This is a port of the Synthetic Visual Reasoning Test problems to the pytorch framework, with an implementation of two convolutional networks to solve them.
Executing
make -j -k
./test-svrt.py
should generate an image
example.png in
the current directory.
Note that the image generation does not take advantage of GPUs or multi-core, and can be as fast as 10,000 vignettes per second and as slow as 40 on a 4GHz i7-6700K.
The file svrtset.py implements the classes VignetteSet and
CompressedVignetteSet both with a constructor
__init__(problem_number, nb_samples, batch_size, cuda = False, logger = None)
and a method
(torch.FloatTensor, torch.LongTensor) get_batch(b)
which returns a pair composed of a 4d 'input' Tensor (i.e. single channel 128x128 images), and a 1d 'target' Tensor (i.e. Boolean labels).
The main function for generating vignettes is
torch.ByteTensor svrt.generate_vignettes(int problem_number, torch.LongTensor labels)
where
problem_numberindicates which of the 23 problem to uselabelsindicates the boolean labels of the vignettes to generate
The returned ByteTensor has three dimensions:
- Vignette index
- Pixel row
- Pixel col
The two additional functions
torch.ByteStorage svrt.compress(torch.ByteStorage x)
and
torch.ByteStorage svrt.uncompress(torch.ByteStorage x)
provide a lossless compression scheme adapted to the ByteStorage of the vignette ByteTensor (i.e. expecting a lot of 255s, a few 0s, and no other value).
This compression reduces the memory footprint by a factor ~50, and may be usefull to deal with very large data-sets and avoid re-generating images at every batch. It induces a little overhead for decompression, and moving from CPU to GPU memory.
See vignette_set.py for a class CompressedVignetteSet using it.
The file
cnn-svrt.py
provides the implementation of two deep networks designed by Afroze
Baqapuri during an internship at Idiap, and allows to train them with
several millions vignettes on a PC with 16Gb and a GPU with 8Gb.