Fully Convolutional Visual Question Answering

This is an attention based model for VQA using a dilated convolutional neural network for modelling the question and a resnet for visual features. The text model is based on the recent convolutional architecture ByteNet. Stacked attention distributions over the images are then used to compute weighted image features, which are concatenated with the text features to predict the answer. Following is the rough diagram for the described model.

Requirements

• Python 2.7.6
• Tensorflow 1.3.0
• nltk

Datasets and Paths

• The model is can be trained either on VQA 1.0 or VQA 2.0. Download the dataset by running sh download.sh in Data directory. Unzip the downloaded files and create the directory Data/CNNModels. Download the pretrained Resnet-152 from here to Data/CNNModels.
• Make 2 empty directories Data/Models1, Data/Models2 for saving the checkpoints while training VQA 1.0 and 2.0 respectively.

Usage

Extract the Image features

• Extract the image features as per the following
  • DEFAULT - Resnet (14,14,2048) block4 features(attention model) - python extract_conv_features.py --feature_layer="block4"
  • VGG (14,14,512) pool5 features(attention model) - python extract_conv_features.py --feature_layer="pool5"
  • VGG fc7 features (4096,) - python extract_conv_features.py --feature_layer="fc7"

Preprocess Questions/Answers

• Tokeinze the questions/answers using python data_loader.py --version=VQA_VERSION (1 or 2)

Training the attention model

• Train using python train_evaluate.py --version=VQA_VERSION
• Following are the customizable model options
  • residual_channels : Number channels in the residual block of bytenet/state of the lstm. Default 512.
  • batch_size : Default 64.
  • learning_rate : default 0.001
  • epochs : Default 25
  • version : VQA dataset version 1 or 2
  • sample_every : sample attention distributions/answers every x steps. Default 200.
  • evaluate_every : Evaluate over validation set every x steps. Default 6000.
  • resume_model : Resume training the model from a checkpoint file.
  • training_log_file : Log accuracy/steps in this filepath. Default 'Data/training_log.json' .
  • feature_layer : Which conv features to use. Default block4 of resnet.
  • text_model : Text model to use : LSTM or bytenet. Default is bytenet

Evaluating a trained model

• The accuracy on the validation is logged every evaluate_every steps while training the model in Data/training_log.json.
• Use python train_evaluate.py --evaluate_every=1 --max_steps=1 --resume_model="Trained Model Path (Data/Models<vqa_-version>/model<epoch>.ckpt)" to evaluate a checkpoint.

Generating Answers/Attention Distributions

Pretrained Model

You may download the pretrained model from here. Save the files in Data/Models1.

• Use python generate.py --question="<QUESTION ABOUT THE IMAGE>" --image_file="<IMAGE FILE PATH>" --model_path="<PATH_TO_CHECKPOINT = Data/Models1/model10.ckpt>" to generate answer/attention distributions in Data/gen_samples.

Sample Results

Image	Question	Attention1	Attention2	Predicted Answer
	is she going to eat both pizza			No
	What color is the traffic light			green
	is the persons hair short			Yes
	what musical instrument is beside the laptop			keyboard
	what color hat is the boy wearing			blue
	what are the men doing			eating
	what type of drink is in the glass			orange juice
	is there a house			yes

References

• Show, Ask, Attend, and Answer: A Strong Baseline For Visual Question Answering
• Neural Machine Translation in Linear Time
• Stacked Attention Networks for Image Question Answering
• And other papers/codes from https://github.com/JamesChuanggg/awesome-vqa