This is an experimental Torch implementation of the VIS + LSTM visual question answering model from the paper Exploring Models and Data for Image Question Answering by Mengye Ren, Ryan Kiros & Richard Zemel.

Requirements:

• Torch
• loadcaffe

Download the MSCOCO train+val images and VQA data using sh data/download_data.sh. Extract all the downloaded zip files inside the data folder.

unzip Annotations_Train_mscoco.zip
unzip Questions_Train_mscoco.zip
unzip train2014.zip

unzip Annotations_Val_mscoco.zip
unzip Questions_Val_mscoco.zip
unzip val2014.zip

If you had them downloaded already, copy over the train2014 and val2014 image folders and VQA JSON files to the data folder.

Download the VGG-19 Caffe model and prototxt using sh models/download_models.sh.

• To avoid memory issues with LuaJIT, install Torch with Lua 5.1 (TORCH_LUA_VERSION=LUA51 ./install.sh). More instructions here.
• If working with plain Lua, luaffifb may be needed for loadcaffe, unless using pre-extracted fc7 features.

th extract_fc7.lua -split train
th extract_fc7.lua -split val

• batch_size: Batch size. Default is 10.
• split: train/val. Default is train.
• gpuid: 0-indexed id of GPU to use. Default is -1 = CPU.
• proto_file: Path to the deploy.prototxt file for the VGG Caffe model. Default is models/VGG_ILSVRC_19_layers_deploy.prototxt.
• model_file: Path to the .caffemodel file for the VGG Caffe model. Default is models/VGG_ILSVRC_19_layers.caffemodel.
• data_dir: Data directory. Default is data.
• feat_layer: Layer to extract features from. Default is fc7.
• input_image_dir: Image directory. Default is data.

th train.lua

• rnn_size: Size of LSTM internal state. Default is 512.
• num_layers: Number of layers in LSTM
• embedding_size: Size of word embeddings. Default is 512.
• learning_rate: Learning rate. Default is 4e-4.
• learning_rate_decay: Learning rate decay factor. Default is 0.95.
• learning_rate_decay_after: In number of epochs, when to start decaying the learning rate. Default is 15.
• alpha: Alpha for adam. Default is 0.8
• beta: Beta used for adam. Default is 0.999.
• epsilon: Denominator term for smoothing. Default is 1e-8.
• batch_size: Batch size. Default is 64.
• max_epochs: Number of full passes through the training data. Default is 15.
• dropout: Dropout for regularization. Probability of dropping input. Default is 0.5.
• init_from: Initialize network parameters from checkpoint at this path.
• save_every: No. of iterations after which to checkpoint. Default is 1000.
• train_fc7_file: Path to fc7 features of training set. Default is data/train_fc7.t7.
• fc7_image_id_file: Path to fc7 image ids of training set. Default is data/train_fc7_image_id.t7.
• val_fc7_file: Path to fc7 features of validation set. Default is data/val_fc7.t7.
• val_fc7_image_id_file: Path to fc7 image ids of validation set. Default is data/val_fc7_image_id.t7.
• data_dir: Data directory. Default is data.
• checkpoint_dir: Checkpoint directory. Default is checkpoints.
• savefile: Filename to save checkpoint to. Default is vqa.
• gpuid: 0-indexed id of GPU to use. Default is -1 = CPU.

th predict.lua -checkpoint_file checkpoints/vqa_epoch23.26_0.4610.t7 -input_image_path data/train2014/COCO_train2014_000000405541.jpg -question 'What is the cat on?'

• checkpoint_file: Path to model checkpoint to initialize network parameters from
• input_image_path: Path to input image
• question: Question string

Randomly sampled image-question pairs from the VQA test set, and answers predicted by the VIS+LSTM model.

Q: What animals are those? A: Sheep

Q: What color is the frisbee that's upside down? A: Red

Q: What is flying in the sky? A: Kite

Q: What color is court? A: Blue

Q: What is in the standing person's hands? A: Bat

Q: Are they riding horses both the same color? A: No

Q: What shape is the plate? A: Round

Q: Is the man wearing socks? A: Yes

Q: What is over the woman's left shoulder? A: Fork

Q: Where are the pink flowers? A: On wall

• Last hidden layer image features from VGG-19
• Zero-padded question sequences for batched implementation
• Training questions are filtered for top_n answers, top_n = 1000 by default (~87% coverage)

To reproduce results shown on this page or try your own image-question pairs, download the following and run predict.lua with the appropriate paths.

• vqa_epoch23.26_0.4610.t7 (Serialized using Lua51) [GPU] [CPU]
• answers_vocab.t7
• questions_vocab.t7
• data.t7

• Exploring Models and Data for Image Question Answering, Ren et al., NIPS15
• VQA: Visual Question Answering, Antol et al., ICCV15

MIT