Skip to content
'DenseCap: Fully Convolutional Localization Networks for Dense Captioning' replicate study with `tensorflow`
Branch: master
Clone or download
rampage644 Update Readme
Discontinue project.
Latest commit 125b1dc Jan 11, 2017
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
densecap Restore box loss, tweak total loss and learning rate Jan 10, 2017
tests Split model Jan 10, 2017
.gitignore Initial commit Dec 10, 2016
LICENSE Initial commit Dec 10, 2016
README.md Update Readme Jan 11, 2017
build_vg_data.py Fix TFRecords create Dec 16, 2016
infer.py Update infer routine Jan 10, 2017
playground.py Fix the model. Dec 23, 2016
train.py Fix wrong ground truth input Jan 10, 2017

README.md

Description

Replicate study (that is, implementation) of 'DenseCap: Fully Convolutional Localization Networks for Dense Captioning' paper with tensorflow

Current state

Project is frozen. Only RPN part is implemented, but it's now working as intended. It was started as experiment but now it takes too much effort with no or little progress. Better to switch and (maybe) return to it later with more experience and knowledge. Current known problems are:

  • Converting sparse IndexedSlices to a dense Tensor of unknown shape. This may consume a large amount of memory message somewhere inside _box_params_loss fucntion.
  • Memory leak
  • No Localization Layer
  • No Language Layer
  • Model not actually working ;)

Links

Helpful links

You can’t perform that action at this time.