resnet improvements

examples/cifar10_resnet.py

@@ -20,9 +20,13 @@ from tensorpack.dataflow import imgaug
 
 """
 CIFAR10-resnet example.
-I can reproduce the results in:
 Deep Residual Learning for Image Recognition, arxiv:1512.03385
-for n=5 and 18 (6.5% val error)
+using the variants proposed in:
+Identity Mappings in Deep Residual Networks, arxiv::1603.05027
+
+I can reproduce the results
+for n=5 (about 7.7% val error) and 18 (about 6.4% val error)
+This model uses the whole training set instead of a 95:5 train-val split.
 """
 
 BATCH_SIZE = 128
@@ -44,9 +48,9 @@ class Model(ModelDesc):
         def conv(name, l, channel, stride):
             return Conv2D(name, l, channel, 3, stride=stride,
                           nl=tf.identity, use_bias=False,
-                          W_init=tf.random_normal_initializer(stddev=2.0/9/channel))
+                          W_init=tf.random_normal_initializer(stddev=np.sqrt(2.0/9/channel)))
 
-        def residual(name, l, increase_dim=False):
+        def residual(name, l, increase_dim=False, first=False):
             shape = l.get_shape().as_list()
             in_channel = shape[3]
 
@@ -58,25 +62,29 @@ class Model(ModelDesc):
                 stride1 = 1
 
             with tf.variable_scope(name) as scope:
-                c1 = conv('conv1', l, out_channel, stride1)
-                b1 = BatchNorm('bn1', c1, is_training)
+                if not first:
+                    b1 = BatchNorm('bn1', l, is_training)
                     b1 = tf.nn.relu(b1)
-                c2 = conv('conv2', b1, out_channel, 1)
-                b2 = BatchNorm('bn2', c2, is_training)
+                else:
+                    b1 = l
+                c1 = conv('conv1', b1, out_channel, stride1)
+                b2 = BatchNorm('bn2', c1, is_training)
+                b2 = tf.nn.relu(b2)
+                c2 = conv('conv2', b2, out_channel, 1)
 
                 if increase_dim:
                     l = AvgPooling('pool', l, 2)
                     l = tf.pad(l, [[0,0], [0,0], [0,0], [in_channel//2, in_channel//2]])
 
-                l = b2 + l
-                l = tf.nn.relu(l)
+                l = c2 + l
                 return l
 
 
-        l = conv('conv1', image, 16, 1)
-        l = BatchNorm('bn1', l, is_training)
+        l = conv('conv0', image, 16, 1)
+        l = BatchNorm('bn0', l, is_training)
         l = tf.nn.relu(l)
-        for k in range(self.n):
+        l = residual('res1.0', l, first=True)
+        for k in range(1, self.n):
             l = residual('res1.{}'.format(k), l)
         # 32,c=16
 
@@ -88,6 +96,8 @@ class Model(ModelDesc):
         l = residual('res3.0', l, increase_dim=True)
         for k in range(1, self.n):
             l = residual('res3.' + str(k), l)
+        l = BatchNorm('bnlast', l, is_training)
+        l = tf.nn.relu(l)
         # 8,c=64
         l = GlobalAvgPooling('gap', l)
         logits = FullyConnected('linear', l, out_dim=10, summary_activation=False,
@@ -123,6 +133,7 @@ def get_data(train_or_test):
             imgaug.RandomCrop((32, 32)),
             imgaug.Flip(horiz=True),
             imgaug.BrightnessAdd(20),
+            #imgaug.Contrast((0.6,1.4)),
             imgaug.MapImage(lambda x: x - pp_mean),
         ]
     else:
@@ -145,7 +156,6 @@ def get_config():
 
     sess_config = get_default_sess_config(0.9)
 
-    # warm up with small LR for 1 epoch
     lr = tf.Variable(0.01, trainable=False, name='learning_rate')
     tf.scalar_summary('learning_rate', lr)
 
@@ -157,7 +167,7 @@ def get_config():
             PeriodicSaver(),
             ValidationError(dataset_test, prefix='test'),
             ScheduledHyperParamSetter('learning_rate',
-                                      [(1, 0.1), (82, 0.01), (123, 0.001), (300, 0.0001)])
+                                      [(1, 0.1), (82, 0.01), (123, 0.001), (300, 0.0002)])
         ]),
         session_config=sess_config,
         model=Model(n=18),

tensorpack/train/base.py

@@ -49,7 +49,7 @@ class Trainer(object):
         if not hasattr(logger, 'LOG_DIR'):
             raise RuntimeError("Please use logger.set_logger_dir at the beginning of your script.")
         self.summary_writer = tf.train.SummaryWriter(
-            logger.LOG_DIR, graph_def=self.sess.graph)
+            logger.LOG_DIR, graph=self.sess.graph)
         self.summary_op = tf.merge_all_summaries()
         # create an empty StatHolder
         self.stat_holder = StatHolder(logger.LOG_DIR, [])