There is no comprehensive documentation for this code. I'm assuming that usage of this code is limited to, for example, either reproducing the xICFG experiments in our paper [Johnson+Zhang,ICML2018] or using it as a reference for developing your own implementation in some convenient platform. And so the minimum information for these purposes is given here.
Image files used as input to training should consist of a 16-byte header and pixels of all the images. An image file for MNIST is provided, and one for SVHN can be downloaded by a script at test/. The files for LSUN datasets are not provided as they are so large. If needed, please make them as follows.
Header
offset 0: number of channels
offset 4: width of each image
offset 8: height of each image
offset 12: number of images
Pixels
Each value of pixels should be represented by one byte (0x00-0xFF), and the pixels of the second image should immediately follow the pixels of the first image, and so forth. The order within an image must be the same as in ppm or pgm files.
That is, an image file should be a 16-byte header immediately followed by the pixel components of ppm/pgm files concatenated.
File naming conventions
A dataset can consist of more than one image file. When a dataset consists of one image file (i.e., one batch), the filename should end with ".xbin". When a dataset consists of, for example, 10 image files (i.e., 10 batches), the files should be named as name.xbin.1of10, name.xbin.2of10, ..., and name.xbin.10of10 (replacing name with whatever you like), and num_batches=10 should be specified in the training parameters.
In our experiments, we made 10 batches (each with 130K bedrooms and 130K living rooms) for LSUN BR+LR and
7 batches (each with 100K towers and 100K bridges) for LSUN T+B, and it is so assumed in the scripts at test/.
The network configuration parameters for the discriminator and approximator are generated by functions in func.sh. The discriminator parameters start with _d_, and the approximator parameters start with _g_. _d_ or _g_ is followed by a layer id, which is a number starting with 0. For example, _d_0layer_type=Act means that layer#0 of the discriminator is an activation layer. Connection between layers is specified by conn, e.g., _g_conn=0-1-2-3-top,1-3. Some of the network parameters are explained in "3.1.2 Layer parameters" of this (for text categorization), and the rest can be figured out, hopefully. Other types of parameters such as cfg_T are self-explanatory. When training is performed (by test/cfg-dcganx.sh, test/cfg-fc2.sh, etc.), the parameters are shown at the beginning of a log file at test/log.
In case you need a reference for developing your own implementation of xICFG, relevant source code files are at bin/src/img/. Look for the method cfg_train in AzpG_Cfg.cpp there. Other directories under bin/src are for general purposes and should be ignored.
Please be warned that it may be hard to read as I didn't write this for reference purposes.