3 This is a port of the Synthetic Visual Reasoning Test problems to the
4 pytorch framework, with an implementation of two convolutional
5 networks to solve them.
7 # Installation and test #
16 should generate an image example.png in the current directory.
18 Note that the image generation does not take advantage of GPUs or
19 multi-core, and can be as fast as 10,000 vignettes per second and as
20 slow as 40 on a 4GHz i7-6700K.
22 # Vignette generation and compression #
26 The svrtset.py implements the classes `VignetteSet` and
27 `CompressedVignetteSet` with the following constructor
30 __init__(problem_number, nb_samples, batch_size, cuda = False, logger = None)
33 and the following method to return one batch
36 (torch.FloatTensor, torch.LongTensor) get_batch(b)
39 ## Low-level functions ##
41 The main function for genering vignettes is
44 torch.ByteTensor svrt.generate_vignettes(int problem_number, torch.LongTensor labels)
49 * `problem_number` indicates which of the 23 problem to use
50 * `labels` indicates the boolean labels of the vignettes to generate
52 The returned ByteTensor has three dimensions:
58 The two additional functions
61 torch.ByteStorage svrt.compress(torch.ByteStorage x)
67 torch.ByteStorage svrt.uncompress(torch.ByteStorage x)
70 provide a lossless compression scheme adapted to the ByteStorage of
71 the vignette ByteTensor (i.e. expecting a lot of 255s, a few 0s, and
74 This compression reduces the memory footprint by a factor ~50, and may
75 be usefull to deal with very large data-sets and avoid re-generating
76 images at every batch. It induces a little overhead for decompression,
77 and moving from CPU to GPU memory.
79 See vignette_set.py for a class CompressedVignetteSet using it.
81 # Testing convolution networks #
89 provides the implementation of two deep networks, and use the
90 compressed vignette code to allow the training with several millions
91 vignettes on a PC with 16Gb and a GPU with 8Gb.
93 The networks were designed by Afroze Baqapuri during an internship at