DIGITS Versions Save

New Features

• Add GoogLeNet as a default network (#11)
• "Classify Many Images" shows classification results of many images at once (#61)
• Show statistics (mean, standard deviation, histogram of values) for each layer of the network at inference time (#67)
• Allow saving images in database with PNG encoding (#73)
• Optionally turn off shuffling when creating a dataset (#72)
• Optionally provide a random seed to caffe (73fe257)

Changes

• Upgrade to NVIDIA/caffe version 0.11.0 (e2bcb27)
• Update pip requirements list to match packages available on Ubuntu 14.04 where possible (4162db4, 133213d)
• Use C3.js instead of Google Charts to enable DIGITS to run without an internet connection (#34)
• Change default image resize mode from HALF_CROP to SQUASH (b4f3261)

Bugfixes

• Save images in BGR order instead of RGB because caffe uses OpenCV to read encoded images (#59)
• Scale the LeNet standard network by the standard deviation of MNIST (~80) during train, val and test phases (5a38aa5, 23c1a78)
• Use a white background when removing transparency from images (#85)

Known Issues

• The GoogLeNet standard network is not behaving correctly when trained on the full ImageNet dataset (#82)
• "Classify Many Images" may timeout if too many images are uploaded and the server takes too long to respond (#70)