1

I have 2 sets of image patches data i.e. training and testing sets. Both of these have been written to LMDB files. I am running convolutional neurall network on this data using Caffe.

The problem is that the data stored on hard disk is occupying considerable amount of space and is hampering my efforts to introduce more training data with deliberate noise addition to make my model more robust.

Is there a way where I can send image patches from my program directly to the CNN (in Caffe) without storing them in LMDB? I am currently using python to generate patches from the images for the training data set.

shripati007
  • 87
  • 1
  • 2
  • 13
  • Have you tried using an ImageData layer? – malreddysid Aug 22 '16 at 11:02
  • When converting images to LMDB files, you can encode them by `convert_imageset --encoded=true --encoded_type=png` to save space details can be found [here](http://stackoverflow.com/a/31431716/6281477). Also, you can consider to use ImageData layer and try multi-core version of OpenCV to speed up reading images according to [this post](http://stackoverflow.com/a/35715139/6281477) – Dale Aug 23 '16 at 01:31

2 Answers2

2

You can write your own python data layer. See discussions here and implementation for of input data layer for video stream here.

Basically you will need add to you network description layer like:

layer {
  type: 'Python'
  name: 'data'
  top: 'data'
  top: 'label'
  python_param {
    # the module name -- usually the filename -- that needs to be in $PYTHONPATH
    module: 'filename'
    # the layer name -- the class name in the module
    layer: 'CustomInputDataLayer'
  }
}

and implement the layer interface in Python:

class CustomInputDataLayer(caffe.Layer):
    def setup(self):
         ...

    def reshape(self, bottom, top)
        top[0].reshape(BATCH_SIZE, your_data.shape)
        top[1].reshape(BATCH_SIZE, your_label.shape)

    def forward(self, bottom, top):
        # assign output
        top[0].data[...] = your_data
        top[1].data[...] = your_label

    def backward(self, top, propagate_down, bottom):
        pass
taarraas
  • 1,483
  • 9
  • 18
2

Other than defining custom python layers, you can use the following options:

  • use ImageData layer: it has a source parameter (source: name of a text file, with each line giving an image filename and label)

  • use MemoryData layer: using which you can load input images directly from memory to your network using ‘setinputarrays‘ method in python. Be cautious about using this layer as it only accepts labels which are single values and you cannot use images as labels (e.g. In semantic segmentation)

  • use a deploy version of your network like this:

    input: "data"
input_shape {
dim: n # batch size
dim: c # number of channels
dim: r # image size1
dim: w # image size2
}

input: "label"
input_shape {
dim: n # batch size
dim: c # number of channels
dim: r # label image size1
dim: w # label image size2
}
 ... #your other layers to follow

  • use an HDF5 input layer (more or less ine lmdb, but lmdb is more computationally efficient)

You can find the details of these layers here: http://caffe.berkeleyvision.org/tutorial/layers.html

There are examples available online as well.

Amir
  • 2,259
  • 1
  • 19
  • 29
  • What if we have to extract patches online for semantic segmentation? I have a python layer that is doing a specific job, and then I need to extract patches from a specific region and then send it to another network. Is it possible to do it? What are the options? Thanks – S.EB May 19 '18 at 01:41