Train a deeper LSTM and normalized CNN Visual Question Answering model. This current code can get 58.16 on Open-Ended and 63.09 on Multiple-Choice on test-standard split. You can check Codalab leaderboard for more details.
This code is written in Lua and requires Torch. The preprocssinng code is in Python, and you need to install NLTK if you want to use NLTK to tokenize the question.
We have prepared everything for you.
If you want to train on train set and evaluate on validation set, you can download the features from here, and the pretrained model from here, put it under the main folder, run
$ th eval.lua -input_img_h5 data_img.h5 -input_ques_h5 data_prepro.h5 -input_json data_prepro.h5 -model_path model/lstm.t7
This will generate the answer json file both on Open-Ended and Multiple-Choice. To evaluate the accuracy of generate result, you need to download the VQA evaluation tools.
If you want to train on train + validation set and evaluate on test set, you can download the feature from here, and the pretrained model from here. The rest is the same as previous one, except you need to evaluate the generated results on testing server.
The first thing you need to do is to download the data and do some preprocessing. Head over to the data/ folder and run
$ python vqa_preprocessing.py --download True --split 1
--download Ture means you choose to download the VQA data from the VQA website and --split 1 means you use COCO train set to train and validation set to evaluation. --split 2 means you use COCO train+val set to train and test set to evaluate. After this step, it will generate two files under the data folder. vqa_raw_train.json and vqa_raw_test.json
Once you have these, we are ready to get the question and image features. Back to the main folder, run
$ python prepro.py --input_train_json data/vqa_raw_train.json --input_test_json data/vqa_raw_test.json --num_ans 1000
to get the question features. --num_ans specifiy how many top answers you want to use during training. You will also see some question and answer statistics in the terminal output. This will generate two files in your main folder, data_prepro.h5 and data_prepro.json. To get the image features, run
$ th prepro_img.lua -input_json data_prepro.json -image_root path_to_image_root -cnn_proto path_to_cnn_prototxt -cnn_model path to cnn_model
Here we use VGG_ILSVRC_19_layers model. After this step, you can get the image feature data_img.h5. We have prepared everything and ready to launch training. You can simply run
$ th train.lua
with the default parameter, this will take about 12 hours on a sinlge Tesla k40 GPU, and will generate the model under model/save/
Jiasen Lu
Xiao Lin
BSD License.