Skip to content
Towards Diverse and Natural Image Descriptions via a Conditional GAN
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
utils
.gitignore
LICENSE
README.md
dDotprod.lua basic codes added Dec 2, 2017
dLSTM.lua basic codes added Dec 2, 2017
dSeq.lua
decoderDotprod.lua
decoderHn.lua
extractFeatFromCNN.lua basic codes added Dec 2, 2017
gLSTM.lua basic codes added Dec 2, 2017
gNoise.lua
gSeq.lua basic codes added Dec 2, 2017
gSeqCrit.lua
policyCrit.lua
trainGANPolicyFeatNoise.lua
trainGANSeqPolicy.lua

README.md

Code of Towards Diverse and Natural Image Descriptions via a Conditional GAN

Based on Neuraltalk2 and SeqGAN. Special thanks to the authors!

Testing codes and related codes will be added gradually.

Brief Explanation

  • Without images as conditions

    • utils/DataLoaderSeq.lua
    • utils/netUtilsSeq.lua
    • utils/netTrainSeq.lua
    • dSeq.lua
    • gSeq.lua
    • trainGANSeqPolicy.lua
    • Can be used to validate the codes for rollouts. If there are bugs, g will eventually generate nonsense
  • With images as conditions, without finetuning the image encoder

    • utils/DataLoader.lua
    • utils/netUtils.lua
    • utils/netTrain.lua
    • dDotprod.lua
    • gNoise.lua
    • trainGANPolicyFeatNoise.lua
You can’t perform that action at this time.