hoseong-kim / sae-pytorch

PyTorch Implementation of SAE from AtoZ

sae-pytorch

Original : [MATLAB version]

PyTorch implementation of Semantic AutoEncoder (SAE).

How to Run

1. git clone https://github.com/hoseong-kim/sae-pytorch.git
2. Download 'awa_demo_data.mat'
   • https://drive.google.com/file/d/1l0UVhhIU-SmtJ9hqk7OVOG9zNga9qt_I/view
3. python sae.py

An Implementation of SAE in PyTorch

1. Set CUB, AwA, aP&Y, SUN, and ImageNet datasets.
   • Partially done (only for AwA dataset).
   • Other datasets will also be available soon.
2. Extract deep features from various deep models, e.g., AlexNet, VGG16, VGG19, GoogleNet, Inception_v3, ResNet-18, ResNet-34, ResNet-50, ResNet-101, ResNet-152.
   • Done, but tuning my source code to achieve results in this paper.
   • The source code will be available after reproducing.
3. Train a Semantic AutoEncoder (SAE).
   • Done.
4. Test unseen class data.
   • Done.

Release Note

v1.0

• Bug fix

Download Paper

Semantic Autoencoder for Zero-shot Learning: [Paper Link (arXiv)]

sae-pytorch's Issues

KeyError: 'test_labels'

Hello,
I wanted to redo your experiment, I found some problems. 'test_labels' is undefined in parse_args.

Error in loading awa_demo_data.mat file

Hello @hoseong-kim

Thank you for making this code and data available.

When I run the code it gives error for line 73 saying:
Unknown mat file type, version 117, 98

Could you please provide the .mat file by saving it in MATLAB v7 version?
save('awa_demo_data.mat','-v7')

How to increase accuracy?

How can I increase the accuracy? What are the parameters? Is there any way to add a training loop?

Code breaks for the dataset

Feature type

Did you extract feature in Pytorch pre-trained model?

some question about awa_demo_data.mat

Dear author:
Thank you for your contribution！
I have download the awa_demo_data.mat. I find the dimension of train_data and test_data is [N, d]. I see the d is feature extracted by CNN. You provide the feature directly but I want to know more about the whole feature extraction network. Could you please add this part in project?
Yours sincerely