'annotations']
-# json.dump(test_data, open('cub_dataset/K_train_annotation.json', 'w'))
-pk.dump(train_data, open('cub_dataset/K_train_annotation.pkl', 'w'))
+json.dump(test_data, open('cub_data/K_train_annotation.json', 'w'))
+pk.dump(train_data, open('cub_data/K_train_annotation.pkl', 'w'))
 
 print("Processing val_data")
 for i in dataset['val_id']:
@@ -89,4 +89,5 @@
             caps.append(tmp)
 
         val_data[i] = caps
-pk.dump(val_data, open('cub_dataset/K_val_annotation.pkl', 'w'))
+pk.dump(val_data, open('cub_data/K_val_annotation.pkl', 'w'))
+json.dump(test_data, open('cub_data/K_val_annotation.json', 'w'))

data-prepro/MSCOCO_preprocess/extract_resnet_cub.py

@@ -81,15 +81,16 @@ def extract_image(net, image_file):
 
 def split(split, net, feat_dict):
         print 'load ' + split 
-	img_dir = '/home/smelly/projects/show-adapt-and-tell/data-prepro/CUB200_preprocess/cub_dataset/CUB_200_2011'
-	img_path = os.path.join(img_dir, split)
+	img_dir = '/home/smelly/projects/show-adapt-and-tell/data-prepro/CUB200_preprocess/cub_data/CUB_200_2011/images'
+        text_path = '/home/smelly/projects/show-adapt-and-tell/data-prepro/CUB200_preprocess/ECCV16_explanations_splits' 
+	# img_path = os.path.join(img_dir, split)
 	# img_list = os.listdir(img_path)
-        with open(os.path.join(img_dir, 'images.txt'), 'r') as f:
+        with open(os.path.join(text_path, split + '.txt'), 'r') as f:
             img_list = [x.split()[-1] for x in f.readlines()]
 	pool5_list = []
 	prob_list = []
         for k in tqdm(img_list):
-		blobs_out_pool5 = extract_image(net, os.path.join(img_path,k))
+		blobs_out_pool5 = extract_image(net, os.path.join(img_dir, k))
 		feat_dict[k.split('.')[0]] = np.array(blobs_out_pool5)
 
 	return feat_dict
@@ -107,6 +108,10 @@ def split(split, net, feat_dict):
 	net.name = os.path.splitext(os.path.basename(args.caffemodel))[0]
 
 	feat_dict = {}
-	split('images', net, feat_dict)
+	split('train_noCub', net, feat_dict)
+	split('val', net, feat_dict)
 	pk.dump(feat_dict, open('./cub_data/cub_trainval_feat.pkl','w'))
+	feat_dict = {}
+	split('test', net, feat_dict)
+	pk.dump(feat_dict, open('./cub_data/cub_test_feat.pkl','w'))
 

show-adapt-tell/data_loader.py

@@ -110,7 +110,7 @@ def __init__(self, conf=None):
         # train img feature
         self.dataset_name = 'cub'
         # target data
-        flickr_img_path = './cub/cub_train_resnet.pkl'
+        flickr_img_path = './cub/cub_trainval_feat.pkl'
         self.train_flickr_img_feat = utils.unpickle(flickr_img_path)
         self.num_train_images_filckr = len(self.train_flickr_img_feat.keys())
         self.train_img_idx = self.train_flickr_img_feat.keys()
@@ -119,17 +119,19 @@ def __init__(self, conf=None):
         self.flickr_caption_train = flickr_caption_train_data['tokenized_caption_list']
         self.flickr_caption_idx_train = flickr_caption_train_data['filename_list']
         self.num_flickr_train_caption = self.flickr_caption_train.shape[0]
-        flickr_img_path = './cub/cub_test_resnet.pkl'
+        flickr_img_path = './cub/cub_test_feat.pkl'
         self.test_flickr_img_feat = utils.unpickle(flickr_img_path)
         self.flickr_random_shuffle()    # shuffle the text data
 
         # MSCOCO data
         img_feat_path = './data/coco_trainval_feat.pkl'
         self.img_feat = utils.unpickle(img_feat_path)
         train_meta_path = './data/K_annotation_train2014.pkl'
+        # train_meta_path = './data/K_train_annotation.pkl'
         train_meta = utils.unpickle(train_meta_path)
         self.filename2id = train_meta['filename2id']
 	val_meta_path = './data/K_annotation_val2014.pkl'
+	# val_meta_path = './data/K_val_annotation.pkl'
 	val_meta = utils.unpickle(val_meta_path)
 	self.id2filename = val_meta['id2filename']
         # train caption

show-adapt-tell/highway.py

@@ -57,7 +57,9 @@ def main(_):
 	if FLAGS.G_is_pretrain:
 	    G_pretrained_model = G_pretrained(sess, dataset, conf=FLAGS)
 	    if FLAGS.is_train:
+                print('G_pretrained_model train')
 		G_pretrained_model.train()
+            print('G_pretrained_model evaluate')
 	    G_pretrained_model.evaluate('test', 0, )
 	if FLAGS.D_is_pretrain:
 	    negative_dataset = mscoco_negative(dataset, FLAGS)