DIGITS Versions Save

Deep Learning GPU Training System

v3.0.0

8 years ago

See release notes for v3.0 RC.

New since 3.0 RC

  • Fix handling of unencoded LMDBs in Torch (#475)
  • Significant performance enhancement for creating datasets (#491)
  • Various documentation fixes / updates

v3.1.0

8 years ago

New Features

  • Enable multi-GPU for Torch (#480)
  • Add basic login functionality (#463)
  • Allow Torch to fine-tune pretrained models (#499)
  • Allow Caffe to fine-tune from multiple pretrained models (#498)
  • New tutorials
    • Fine-tuning (#500)
    • Siamese networks (#453)
    • Weight initialization (#522)
  • Allow optional specification of image folder during multiple inference (#526)

Changes

  • Torch performance improvements (#368, #390, #441, #339)
  • Disable colormap for "Top N" feature (#481)
  • Better real-time updates for dataset creation (#473)
  • Better display for device_query tool (#497)
  • Display the job directory for all job types (#469)
  • Use Flask "Blueprints" to cleanup routing code (#507)
  • Cleanup and alphabetize imports throughout the project (#501)
  • Removed docs/API.md and docs/FlaskRoutes.md (a05356ebfe0fe462f20143625ec8c942847348de)

Bugfixes

  • Enable importing of LMDBs created with Caffe's convert_imageset tool (#517)

v3.0.0-rc.3

8 years ago

New Features

  • Add Torch7 as an alternative backend to Caffe (#324, #345)
  • Make using python layers easier by [optionally] attaching a python file to each model (#329)
  • Add the ability to clone previous jobs with a click (#334)
  • Update the homepage to show job updates in real-time (#240)
  • Enable mean subtraction by subtracting the mean file as well as subtracting the mean pixel (#321)
  • Support NVcaffe v0.14 (#341, #336)
  • Display the job directory size for each DatasetJob and ModelJob (#309)
  • Add a backend badge (LMDB/HDF5) to DatasetJobs on the homepage (#323)
  • Explore images in LMDB datasets (#331)

Changes

  • Use port 34448 for the digits-server instead of port 8080 (#392)
  • Remove digits-walkthrough (#352)
  • Enforce standard UI for file input fields across different browsers (#325)

Bugfixes

  • Fix PicklingErrors issues on all platforms (#307)
  • Fix issue when running inference on many images at once (#361)

Known Issues

  • Large inference requests (i.e. "Classify many") may cause timeouts or even crashes (#479)
  • Incorrect handling of unencoded LMDB in Torch wrapper (#477)

v2.2.1

8 years ago

New since 2.2.0

  • Fixed snapshot list for previous networks (#285)
  • Fixed parameter counting (#317)

v2.2.0

8 years ago

New Features

  • Add [initial] support for HDF5 datasets (#226)
  • Zoom in on weight/activation visualizations (#267)
  • Add a new page for comparing training results (#195)
  • Add notes to jobs (#283)

Changes

  • Open inference results in a new browser tab (#244)
  • Various improvements for using prebuilt LMDBs (#268)
  • Sort subfolders when parsing a folder of images (#296)
  • Use input_shape instead of input_dim for deploy network prototxt (#231)

Known Issues

  • Using a snapshot from a previous network doesn't work unless the network is on the first page (#285)
  • Parameter counting fails for some layer types (like PReLU) (#317)

v2.1.0

8 years ago

New Features

  • Add support for "Generic Inference" (i.e. non-classification) networks (#189)
  • Display number of learned parameters in a model (#221)
  • Show ground truth in "Classify Many" if provided (#110)
  • Zoom in on a selection of the loss/accuracy graph (#113)
  • Add autocomplete for server-side path input fields (#183)
  • Select max/min images per class when parsing a folder of images (#161)
  • Allow user to download log from CreateDb tasks (#221)
  • Show number of available GPUs on home page (#207)
  • Allow local file upload for image lists (#106)
  • Display DIGITS version in top right of page header (#153) and in the console output (c181797cdf3ce27bf65a22fd39fbc61b95ecaab6)

Changes

  • Double the LMDB map_size when running out of memory instead of setting to 1TB (#209)
    • requires py-lmdb 0.87
  • Rename default GoogLeNet layers and tops (9ff246eed47ec04461956b133495260855168e2e)
  • Add pagination to Previous Networks list (c181797cdf3ce27bf65a22fd39fbc61b95ecaab6)
  • Various changes that help with Windows compatibility (#199)
  • Major refactoring of tests (#192)

Known issues

  • Parameter counting fails for some layer types (like PReLU) (#317)

v2.0.0

8 years ago

New Features

  • Enabled support for multi-GPU Caffe (#92)
    • Select multiple and/or specific GPUs for training (#92, #104)
  • Created new routes for JSON REST API (#134, #136)
  • Started using GPU for inference (#66)
  • Added NVML info about GPU memory/utilization (#93)
  • Enabled ADAGRAD and NESTEROV as alternative solver types (@drozdvadym in #102)
  • Added scripts to download standard datasets MNIST and CIFAR
  • Added option to set server name (#111)
  • Added support for PPM images (#123)
  • Enabled path autocompletion while setting values in the configuration (#96)

Changes

  • Added a python classification example (#147)
  • Subtract mean pixel during training (#169)
  • Added TravisCI integration to run tests (#28)
  • Added Coveralls integration for test coverage
  • Added Landscape integration to inspect code
  • Added auto-generated documentation of the webapp’s HTTP routes
  • Switched to loading config files from new, more logical locations (#96)
  • Started suppressing most of Caffe’s raw output (b382e99b8a143c9bbbf659ba74e67bf2ef12718e, 019bc6ca750601396a502ad0fd2b0d47b239f0d7)
  • Added a CLA

Bugfixes

  • Fixed various OSX platform-specific issues (#32, @trivedigaurav in #94)

Known Issues

v2.0.0-rc3

8 years ago

See release notes for v2.0.0-preview.

New since 2.0 Preview

v2.0.0-preview

8 years ago

New Features

  • Enabled support for multi-GPU Caffe (#92)
    • Select multiple and/or specific GPUs for training (#92, #104)
  • Created new routes for JSON REST API (#134, #136)
  • Started using GPU for inference (#66)
  • Added NVML info about GPU memory/utilization (#93)
  • Enabled ADAGRAD and NESTEROV as alternative solver types (@drozdvadym in #102)
  • Added scripts to download standard datasets MNIST and CIFAR
  • Added option to set server name (#111)
  • Added support for PPM images (#123)
  • Enabled path autocompletion while setting values in the configuration (#96)

Changes

Bugfixes

  • Fixed various OSX platform-specific issues (#32, @trivedigaurav in #94)

Known Issues

v1.1.2

8 years ago

See release notes for v1.1.0.

New since 1.1.0

  • Fixed a few things in the documentation (6ab2d6f8e0541fb92cf157b8d95072f057fa2459)
  • Fixed upgrade path for datasets and jobs created with older versions of DIGITS (6a838e0b44a4352480b889efac848845bccad5fc)