DataSet Could not be found

[Vrishod]

Couldnot find the dataset used for training

[yonkshi]

Hi, the dataset we used were:

1. Pretrained Imagenet
2. Oxford-102 flowers
3. Oxford-102 description: https://github.com/reedscot/cvpr2016 (The author of the original paper used Amazon Mechanical Turks to transcribe the Oxford 102 flowers)

[SreenijaK]

@yonkshi Incase I want to work for my own dataset how do i do the embeddings. How do i convert the textfiles to .t7 format?

[yonkshi]

@SreenijaK You may check out the original GAN-CLS paper, the author mentioned that you can either pretrain the text embedding or train it end to end with the GAN. To pretrain the text embedding you will need to train the text encoder using the pretrained LeNet's embedding layer. LeNet already contains a well defined embedding for images, and you want train your text embedding to be as close to that as possible.

As for the .t7 file, it's simply a torch dataset file (kind of like an h5 file if you are familiar), you can write a custom dataloader if you plan on using t7, I am not using t7 file as I would need to install torch for that to work.