Release Notes for version 1.2

Demos

  • Fixing and cleaned mnist.conf demo, added comments and l2pool. run_type was missing. Disabled training display crashing.
  • Cleaned face demo, added comments. Fixed best_cam.conf for face detection demo (demos/face)

EBLearn and Idx Core Library

  • Introduced a much simpler state mechanism
    • Fixed memory leaks introduced by the much simpler state mechanism :)
  • Added linear_tanh module
  • Added rgb_to_rgb module (doing global normalization only)
  • Added dropout module
  • Added linear_merge_module which is a much more efficient merge+linear module
  • Added saving matrices to strings
  • Added custom picking probabilities for classes in datasource
  • Added eblprint and eblprinto macros and changed all core code to use it.
  • Datasource tries to read as csv if files are not matrices
  • Moved string utils from libeblearntools to libidx to allow csv answers saving in datasource and train
  • Moving resize_module as parent of resizepp_module and adding a resize_as another module capability to resize_module
  • Fixed preprocessing bug for some platforms when all the values of the tensor are the same.
  • Removing global norm tuning for normalized layers in ebl_preprocessing. Global norm is always performed for lcn layers and a tunable parameter for other layers.
  • Removing global norm in new_norm() in ebl_preprocessing, since it is now done in the pp modules directly (changed recently), adding more describe details in modules for better debugging, changing default gaussian coeff in pp norm from 0 to DEFAULT_GAUSSIAN_COEFF (should not make much difference)
  • Fixing bug in merge_module
  • Fixing cross entropy loss
  • Fixing std::pow call ambiguity on some systems
  • Adding valid convolution option to normalizations, adding interface for rgb_to_* modules in netconf, and other minor tweaks
  • Adding automatic setting of replicate_inputs in ms_module so that user does not have to set it
  • Fixed dynset to dataset conversion

EBLearn Tools Library

  • Adding camera_datasource
  • Fixing scale_mode saving
  • Fixing patch dataset
  • Fixing pascalbg_dataset for dynset

CUDA

  • Added cuda_thres
  • Fixing memory leak in cuda div
  • Added cuda fsum
  • added cuda optimizations for linear_merge

Windows

  • Windows fixes for compilation
  • Added face detector test for Windows (previously not implemented)
  • CMake fixes for Windows
  • Disabling scales print in dsdisplay for windows

MAC OSX

  • Compile fixes

Android

  • Android fixes for new NDK, and now it uses STL from the NDK
  • Changed android demo to use new state mechanism
  • Added full configuration files support and libeblearntools support for android (makes it way more functional)
  • Added relevant logging macros, now logcat has same output as console output on PCs
  • Updated android demo to use detection thread and configuration files (makes it way more generic)
  • Fixing rectangle drawing in android demo
  • Various bug fixes

Detect tool

  • Fixed smoothing bug
  • Adding optional specification of manual scales for individual images in detector
  • Fixing detector's auto-alignment of merging
  • Fixing estimation of network stride in detector
  • Adding datasource capability to detect
  • Fixed a functionality bug in bbox_file usage
  • Changing default bbox saving period to 500 to limit wasted I/O time

Metaparse & Metarun tools

  • Adding saving of each job's best model in metaparse
  • Adding possibility to pass extra conf arguments to metarun
  • Minor fix for plotting in metaparser
  • Fixing metarun not running when max_jobs not set

Tester

  • Fixing face detection test, a few dead stores, a few possible logical errors
  • Added CUDA convolution test

Other

  • Added mconvert tool
  • Fixing matlab output format to handle >2gb files
  • Added ebl2matlab tool and added output wrappers for matio library
  • Adding weights permutation option for parameter
  • Added a text mode to maketable tool
  • Adding static saving in dsfprop
  • Fixed loading of single-matrix datasets in dscompile
  • Fixed partial weights loading

Boost Fixes

  • Changed filesystem code to BOOST_FILESYSTEM_VERSION 3
  • Removed depreceated boost functions for newer boost versions
  • Adding try/catch boost exception on file finding function
  • Committed newer version of FindBoost.cmake for cmake to find newer boost versions
release_notes_1.2.txt · Last modified: 2013/01/17 12:33 by soumith