MXNet Scala Data Loading API

This topic introduces the data input method for MXNet. MXNet uses an iterator to provide data to the neural network. Iterators do some preprocessing and generate batches for the neural network.

MXNet provides basic iterators for MNIST and Recordio images. To hide the cost of I/O, MXNet uses a prefetch strategy that enables parallelism for the learning process and data fetching. Data is automatically fetched by an independent thread.


Data Iterator Parameters

To create a data iterator, you typically need to provide five parameters:

  • Dataset Param provides basic information about the dataset, e.g., file path, input shape.
  • Batch Param provides information required to form a batch, e.g., batch size.
  • Augmentation Param tells MXNet which augmentation operations (e.g., crop or mirror) to perform on an input image.
  • Backend Param controls the behavior of the back-end threads to hide the cost of data loading.
  • Auxiliary Param provides options for checking and debugging.

You must provide the Dataset Param and Batch Param, otherwise MXNet can’t create the data batch. Provide other parameters as required by your algorithm and performance needs. We provide a detailed explanation and examples of the options later.

Create a Data Iterator

The IO API provides a simple way to create a data iterator in Scala. The following example code shows how to create a CIFAR data iterator.

    scala> val dataiter = IO.ImageRecordIter(Map(
    scala>        // Utility Parameter
    scala>        // Optional
    scala>        // Name of the data, should match the name of the data input of the network
    scala>        // data_name='data',
    scala>        // Utility Parameter
    scala>        // Optional
    scala>        // Name of the label, should match the name of the label parameter of the network.
    scala>        // Usually, if the loss layer is named 'foo', then the label input has the name
    scala>        // 'foo_label', unless overwritten
    scala>        // label_name='softmax_label',
    scala>        // Dataset Parameter
    scala>        // Impulsary
    scala>        // indicating the data file, please check the data is already there
    scala>        "path_imgrec" -> "data/cifar/train.rec",
    scala>        // Dataset Parameter
    scala>        // Impulsary
    scala>        // indicating the image size after preprocessing
    scala>        "data_shape" -> "(3,28,28)",
    scala>        // Batch Parameter
    scala>        // Impulsary
    scala>        // tells how many images in a batch
    scala>        "batch_size" -> "100",
    scala>        // Augmentation Parameter
    scala>        // Optional
    scala>        // when offers mean_img, each image will subtract the mean value at each pixel
    scala>        "mean_img" -> "data/cifar/cifar10_mean.bin",
    scala>        // Augmentation Parameter
    scala>        // Optional
    scala>        // randomly crop a patch of the data_shape from the original image
    scala>        "rand_crop" -> "True",
    scala>        // Augmentation Parameter
    scala>        // Optional
    scala>        // randomly mirror the image horizontally
    scala>        "rand_mirror" -> "True",
    scala>        // Augmentation Parameter
    scala>        // Optional
    scala>        // randomly shuffle the data
    scala>        "shuffle" -> "False",
    scala>        // Backend Parameter
    scala>        // Optional
    scala>        // Preprocessing thread number
    scala>        "preprocess_threads" -> "4",
    scala>        // Backend Parameter
    scala>        // Optional
    scala>        // Prefetch buffer size
    scala>        "prefetch_buffer" = "1"))

First, explicitly specify the kind of data (MNIST, ImageRecord, etc.) to fetch. Then, provide the options for the dataset, batching, image augmentation, multi-tread processing, and prefetching operations. The code automatically validates the parameters. If a required parameter is missing, MXNet returns an error.

How to Get Data

We provide a scripts to download MNIST data and CIFAR10 ImageRecord data. If you want to create your own dataset, we recommend using the Image Recordio data format.

Create a Dataset Using Recordio

Recordio implements a file format for a sequence of records. We recommend storing images as records and packing them together. The benefits include:

  • Storing images in a compact format–e.g., JPEG, for records–greatly reduces the size of the dataset on the disk.
  • Packing data together allows continuous reading on the disk.
  • Recordio has a simple way to partition, simplifying distributed setting. We provide an example later.

We provide the im2rec tool so you can create an Image Recordio dataset by yourself. The following walkthrough shows you how.


Download the data. You don’t need to resize the images manually. You can use im2rec to resize them automatically. For details, see the “Extension” topic.

Step 1. Make an Image List File

After you download the data, you need to make an image list file. The format is:

    integer_image_index \t label_index \t path_to_image

Typically, the program takes the list of names of all of the images, shuffles them, then separates them into two lists: a training files name list and a testing file name list. Write the list in the right format.

This is example file:

    95099  464     n04467665_17283.JPEG
    10025081        412     ILSVRC2010_val_00025082.JPEG
    74181   789     n01915811_2739.JPEG
    10035553        859     ILSVRC2010_val_00035554.JPEG
    10048727        929     ILSVRC2010_val_00048728.JPEG
    94028   924     n01980166_4956.JPEG
    1080682 650     n11807979_571.JPEG
    972457  633     n07723039_1627.JPEG
    7534    11      n01630670_4486.JPEG
    1191261 249     n12407079_5106.JPEG

Step 2. Create the Binary File

To generate a binary image, use im2rec in the tool folder. im2rec takes the path of the image list file you generated, the root path of the images, and the output file path as input. This process usually takes several hours, so be patient.

A sample command:

    ./bin/im2rec image.lst image_root_dir output.bin resize=256

For more details, run ./bin/im2rec.

Extension: Multiple Labels for a Single Image

The im2rec tool and IO.ImageRecordIter have multi-label support for a single image. For example, if you have four labels for a single image, you can use the following procedure to use the Recordio tools.

  1. Write the the image list files as follows:

        integer_image_index \t label_1 \t label_2 \t   label_3 \t label_4 \t path_to_image
  2. Run im2rec, adding a ‘label_width=4’ to the command argument, for example:

        ./bin/im2rec image.lst image_root_dir output.bin resize=256 label_width=4
  3. In the iterator generation code, set label_width=4 and path_imglist=<<The PATH TO YOUR image.lst>>, for example:

        val dataiter = IO.ImageRecordIter(Map(
          "path_imgrec" -> "data/cifar/train.rec",
          "data_shape" -> "(3,28,28)",
          "path_imglist" -> "data/cifar/image.lst",
          "label_width" -> "4"

Next Steps

  • NDArray API for vector/matrix/tensor operations
  • KVStore API for multi-GPU and multi-host distributed training