add attribute file

.gitignore

@@ -1,11 +1,11 @@
-#忽略build文件夹
+# ignore folder
 build
 log
 
-# 忽略所有生成的备份文件
+# backup files
 *.*~
 
-#忽略所有.o 和 .a文件.
+#.o and .a files
 *.[oa]
 
 *.DS_Store

CUB.py

@@ -32,7 +32,7 @@ def load_data(data_folder, target_size=(224, 224), bounding_box=True):
     attribute_name_file = data_folder+'/attributes.txt'
     processed_attribute_file = data_folder+'/processed_attributes.txt'
     # train test split
-    split_rf = open(split_file,'rb')
+    split_rf = open(split_file,'r')
     train_test_list = []
     train_idx = []
     test_idx = []
@@ -47,7 +47,7 @@ def load_data(data_folder, target_size=(224, 224), bounding_box=True):
         i+=1
     split_rf.close()
     # bb
-    bb_rf = open(bb_file,'rb')
+    bb_rf = open(bb_file,'r')
     bb_list = []
     for line in bb_rf.readlines():
         strs = line.strip().split(' ')
@@ -56,7 +56,7 @@ def load_data(data_folder, target_size=(224, 224), bounding_box=True):
     bb_rf.close()
     # images
     i = 0
-    images_rf = open(images_file,'rb')
+    images_rf = open(images_file,'r')
     for line in images_rf.readlines():
         strs = line.strip().split(' ')
         img = image.load_img(data_folder+'/images/'+strs[1])
@@ -73,7 +73,7 @@ def load_data(data_folder, target_size=(224, 224), bounding_box=True):
             print(i,' images load.')
     images_rf.close()
     # label
-    label_rf = open(label_file,'rb')
+    label_rf = open(label_file,'r')
     for line in label_rf.readlines():
         strs = line.strip().split(' ')
         if(train_test_list[int(strs[0])-1]=='1'):
@@ -87,7 +87,7 @@ def load_data(data_folder, target_size=(224, 224), bounding_box=True):
     A_test = A_all[test_idx]
     # class attributes
     C_A = np.zeros((200,312))
-    class_attr_rf = open(class_attributes_file,'rb')
+    class_attr_rf = open(class_attributes_file,'r')
     i = 0
     for line in class_attr_rf.readlines():
         strs = line.strip().split(' ')

cub_demo.py

@@ -20,16 +20,18 @@
 from keras.utils import np_utils
 from keras import backend as K
 from keras.models import load_model
-from keras.utils.visualize_util import plot
-#from deep_learning_models.vgg16 import VGG16
+#from keras.utils.visualize_util import plot
 from keras.applications.vgg16 import VGG16
 from keras.applications.resnet50 import ResNet50
-import CUB
 import scipy.misc
 from sklearn import preprocessing
+import CUB
 
-# model config
+# args
 net = sys.argv[1]
+data_folder = sys.argv[2]
+
+# model config
 flag_test = False
 batch_size = 10
 nb_epoch = 10
@@ -44,12 +46,11 @@
 
 # dataset config
 dataset = 'CUB'
-data_folder = '/home/hankai/data/CUB_200_2011'
 nb_classes = 200
 nb_attributes = [10, 16, 16, 16, 5, 16, 7, 16, 12, 16, 16, 15, 4, 16, 16, 16, 16, 6, 6, 15, 5, 5, 5, 16, 16, 16, 16, 5]
 img_rows, img_cols = 448, 448
 L = 14*14
-lr_list = [0.001,0.001,0.001,0.001,0.001,0.001,0.0001]
+lr_list = [0.001,0.003,0.001,0.001,0.001,0.001,0.001,0.0001]
 
 def init_classification(input_fea_map, dim_channel, nb_class, name=None):
     # conv

readme.md

@@ -4,23 +4,25 @@ Code for paper: Attribute-Aware Attention Model for Fine-grained Representation
 ![](./fig/a3m.png)
 
 ### Usage
-Requires: Keras 1.2.1 with Theano backend
+Requires: Keras 1.2.1 ("image_data_format": "channels_first")
 
-1. Download CUB-200-2011 dataset [here](http://www.vision.caltech.edu/visipedia/CUB-200-2011.html) and unzip it to `$CUB`; Change `data_dir` in `cub_demo.py` to `$CUB`.
+1. Download CUB-200-2011 dataset [here](http://www.vision.caltech.edu/visipedia/CUB-200-2011.html) and unzip it to `$CUB`; Copy file `tools/processed_attributes.txt` in `$CUB`.
 
-2. `sh cub_run.sh`
+2. Change `data_dir` in `run.sh` to `$CUB`, run the scprit `sh run.sh` to obtain the result.
 
-Result on CUB dataset
+- Result on CUB dataset
 ![](./fig/result.png)
 
 ### Citation
 
 Please use the following bibtex to cite our work:
 ```
-@inproceedings{a3m_2018_kai,
-  author = {Kai Han and Jianyuan Guo and Chao Zhang and Mingjian Zhu},
-  title = {Attribute-Aware Attention Model for Fine-grained Representation Learning},
-  year = {2018},
-  booktitle = {ACM Multimedia Conference}
-  }
+@inproceedings{han2018attribute,
+  title={Attribute-Aware Attention Model for Fine-grained Representation Learning},
+  author={Han, Kai and Guo, Jianyuan and Zhang, Chao and Zhu, Mingjian},
+  booktitle={2018 ACM Multimedia Conference on Multimedia Conference},
+  pages={2040--2048},
+  year={2018},
+  organization={ACM}
+}
 ```

cub_run.sh → run.sh

@@ -4,7 +4,8 @@ datetime=`date -d @$timestamp +"%Y-%m-%d-%H:%M:%S"`
 #net=VGG16
 #net=InceptionV3
 net=ResNet50
+data_dir=/home/hankai/data/CUB_200_2011
 gpu_id=0
-THEANO_FLAGS='device=gpu'$gpu_id',floatX=float32,lib.cnmem=0.6' python cub_demo.py $net | tee "./log/"$net"-"$datetime".log.txt"
+THEANO_FLAGS='device=gpu'$gpu_id',floatX=float32,lib.cnmem=0.6' python cub_demo.py $net $data_dir | tee "./log/"$net"-"$datetime".log.txt"
 
 

tools/attributes_process.py

@@ -0,0 +1,40 @@
+#coding=utf-8
+import re
+import numpy as np
+
+# get attribute cluster idx
+attribute_name_file = 'attributes.txt'
+f1 = open(attribute_name_file, 'rb')
+start_idxs = []
+last_attr = ''
+for line in f1.readlines():
+	strs = re.split(' |::', line)
+	if(strs[1]!=last_attr):
+		start_idxs.append(int(strs[0]))
+	last_attr = strs[1]
+start_idxs.append(int(strs[0])+1)
+print(start_idxs)
+a = np.array(start_idxs)
+nums = a[1:]-a[:-1]+1
+print(np.sum(nums))
+print(nums.tolist())
+
+# transform binary attribute to clustered attribute
+nb_attr = len(start_idxs)-1
+A_all = np.zeros((11788,nb_attr))
+image_attribute_file = 'attributes/image_attribute_labels.txt'
+f2 = open(image_attribute_file,'rb')
+for line in f2.readlines():
+	strs = re.split(' ', line)
+	img_id = int(strs[0])-1
+	attr_id = int(strs[1])
+	is_present = int(strs[2])
+	if(is_present>0):
+		for i in range(len(start_idxs)):
+			if(attr_id<start_idxs[i]):
+				break
+		A_all[img_id][i-1] = attr_id-start_idxs[i-1]+1 # 0 mean no attr
+print(A_all[1])
+
+new_attr_file = 'processed_attributes.txt'
+np.savetxt(new_attr_file,A_all,fmt='%d')