yuzcccc / vqa-mfb Goto Github PK

The result reported is 68.76%. I use the features from Bottom-Up Attentionhttps://github.com/peteanderson80/bottom-up-attention, and use the same architecture as yours MFH, but only get 68.58% in the test-dev set. Is there anything different from the original MFH+CoAtt+GloVe? What should I pay attention to when re-implement? Thank you very much! Your code is very helpful!

Thank you for your work !

Can you release the pretrained model of MFH+RCN as well?

Hi,
Thanks for the excellent code base.
can i get the feature trained with faster rcnn and backbone resnet152?
I'm waiting for your reply.

Hi Yu, when I am trying to run the code, I noticed that the config.py in mfh_coatt_glove contains some data files such as 'OpenEnded_mscoco_train2014_questions.json', but I have not found these files in the link you provided (http://visualqa.org/download.html). Could you please tell me where can I find those files? Thanks a lot.

The original vqa-mcb didn't have build in support for restarting training which was due to the compact bilinear layer. Can this caffe implementation resume training from the last iteration?

In others paper, such as Bottom-Up Attention,training one network usually converges in 12-18 epochs, which takes in the order of 12-18 hours on a single Nvidia K40 GPU. However, this model takes me at least 9 hours to train an epoch, how many hours do you take in your side? Thanks!

@yuzcccc Hi, Yu. When I do the preprocess for extracting features, I use the workstation with 2 TITAN. But the code use the GPU memory only 4G, the remaining 20G is not used, the code running speed is very slow. Could you tell me how to make full use of the 2 TITAN, and improve the code running speed?

Hi!
How did you generate the FRCN image features? Would you be willing to share that code?

Hi,
Thanks for the excellent code base.
can i get the feature trained with faster rcnn and backbone resnet152?
I'm waiting for your reply.

[ RUN ] LayerFactoryTest/1.TestCreateLayer
src/caffe/test/test_layer_factory.cpp:47: Failure
Value of: layer->type()
Actual: "MatMul2"
Expected: iter->first
Which is: "MatMul"
[ FAILED ] LayerFactoryTest/1.TestCreateLayer, where TypeParam = caffe::CPUDevice (195 ms)

Any advice? I guess the error is due to the custom layer.

从OpenEnded questions中的annotations中怎么选取3000类答案

This is a confirmation question about MFB Baseline. According to paper there should be two layers of LSTM with 1024-D hidden units each, but in the implementation only one LSTM layer with 1024-D is used. Kindly confirm.

Thanks