To compat higher version of python and Pytorch, I changed some part of the hengyuan-hu's implementation: https://github.com/hengyuan-hu/bottom-up-attention-vqa . Thanks for him generously sharing the code.
Most importantly, the results of this code is able to get 63.60 now. The accuracy is calculated using the VQA evaluation metric. Due to time restrictions, I just trained few times. If you are interested, I believe it's able to get the higher performance.
By the way, I find a problem exist in weight_norm of Pytorch v1.0, which leads to the performance decline in Python3. And there's the link about this problem in Pytorch forum: https://discuss.pytorch.org/t/is-weight-norm-different-in-different-python-versions/71108. If someone know or interested about this, please don't hesitate to contact me.
Thanks for ptrblck's answer, this issue is solved in higher version of Pytorch, like v1.4. So weight_norm is safe to be used in this model. And now, whatever you train with Python 3.7 or 2.7, you can get the same results.
There are some changes in this code:
- All codes are reformatted and some unnecessary part are trimmed.
- The ways to process and load images are edited.
- To keep succinct, some functions are changed in dataset.
- This code is compatible with 4 dataset: VQA-v1, VQA-v1, CP-v1, CP-v2.
- Small fraction is changed in model to compatible with Pytorch v1.0.0 and python3.
- The ways to train and eval are edited, refer to this code: https://github.com/guoyang9/vqa-prior. Also great thanks!
- Add test support and save_results function.
- tqdm, h5py
- python 3.7
- pytorch (v1.4 is suggested)
- Extract pre-trained image features.
python preprocess-image-rcnn.py - Preprocess questions and answers.
bash tools/process.sh
python main.py --output baseline --gpu 0
python main.py --test --output baseline --gpu 0