Image processing tools
python setup.py install
Takes list of filenames, writes filename + descriptor to TSV.
# FC2 features
cat filenames | python -m tdesc --model vgg16 > vgg16-fc2.descriptors
# CROW features (sum pooled conv5 features)
cat filenames | python -m tdesc --model vgg16 --crow > vgg16-crow.descriptors
Takes list of filenames, writes filename + descriptor to TSV.
cat filenames | python -m tdesc --model dlib_face --outpath faces.h5
It seems like the AVX_INSTRUCTIONS option in dlib makes a big difference (8x on my box?).
The threaded IO can give a pretty big speedup.
These are benchmarks on my system, reading images over the network. The boost you get from this depends on lots of factors.
# Naive IO
cat urls | ./naive-runner.py --crow > /dev/null
Using Theano backend.
Using cuDNN version 5110 on context None
Mapped name None to device cuda0: TITAN X (Pascal) (0000:04:00.0)
VGG16Worker: ready
0 images | 0.055656 seconds
100 images | 4.867988 seconds
200 images | 10.684320 seconds
300 images | 14.986690 seconds
400 images | 19.420027 seconds
500 images | 25.329047 seconds
600 images | 30.432902 seconds
700 images | 35.608703 seconds
800 images | 40.109027 seconds
900 images | 44.767013 seconds
999 images | 53.446249 seconds
# One thread
cat urls | python -m tdesc --crow --io-threads 1 > /dev/null
Using Theano backend.
Using cuDNN version 5110 on context None
Mapped name None to device cuda0: TITAN X (Pascal) (0000:04:00.0)
VGG16Worker: ready
100 images | 4.254059 seconds
200 images | 7.766741 seconds
300 images | 10.954665 seconds
400 images | 14.537050 seconds
500 images | 18.294582 seconds
600 images | 23.202393 seconds
700 images | 28.520624 seconds
800 images | 32.765243 seconds
900 images | 36.555996 seconds
1000 images | 41.932386 seconds
# Two threads
cat urls | python -m tdesc --crow --io-threads 2 > /dev/null
...
900 images | 15.679864 seconds
1000 images | 17.494891 seconds
# Four threads
cat urls | python -m tdesc --crow --io-threads 4 > /dev/null
...
900 images | 7.591560 seconds
1000 images | 8.437886 seconds
Might be able to get a (modest) speedup by predicting on batches of images, instead of predicting one-by-one.