https://github.com/Microsoft/CNTK
Revision 13a4f80fcb9998d8ec70345d5923c915e55100a9 authored by Momcilo Vasilijevic on 16 September 2016, 16:20:56 UTC, committed by Momcilo Vasilijevic on 31 October 2016, 17:23:19 UTC
There are two types of problems with reporting mean IOU error
1) Final error in CNTK is calculated as average of mini-batch errors.
Since mean IOU is calculated by accumulating intersection and union,
result for the last mini batch should be used instead of average.
2) Reported error is averged with number of processed samples (minibatch
size) which causes problems when working with minibatch size bigger than
1. In that case, reported error will be smaller <minibatch_size> times.
Adding logic that decides whether evaluation nodes accumulated error based
on network topology.
In case of distributed calculation, average of aggregate errors calculated
on different nodes will be reported.
Refactor criteria accumulation. Criteria now knows if evaluation node
is accumulating error or not.
1 parent 2bc795c
Raw File
Tip revision: 13a4f80fcb9998d8ec70345d5923c915e55100a9 authored by Momcilo Vasilijevic on 16 September 2016, 16:20:56 UTC
Fix error reporting for mean IOU
Tip revision: 13a4f80
.gitattributes
*.ini text
*.simple text
*.counts text
*.labels text
*.feats text
*.ctf text
*.post text
*.cpu text
*.gpu text
*.rst text

.gitattributes text
.gitignore text
.gitmodules text

.clang-format text

*.md text
*.txt text
*.TXT text
*.html text
*.lyx text
*.bib text
.gitkeep text
KaldiReaderReadme text
Readme text

*.bat text
*.cmd text
*.py text
*.ipynb text
*.pl text
*.ps1 text
*.ps text
*.i text

*.sh text eol=lf
build-and-test text eol=lf
configure text eol=lf
generate_build_info text eol=lf
run-test text eol=lf
run-test-common text eol=lf
run-timit-test-common text eol=lf
run-boost-test-common text eol=lf
make_binary_drop_linux text eol=lf

# Used from Unix / Cygwin 'md5sum -c', needs to have LF line endings:
Tests/EndToEndTests/Examples/Speech/TIMIT/WriteBottleneck/expected_output_md5sum.*.txt eol=lf
Tests/EndToEndTests/Examples/Speech/TIMIT/WriteScaledLogLike/expected_output_md5sum.*.txt eol=lf

# Used by reader unit test, needs to keep LF line endings.
Tests/UnitTests/ReaderTests/Data/CNTKTextFormatReader/invalid_inputs.txt eol=lf

Makefile text
*.sln text
*.pyproj text
*.vcxproj text
*.vcxproj.filters text
*.vssettings text
*.csproj text
*.props text
*.asax text

*.h text
*.cpp text
*.cc text
*.cu text
*.cuh text
*.proto text

*.cs text

*.yml text
*.yml.bak text

*.config text
*.cntk text
*.bs text
*.ndl text
*.mel text
*.xml text
*.m text
*.mlf text
*.scp text
*.list text

*.log text

Dockerfile* text

# Speech data
mean.363 text
var.363 text
prior.132 text

# dot (graph description language) file 
*.dot text 

# AMI-specific
Results text
40fbank.conf text
80fbank.conf text

# SLU-specific
atis.*.simple

# TIMIT-specific
*.fbank24_zda text
TIMIT*.100 text
TIMIT*.20 text
TIMIT*.arpa text
TIMIT*.cntk text
TIMIT*.dnn_map text
TIMIT*.fbank text
TIMIT*.fullpath text
TIMIT*.lookahead text
TIMIT*.rnn text
TIMIT*.statelist text
TIMIT*.tfsa text
TIMIT*.transitions text

Examples/Text/ATIS/Data/ATIS.* text
Examples/Text/ATIS/*.wl text

Examples/SequenceToSequence/CMUDict/Data/cmudict-0.7b* text

# Binary extensions:
*.ark binary
*.chunk binary
*.cmf binary
*.docx binary
*.jpg binary
*.pdf binary
*.png binary
*.pptx binary
*.snk binary
*.vsdm binary
*.zip binary
*.dnn binary
Examples/Image/Detection/FastRCNN/fastRCNN/*/*.pyd binary
back to top