update dch

DeepHash/model/dch/__init__.py

@@ -0,0 +1,16 @@
+from .util import Dataset
+from .dch import DCH
+
+def train(train_img, database_img, query_img, config):
+    model = DCH(config)
+    img_database = Dataset(database_img, config.output_dim)
+    img_query = Dataset(query_img, config.output_dim)
+    img_train = Dataset(train_img, config.output_dim)
+    model.train(img_train)
+    return model.save_dir
+
+def validation(database_img, query_img, config):
+    model = DCH(config)
+    img_database = Dataset(database_img, config.output_dim)
+    img_query = Dataset(query_img, config.output_dim)
+    return model.validation(img_query, img_database, config.R)

DeepHash/model/prunehash/prunehash.py → DeepHash/model/dch/dch.py

@@ -16,44 +16,28 @@
 import model.plot as plot
 from architecture.single_model import img_alexnet_layers
 from evaluation import MAPs
-from .util import Dataset
 
 
-class PruneHash(object):
-    def __init__(self, config, stage):
+class DCH(object):
+    def __init__(self, config):
         ### Initialize setting
         print ("initializing")
         np.set_printoptions(precision=4)
-        self.stage = stage
-        self.device = config['device']
-        self.output_dim = config['output_dim']
-        self.n_class = config['label_dim']
-        self.cq_lambda = config['cq_lambda']
-        self.alpha = config['alpha']
-        self.bias = config['bias']
-        self.gamma = config['gamma']
-
-        self.batch_size = config['batch_size'] if self.stage == "train" else config['val_batch_size']
-        self.max_iter = config['max_iter']
-        self.img_model = config['img_model']
-        self.loss_type = config['loss_type']
-        self.learning_rate = config['learning_rate']
-        self.learning_rate_decay_factor = config['learning_rate_decay_factor']
-        self.decay_step = config['decay_step']
-
-        self.finetune_all = config['finetune_all']
 
+        with tf.name_scope('stage'):
+            # 0 for training, 1 for validation
+            self.stage = tf.placeholder_with_default(tf.constant(0), [])
+        for k, v in vars(config).items():
+            setattr(self, k, v)
         self.file_name = 'loss_{}_lr_{}_cqlambda_{}_alpha_{}_bias_{}_gamma_{}_dataset_{}'.format(
                 self.loss_type,
-                self.learning_rate,
-                self.cq_lambda,
+                self.lr,
+                self.q_lambda,
                 self.alpha,
                 self.bias,
                 self.gamma,
-                config['dataset'])
-        self.save_dir = config['save_dir']
-        self.save_file = os.path.join(config['save_dir'], self.file_name + '.npy')
-        self.log_dir = config['log_dir']
+                self.dataset)
+        self.save_file = os.path.join(self.save_dir, self.file_name + '.npy')
 
         ### Setup session
         print ("launching session")
@@ -63,27 +47,25 @@ def __init__(self, config, stage):
         self.sess = tf.Session(config=configProto)
 
         ### Create variables and placeholders
+        self.img = tf.placeholder(tf.float32, [None, 256, 256, 3])
+        self.img_label = tf.placeholder(tf.float32, [None, self.label_dim])
+        self.img_last_layer, self.deep_param_img, self.train_layers, self.train_last_layer = self.load_model()
 
-        with tf.device(self.device):
-            self.img = tf.placeholder(tf.float32, [self.batch_size, 256, 256, 3])
-            self.img_label = tf.placeholder(tf.float32, [self.batch_size, self.n_class])
-
-            if self.stage == 'train':
-                self.model_weights = config['model_weights']
-            else:
-                self.model_weights = self.save_file
-            self.img_last_layer, self.deep_param_img, self.train_layers, self.train_last_layer = self.load_model()
-
-            self.global_step = tf.Variable(0, trainable=False)
-            self.train_op = self.apply_loss_function(self.global_step)
-            self.sess.run(tf.global_variables_initializer())
+        self.global_step = tf.Variable(0, trainable=False)
+        self.train_op = self.apply_loss_function(self.global_step)
+        self.sess.run(tf.global_variables_initializer())
         return
 
     def load_model(self):
         if self.img_model == 'alexnet':
             img_output = img_alexnet_layers(
-                    self.img, self.batch_size, self.output_dim,
-                    self.stage, self.model_weights)
+                    self.img,
+                    self.batch_size,
+                    self.output_dim,
+                    self.stage,
+                    self.model_weights,
+                    self.with_tanh,
+                    self.val_batch_size)
         else:
             raise Exception('cannot use such CNN model as ' + self.img_model)
         return img_output
@@ -139,7 +121,7 @@ def reduce_shaper(t):
             r = tf.reshape(r, [-1, 1])
             ip = r - 2*tf.matmul(u, tf.transpose(u)) + tf.transpose(r)
 
-            ip = tf.constant(self.gamma) / (ip + tf.constant(self.gamma)*tf.constant(self.gamma))
+            ip = self.gamma / (ip + self.gamma ** 2)
         else:
             ip = tf.clip_by_value(tf.matmul(u, tf.transpose(u)), -1.5e1, 1.5e1)
         ones = tf.ones([tf.shape(u)[0], tf.shape(u)[0]])
@@ -158,13 +140,12 @@ def apply_loss_function(self, global_step):
             self.cos_loss = self.cross_entropy(self.img_last_layer, self.img_label, self.alpha, True, True, self.bias)
 
         self.q_loss_img = tf.reduce_mean(tf.square(tf.subtract(tf.abs(self.img_last_layer), tf.constant(1.0))))
-        self.q_lambda = tf.Variable(self.cq_lambda, name='cq_lambda')
-        self.q_loss = tf.multiply(self.q_lambda, self.q_loss_img)
+        self.q_loss = self.q_lambda * self.q_loss_img
         self.loss = self.cos_loss + self.q_loss
 
         ### Last layer has a 10 times learning rate
-        self.lr = tf.train.exponential_decay(self.learning_rate, global_step, self.decay_step, self.learning_rate_decay_factor, staircase=True)
-        opt = tf.train.MomentumOptimizer(learning_rate=self.lr, momentum=0.9)
+        lr = tf.train.exponential_decay(self.lr, global_step, self.decay_step, self.lr, staircase=True)
+        opt = tf.train.MomentumOptimizer(learning_rate=lr, momentum=0.9)
         grads_and_vars = opt.compute_gradients(self.loss, self.train_layers+self.train_last_layer)
         fcgrad, _ = grads_and_vars[-2]
         fbgrad, _ = grads_and_vars[-1]
@@ -174,11 +155,11 @@ def apply_loss_function(self, global_step):
         tf.summary.scalar('loss', self.loss)
         tf.summary.scalar('cos_loss', self.cos_loss)
         tf.summary.scalar('q_loss', self.q_loss)
-        tf.summary.scalar('lr', self.lr)
+        tf.summary.scalar('lr', lr)
         self.merged = tf.summary.merge_all()
 
 
-        if self.stage == "train" and self.finetune_all:
+        if self.finetune_all:
             return opt.apply_gradients([(grads_and_vars[0][0], self.train_layers[0]),
                                         (grads_and_vars[1][0]*2, self.train_layers[1]),
                                         (grads_and_vars[2][0], self.train_layers[2]),
@@ -208,13 +189,10 @@ def train(self, img_dataset):
             shutil.rmtree(tflog_path)
         train_writer = tf.summary.FileWriter(tflog_path, self.sess.graph)
 
-        for train_iter in range(self.max_iter):
+        for train_iter in range(self.iter_num):
             images, labels = img_dataset.next_batch(self.batch_size)
             start_time = time.time()
 
-            assign_lambda = self.q_lambda.assign(self.cq_lambda)
-            self.sess.run([assign_lambda])
-
             _, loss, cos_loss, output, summary = self.sess.run([self.train_op, self.loss, self.cos_loss, self.img_last_layer, self.merged],
                                     feed_dict={self.img: images,
                                                self.img_label: labels})
@@ -224,7 +202,7 @@ def train(self, img_dataset):
             img_dataset.feed_batch_output(self.batch_size, output)
             duration = time.time() - start_time
 
-            if train_iter % 1 == 0:
+            if train_iter % 100 == 0:
                 print("%s #train# step %4d, loss = %.4f, cross_entropy loss = %.4f, %.1f sec/batch"
                         %(datetime.now(), train_iter+1, loss, cos_loss, duration))
 
@@ -236,24 +214,29 @@ def train(self, img_dataset):
 
     def validation(self, img_query, img_database, R=100):
         print("%s #validation# start validation" % (datetime.now()))
-        query_batch = int(ceil(img_query.n_samples / self.batch_size))
+        query_batch = int(ceil(img_query.n_samples / float(self.val_batch_size)))
+        img_query.finish_epoch()
         print("%s #validation# totally %d query in %d batches" % (datetime.now(), img_query.n_samples, query_batch))
         for i in range(query_batch):
-            images, labels = img_query.next_batch(self.batch_size)
+            images, labels = img_query.next_batch(self.val_batch_size)
             output, loss = self.sess.run([self.img_last_layer, self.cos_loss],
-                                   feed_dict={self.img: images, self.img_label: labels})
-            img_query.feed_batch_output(self.batch_size, output)
+                                   feed_dict={self.img: images,
+                                              self.img_label: labels,
+                                              self.stage: 1})
+            img_query.feed_batch_output(self.val_batch_size, output)
             print('Cosine Loss: %s'%loss)
 
-        database_batch = int(ceil(img_database.n_samples / self.batch_size))
+        database_batch = int(ceil(img_database.n_samples / float(self.val_batch_size)))
+        img_database.finish_epoch()
         print("%s #validation# totally %d database in %d batches" % (datetime.now(), img_database.n_samples, database_batch))
         for i in range(database_batch):
-            images, labels = img_database.next_batch(self.batch_size)
+            images, labels = img_database.next_batch(self.val_batch_size)
 
             output, loss = self.sess.run([self.img_last_layer, self.cos_loss],
-                                   feed_dict={self.img: images, self.img_label: labels})
-            img_database.feed_batch_output(self.batch_size, output)
-            #print output[:10, :10]
+                                         feed_dict={self.img: images,
+                                                    self.img_label: labels,
+                                                    self.stage: 1})
+            img_database.feed_batch_output(self.val_batch_size, output)
             if i % 100 == 0:
                 print('Cosine Loss[%d/%d]: %s'%(i, database_batch, loss))
 
@@ -283,15 +266,3 @@ def validation(self, img_query, img_database, R=100):
             'i2i_map_radius_2': mmap,
         }
 
-def train(train_img, config):
-    model = PruneHash(config, 'train')
-    img_dataset = Dataset(train_img, config['output_dim'])
-    model.train(img_dataset)
-    return model.save_file
-
-def validation(database_img, query_img, config):
-    model = PruneHash(config, 'val')
-    img_database = Dataset(database_img, config['output_dim'])
-    img_query = Dataset(query_img, config['output_dim'])
-    return model.validation(img_query, img_database, config['R'])
-

examples/dch/train_val_script.py

@@ -0,0 +1,70 @@
+import os
+import argparse
+import warnings
+import numpy as np
+import scipy.io as sio
+import model.dch as model
+import data_provider.image as dataset
+
+from pprint import pprint
+
+warnings.filterwarnings("ignore", category = DeprecationWarning)
+warnings.filterwarnings("ignore", category = FutureWarning)
+
+os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
+
+parser = argparse.ArgumentParser(description='Triplet Hashing')
+parser.add_argument('--lr', '--learning-rate', default=0.005, type=float)
+parser.add_argument('--output-dim', default=64, type=int)   # 256, 128
+parser.add_argument('--alpha', default=0.5, type=float)
+parser.add_argument('--bias', default=0.0, type=float)
+parser.add_argument('--gamma', default=20, type=float)
+parser.add_argument('--iter-num', default=2000, type=int)
+parser.add_argument('--q-lambda', default=0, type=float)
+parser.add_argument('--dataset', default='cifar10', type=str)
+parser.add_argument('--gpus', default='0', type=str)
+parser.add_argument('--log-dir', default='tflog', type=str)
+parser.add_argument('-b', '--batch-size', default=128, type=int)
+parser.add_argument('-vb', '--val-batch-size', default=16, type=int)
+parser.add_argument('--decay-step', default=10000, type=int)
+parser.add_argument('--decay-factor', default=0.1, type=int)
+parser.add_argument('--loss-type', default='pruned_cross_entropy', type=str)
+
+tanh_parser = parser.add_mutually_exclusive_group(required=False)
+tanh_parser.add_argument('--with-tanh', dest='with_tanh', action='store_true')
+tanh_parser.add_argument('--without-tanh', dest='with_tanh', action='store_false')
+parser.set_defaults(with_tanh=True)
+
+parser.add_argument('--img-model', default='alexnet', type=str)
+parser.add_argument('--model-weights', type=str,
+                    default='../../DeepHash/architecture/single_model/pretrained_model/reference_pretrain.npy')
+parser.add_argument('--finetune-all', default=True, type=bool)
+parser.add_argument('--save-dir', default="./models/", type=str)
+parser.add_argument('-e', '--evaluate', dest='evaluate', action='store_true')
+
+args = parser.parse_args()
+
+os.environ['CUDA_VISIBLE_DEVICES'] = args.gpus
+
+label_dims = {'cifar10': 10, 'cub': 200, 'nuswide_81': 81, 'coco': 80}
+Rs = {'cifar10': 54000, 'nuswide_81': 5000, 'coco': 5000}
+args.R = Rs[args.dataset]
+args.label_dim = label_dims[args.dataset]
+args.img_tr = "/home/caoyue/data/{}/train.txt".format(args.dataset)
+args.img_te = "/home/caoyue/data/{}/test.txt".format(args.dataset)
+args.img_db = "/home/caoyue/data/{}/database.txt".format(args.dataset)
+
+pprint(vars(args))
+
+query_img, database_img = dataset.import_validation(args.img_te, args.img_db)
+
+if not args.evaluate:
+    train_img = dataset.import_train(args.img_tr)
+    model_weights = model.train(train_img, database_img, query_img, args)
+    args.model_weights = model_weights
+
+maps = model.validation(database_img, query_img, args)
+for key in maps:
+    print(("{}\t{}".format(key, maps[key])))
+
+pprint(vars(args))

examples/dtq/train_val_script.py

@@ -37,7 +37,7 @@
 
 parser.add_argument('--img-model', default='alexnet', type=str)
 parser.add_argument('--model-weights', type=str,
-                    default='../../core/architecture/single_model/pretrained_model/reference_pretrain.npy')
+                    default='../../DeepHash/architecture/single_model/pretrained_model/reference_pretrain.npy')
 parser.add_argument('--finetune-all', default=True, type=bool)
 parser.add_argument('--max-iter-update-b', default=3, type=int)
 parser.add_argument('--max-iter-update-Cb', default=1, type=int)