DIGITS Versions Save

Since 6.1.0

Bugfixes

• Update for new TF API (#2014)
• Update CI scripts to add some new deps to Caffe build (#1993)
• Update import and API for pydicom 1.0
• Fix label distribution and its view page (#1916)

Since 6.0

New Features

• Added functionality to integrate DIGITS with S3 Endpoints (#1868)
• Added publish to inference server on classification workflow (#1906)

Bugfixes

• Fix frozen graph issue (#1907)
• Fix 404 error for /datasets/inference-form/... from #1888 (#1889)
• Remove timeout assertion (#1859)

Changes

• Various updates on document

Known Issues

• Out of memory error in the semantic-segmentation example when training the FCN AlexNet model on Tesla P100.

See release notes for the 6.0 release candidate.

Since 6.0 RC1

New Features

• Added support for URL prefix (#1803)

Bugfixes

• Fixed loading/saving tensorflow models (#1794)

Changes

• Various updates on document

Known Issues

• Visualization for Caffe models does not currently work. (#1738)

New Features

• Added TensorFlow backend for DIGITS as an alternate to Caffe and Torch (#1714)
• Added examples and support for GANs (#1714)
• Added support for text classification (#1025)
• Added more viewing options for image segmentation (#1188)

Changes

• HTML embedding now defaults to PNG (#1270)
• Images that causes exceptions will now show the file name (#1636)

Bugfixes

• Fixed softmax visualization issue with scaled images (#1647)
• Documentation was changed for model store with official pictures (#1650)
• Fixed Caffe search path in Windows (#1244)
• Fixed image file entry in Sunnybrook inference form (#1237)
• Fixed bugs when visiting nested image folder (#1477)

Known Issues

• Visualization for Caffe models does not currently work. (#1738)

See release notes for the 5.0 release candidate.

New since 5.0 RC

• Enable the DIGITS Model Store (https://github.com/NVIDIA/DIGITS/pull/1308)
• Fix calculations related to batch accumulation for Caffe (https://github.com/NVIDIA/DIGITS/pull/1307)
• Various documentation updates

279 commits since v4.0.0

New Features

• Import pretrained models from a model "store" (#896, #1077, #1161)
• Support for image segmentation workflows (#830, #961, #1131)
• Online data augmentation with Torch (#777)
• Show CPU and system memory utilization during training (#800)
• Improved bounding-box visualizations for object detection models (#869)
• Create groups of jobs for easier display on the home page (#734)
• Reuse data extensions for inference (#1024)
• Support for plugin extensions (#1093, #927, #947)
• Add documentation for the REST API (#964)

Changes

• Use environment variables for configuration instead of a file (#1091)
• Remove digits-server and dependency on gunicorn (#1127)
• digits-devserver is now just a small shell script instead of a Python script (#1121)
• New design for Torch multi-GPU training (#828)
• Add Ubuntu 16.04 support by updating dependency versions (#965)
• Allow testing of only Caffe or only Torch with the testsuite (#1143)
• Return more info when downloading a model tarball or json (#891)

Bugfixes

• Fix bug with Torch and CUDA_VISIBLE_DEVICES (#1130)
• Fix issues with browsers returning incorrectly cached css and js files (#904)

Known Issues

• Training goes on longer than required when using batch accumulation (#1240)

529 commits since v3.0.0

New Features

• Add support for object-detection networks like DetectNet (#735) with documentation (#803)
• Parameter sweep over batch size and learning rate (#708)
• Show accuracy confusion matrix for "Classify Many" (#608)
• Test a model with an LMDB (#638)
• Add basic login functionality (#463)

Changes

• Major revamp of home page (#728, #790)
• Allow use of BVLC/caffe (#769)
• Run inference jobs in separate processes (#573)

Bugfixes

• Made device_query compatible with CUDA 8.0 (#890)

For more information, see the release notes for v3.1, v3.2, v3.3, and the 4.0 RC.

211 commits since v3.3.0

New Features

• Add support for object-detection networks like DetectNet (#735) with documentation (#803)
• Parameter sweep over batch size and learning rate (#708)
• Add plugin systems for data formats (#731) and inference visualizations (#756)
• Expose Caffe's iter_size solver option (#744)
• Add syntax highlighting when editing custom networks (#751)
• View list of related jobs (#767)
• Explore generic datasets (#822)
• Add example for doing text classification with Torch (#684)

Changes

• Major revamp of home page (#728, #790)
• Allow use of BVLC/caffe (#769)
• New Torch multi-GPU programming model (#732)
• Make small improvements to standard networks (#733, #749)
• Set weight_decay to lr / 100 (#792)
• Make major improvements to TravisCI build system (#766, #788)

New Features

• Show accuracy confusion matrix for "Classify Many" (#608)
• Test a model with an LMDB (#638)
• Use layer stages in network descriptions for full control over train/val/deploy networks (#628)
• Option to limit number of images to use for "Classify/Test Many" (#592)
• Better in-app documentation for Python layers (#651)

Changes

• Run inference jobs in separate processes (#573)
• Path autocompletion returns sorted list (#621)

Bugfixes

• Fixed UI bugs when using Safari (#702)
• Fixed file serving for files with absolute paths (#586)
• Fixed some UI bugs related to permissions (#594, #596)
• Various torch-related bugfixes (#661, #663, #681, #686, #699)
• Windows compatibility fixes (#698)

New Features

• Add support for new solvers - RMSprop, AdaDelta and Adam (#564)
• AlexNet for Torch now works for multiple GPUs (#539)
• New documentation for installing CUDA toolkit, drivers, etc. (#558)

Changes

• Only look in one location for config files (#541)
• Re-use weights when retraining a model on the same dataset (#538)
• Functional improvements and documentation changes for examples/classification (#559, #557, #579, #582)
• Better error-checking for caffe networks referencing invalid layer "bottoms" (#576)

Bugfixes

• Fixes for multistep learning rate (#549, #550)