CODEWITHSUNDEEP
pythonmachine-learningtensorflowdeep-learningconv-neural-network
Hadi GhahremanNezhad
Hadi GhahremanNezhad

Reputation: 2445

How to reduce size of bottleneck features of intermediate layer of VGG16?

I am trying to connect the layer c0nv4_3 of vgg16 network instead of conv5_3 to the RPN network of Faster R-CNN. Here is the python code of vgg16 network. I have changed these lines:

def _image_to_head(self, is_training, reuse=False):
    with tf.variable_scope(self._scope, self._scope, reuse=reuse):
      net = slim.repeat(self._image, 2, slim.conv2d, 64, [3, 3],
                          trainable=False, scope='conv1')
      net = slim.max_pool2d(net, [2, 2], padding='SAME', scope='pool1')
      net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3],
                        trainable=False, scope='conv2')
      net = slim.max_pool2d(net, [2, 2], padding='SAME', scope='pool2')
      net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3],
                        trainable=is_training, scope='conv3')
      net = slim.max_pool2d(net, [2, 2], padding='SAME', scope='pool3')
      net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3],
                        trainable=is_training, scope='conv4')
      net = slim.max_pool2d(net, [2, 2], padding='SAME', scope='pool4')
      net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3],
                        trainable=is_training, scope='conv5')

    self._act_summaries.append(net)
    self._layers['head'] = net

return net

to:

def _image_to_head(self, is_training, reuse=False):
    with tf.variable_scope(self._scope, self._scope, reuse=reuse):
      net = slim.repeat(self._image, 2, slim.conv2d, 64, [3, 3],
                          trainable=False, scope='conv1')
      net = slim.max_pool2d(net, [2, 2], padding='SAME', scope='pool1')
      net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3],
                        trainable=False, scope='conv2')
      net = slim.max_pool2d(net, [2, 2], padding='SAME', scope='pool2')
      net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3],
                        trainable=is_training, scope='conv3')
      net = slim.max_pool2d(net, [2, 2], padding='SAME', scope='pool3')
      net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3],
                        trainable=is_training, scope='conv4')

    self._act_summaries.append(net)
    self._layers['head'] = net

return net

As seen above, I removed the conv5 and pool4 layers; because my objects are small and I hoped to get better results, but the results got worse. I think I need to add a deconv layer to the end of conv4? Or there is another way?

thanks

Upvotes: 0

Views: 292

Answers (1)

Pramod Patil
Pramod Patil

Reputation: 827

There are methods too, for reducing the length of bottleneck features.

Why not to add deconv:

  • You will initialize deconv layer with random values
  • You are not finetuning the network, you are just making forward pass through the network.
  • So the output of deconv will randomize your features of conv4

Pooling Layers:

  • Average pooling(based on the window size, it will return average of that window). So if lets say window(2,2) with values[3,2,4,3] will result into only one value: 6

  • MaxPool(based on the window size, it will result maximum value of that window). So if lets say window(2,2) with values[3,2,4,3] will result into only one value: 3

Check out pooling layers here

Upvotes: 1

Related Questions