我正在TensorFlow中实现SegNet，我用它来将航拍图像分成两类：“建筑”和“不建筑”。我有一个小版本的网络，准确度高达82％mIoU。
但是，我想通过添加多个卷积层来扩展网络，就像原始的SegNet一样，但是我无法让它工作。
这就是我实现有效的小模型的方式：
def inference_basic(images, phase_train, batch_size, keep_prob):
conv1 = conv_layer_with_bn(norm1, [7, 7, images.get_shape().as_list()[3], 64], phase_train, name="conv1")
pool1, pool1_indices = tf.nn.max_pool_with_argmax(conv1, ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1], padding='SAME', name='pool1')
conv2 = conv_layer_with_bn(pool1, [7, 7, 64, 64], phase_train, name="conv2")
pool2, pool2_indices = tf.nn.max_pool_with_argmax(conv2, ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1], padding='SAME', name='pool2')
conv3 = conv_layer_with_bn(pool2, [7, 7, 64, 64], phase_train, name="conv3")
pool3, pool3_indices = tf.nn.max_pool_with_argmax(conv3, ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1], padding='SAME', name='pool3')
conv4 = conv_layer_with_bn(pool3, [7, 7, 64, 64], phase_train, name="conv4")
pool4, pool4_indices = tf.nn.max_pool_with_argmax(conv4, ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1], padding='SAME', name='pool4')
""" End of encoder """
""" start decoder """
upsample4 = deconv_layer(pool4, [2, 2, 64, 64], [batch_size, FLAGS.image_h//8, FLAGS.image_w//8, 64], 2, "up4")
conv_decode4 = conv_layer_with_bn(upsample4, [7, 7, 64, 64], phase_train, False, name="conv_decode4")
upsample3= deconv_layer(conv_decode4, [2, 2, 64, 64], [batch_size, FLAGS.image_h//4, FLAGS.image_w//4, 64], 2, "up3")
conv_decode3 = conv_layer_with_bn(upsample3, [7, 7, 64, 64], phase_train, False, name="conv_decode3")
upsample2= deconv_layer(conv_decode3, [2, 2, 64, 64], [batch_size, FLAGS.image_h//2, FLAGS.image_w//2, 64], 2, "up2")
conv_decode2 = conv_layer_with_bn(upsample2, [7, 7, 64, 64], phase_train, False, name="conv_decode2")
upsample1= deconv_layer(conv_decode2, [2, 2, 64, 64], [batch_size, FLAGS.image_h, FLAGS.image_w, 64], 2, "up1")
conv_decode1 = conv_layer_with_bn(upsample1, [7, 7, 64, 64], phase_train, False, name="conv_decode1")
""" end of decoder """
""" Start Classify """
with tf.variable_scope('conv_classifier') as scope:
kernel = _variable_with_weight_decay('weights',
shape=[1, 1, 64, FLAGS.num_class],
initializer=msra_initializer(1, 64),
wd=0.0005)
conv = tf.nn.conv2d(conv_decode1, kernel, [1, 1, 1, 1], padding='SAME')
biases = _variable_on_cpu('biases', [FLAGS.num_class], tf.constant_initializer(0.0))
conv_classifier = tf.nn.bias_add(conv, biases, name=scope.name)
return conv_classifier
这是扩展模型，结果非常糟糕：
def inference(images, phase_train, batch_size):
conv1_1 = conv_layer_with_bn(images, [7, 7, images.get_shape().as_list()[3], 64], phase_train, name="conv1_1")
conv1_2 = conv_layer_with_bn(conv1_1, [7, 7, 64, 64], phase_train, name="conv1_2")
pool1, pool1_indices = tf.nn.max_pool_with_argmax(conv1_2, ksize=[1, 2, 2, 1],strides=[1, 2, 2, 1], padding='SAME', name='pool1')
conv2_1 = conv_layer_with_bn(pool1, [7, 7, 64, 64], phase_train, name="conv2_1")
conv2_2 = conv_layer_with_bn(conv2_1, [7, 7, 64, 64], phase_train, name="conv2_2")
pool2, pool2_indices = tf.nn.max_pool_with_argmax(conv2_2, ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1], padding='SAME', name='pool2')
conv3_1 = conv_layer_with_bn(pool2, [7, 7, 64, 64], phase_train, name="conv3_1")
conv3_2 = conv_layer_with_bn(conv3_1, [7, 7, 64, 64], phase_train, name="conv3_2")
conv3_3 = conv_layer_with_bn(conv3_2, [7, 7, 64, 64], phase_train, name="conv3_3")
pool3, pool3_indices = tf.nn.max_pool_with_argmax(conv3_3, ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1], padding='SAME', name='pool3')
conv4_1 = conv_layer_with_bn(pool3, [7, 7, 64, 64], phase_train, name="conv4_1")
conv4_2 = conv_layer_with_bn(conv4_1, [7, 7, 64, 64], phase_train, name="conv4_2")
conv4_3 = conv_layer_with_bn(conv4_2, [7, 7, 64, 64], phase_train, name="conv4_3")
pool4, pool4_indices = tf.nn.max_pool_with_argmax(conv4_3, ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1], padding='SAME', name='pool4')
conv5_1 = conv_layer_with_bn(pool4, [7, 7, 64, 64], phase_train, name="conv5_1")
conv5_2 = conv_layer_with_bn(conv5_1, [7, 7, 64, 64], phase_train, name="conv5_2")
conv5_3 = conv_layer_with_bn(conv5_2, [7, 7, 64, 64], phase_train, name="conv5_3")
pool5, pool5_indices = tf.nn.max_pool_with_argmax(conv5_3, ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1], padding='SAME', name='pool5')
""" End of encoder """
""" Start decoder """
upsample5 = deconv_layer(pool5, [2, 2, 64, 64], [batch_size, FLAGS.image_h//16, FLAGS.image_w//16, 64], 2, "up5")
conv_decode5_1 = conv_layer_with_bn(upsample5, [7, 7, 64, 64], phase_train, True, name="conv_decode5_1")
conv_decode5_2 = conv_layer_with_bn(conv_decode5_1, [7, 7, 64, 64], phase_train, True, name="conv_decode5_2")
conv_decode5_3 = conv_layer_with_bn(conv_decode5_2, [7, 7, 64, 64], phase_train, True, name="conv_decode5_3")
upsample4 = deconv_layer(conv_decode5_3, [2, 2, 64, 64], [batch_size, FLAGS.image_h//8, FLAGS.image_w//8, 64], 2, "up4")
conv_decode4_1 = conv_layer_with_bn(upsample4, [7, 7, 64, 64], phase_train, True, name="conv_decode4_1")
conv_decode4_2 = conv_layer_with_bn(conv_decode4_1, [7, 7, 64, 64], phase_train, True, name="conv_decode4_2")
conv_decode4_3 = conv_layer_with_bn(conv_decode4_2, [7, 7, 64, 64], phase_train, True, name="conv_decode4_3")
upsample3 = deconv_layer(conv_decode4_3, [2, 2, 64, 64], [batch_size, FLAGS.image_h//4, FLAGS.image_w//4, 64], 2, "up3")
conv_decode3_1 = conv_layer_with_bn(upsample3, [7, 7, 64, 64], phase_train, True, name="conv_decode3_1")
conv_decode3_2 = conv_layer_with_bn(conv_decode3_1, [7, 7, 64, 64], phase_train, True, name="conv_decode3_2")
conv_decode3_3 = conv_layer_with_bn(conv_decode3_2, [7, 7, 64, 64], phase_train, True, name="conv_decode3_3")
upsample2= deconv_layer(conv_decode3_3, [2, 2, 64, 64], [batch_size, FLAGS.image_h//2, FLAGS.image_w//2, 64], 2, "up2")
conv_decode2_1 = conv_layer_with_bn(upsample2, [7, 7, 64, 64], phase_train, True, name="conv_decode2_1")
conv_decode2_2 = conv_layer_with_bn(conv_decode2_1, [7, 7, 64, 64], phase_train, True, name="conv_decode2_2")
upsample1 = deconv_layer(conv_decode2_2, [2, 2, 64, 64], [batch_size, FLAGS.image_h, FLAGS.image_w, 64], 2, "up1")
conv_decode1_1 = conv_layer_with_bn(upsample1, [7, 7, 64, 64], phase_train, True, name="conv_decode1_1")
conv_decode1_2 = conv_layer_with_bn(conv_decode1_1, [7, 7, 64, 64], phase_train, True, name="conv_decode1_2")
""" End of decoder """
""" Start Classify """
# output predicted class number
with tf.variable_scope('conv_classifier') as scope: #all variables prefixed with "conv_classifier/"
kernel = _variable_with_weight_decay('weights',
shape=[1, 1, 64, FLAGS.num_class],
initializer=msra_initializer(1, 64),
wd=0.0005)
conv = tf.nn.conv2d(conv_decode1_2, kernel, [1, 1, 1, 1], padding='SAME')
biases = _variable_on_cpu('biases', [FLAGS.num_class], tf.constant_initializer(0.0))
conv_classifier = tf.nn.bias_add(conv, biases, name=scope.name)
#logit = conv_classifier = prediction
return conv_classifier
卷积层：
def conv_layer_with_bn(inputT, shape, train_phase, activation=True, name=None):
in_channel = shape[2]
out_channel = shape[3]
k_size = shape[0]
with tf.variable_scope(name) as scope:
kernel = _variable_with_weight_decay('weights',
shape=shape,
initializer=msra_initializer(k_size, in_channel),
wd=None)
conv = tf.nn.conv2d(inputT, kernel, [1, 1, 1, 1], padding='SAME')
biases = _variable_on_cpu('biases', [out_channel], tf.constant_initializer(0.0))
bias = tf.nn.bias_add(conv, biases)
if activation is True:
conv_out = tf.nn.relu(batch_norm_layer(bias, train_phase, scope.name))
else:
conv_out = batch_norm_layer(bias, train_phase, scope.name)
return conv_out
def batch_norm_layer(inputT, is_training, scope):
"""Used in conv_layer_with_bn()"""
return tf.cond(is_training,
lambda: tf.contrib.layers.batch_norm(inputT, is_training=True,
center=False, updates_collections=None, scope=scope+"_bn"),
lambda: tf.contrib.layers.batch_norm(inputT, is_training=False,
updates_collections=None, center=False, scope=scope+"_bn", reuse = True))
扩展模型大约10％mIoU，因为图像中的所有像素都被分类为“非构建”类。任何人都可以帮助我理解为什么会这样吗？我查看了SegNet的caffe implementation，我看不出两种实现之间的区别。
答案 0 :(得分：0)
经过一些测试，我想我可能知道更深层次模型表现不佳的原因。这似乎是权重初始化的一个问题，我认为，这在更深层次的模型中更为关键。我更新了我的模型，使用Delving Deep into Rectifiers,论文中提出的权重初始化器，以及随机梯度下降，学习率为0.1。这似乎解决了这个问题！
我的想法是否正确？在使用更深层次的模型时，权重初始化是否会变得更加重要？