tree: 8040b185884849fbe8b0bc3fb32444055c53ff02 [path history] [tgz]
  1. README.md
  2. process_data.py
  3. train_largedata.py
  4. train_mpi.py
examples/largedataset_cnn/README.md

Image Classification using Convolutional Neural Networks with a dataset from the filesysetm

Examples inside this folder show how to train CNN models using SINGA for image classification.

It reads the dataset from the filesystem defined by process_data.py. Hence, users can modify process_data.py for their perference of dataset format.

In the current setting, ‘classes.txt’ contains the names of the classes at each line. For example, if it is a food dataset containing three classes, ‘classes.txt’ may read like this:

Fish_and_chips
bagel_and_croissant
bak_kut_teh

Then, the directory ‘/Data/’ contains all the folders for images of different classes, while each folder name should be the same as that appeared in ‘classes.txt’. The name of an image file should not be a concern, but it should be placed inside the folder of the class it belongs to. For the same example above, the folder structure may look like this:

Data/
    Fish_and_chips/
        fish1.jpg
        fish2.jpg
        ...
        chip1.jpg
        ...
    bagel_and_croissant/
        bagel.jpg
        ...
        croissant1.jpg
        croissant2.jpg
        ...
    bak_kut_teh/
        photo.jpg
        photo2.jpg
        ...             

Before running the code, the model folder in examples/cnn should be copied to this directory.

  • train_largedata.py is the training script, which controls the training flow by doing BackPropagation and SGD update.

  • train_mpi.py is the script for distributed training (among multiple nodes) using MPI and NCCL for communication.