Implementation of Paper Show, Ask, Attend, and Answer: A Strong Baseline For Visual Question Answering in PyTorch.

This implementation is slgihtly focused on VQA_v2.0 DataSet

pip install git+https://github.com/nikhilkarnwal/CVML.git 

• Download data from VQA_v2.0 for both training and validation and unzip it.

• Import ailabs library, and first preprocess images and text(question and answers) and then start training.

    import json
    import os
    import torch
    from ailabs.vqa import VQANet
    from ailabs import config
    from ailabs.vqa.src import train, FeatureExtractor
    from ailabs.vqa.src.dataset import process, ptext, pimages
    from ailabs.vqa.src.config import Config
  
    config.device = 'cuda'
    config.dir = './saved_models'
    if not os.path.exists(config.dir):
        os.mkdir(config.dir)
    config.train_path='./train2014'
    config.val_path='./val2014'
    config.preprocessed_path = os.path.join(config.dir, 'resnet-14x14.h5')
    config.vocabulary_path = os.path.join(config.dir, 'vocab.json')
    config.data_workers = 8
    vqa_v = 'v2'
    config.vqa_data['train_q'] = '{0}_{1}_{2}_{3}_{4}.json'.format(vqa_v, config.task, config.dataset, 'train2014',
                                                                   'questions')
    config.vqa_data['val_q'] = '{0}_{1}_{2}_{3}_{4}.json'.format(vqa_v, config.task, config.dataset, 'val2014',
                                                                 'questions')
    config.vqa_data['train_a'] = '{0}_{2}_{3}_{4}.json'.format(vqa_v, config.task, config.dataset, 'train2014',
                                                               'annotations')
    config.vqa_data['val_a'] = '{0}_{2}_{3}_{4}.json'.format(vqa_v, config.task, config.dataset, 'val2014',
                                                             'annotations')
    config.qa_path = './'
    config.preprocess_batch_size=6
  
    torch.cuda.empty_cache()
    import gc
    gc.collect()
    print(json.dumps(config.vqa_data))
    net = FeatureExtractor()
    pimages.execute(net)
    ptext.execute()
  
    # params for training
    config.batch_size = 64
    config.epochs = 50
    config.resume = None # change it to checkpoint in case you want to resume from there
  
  
    def getattr(obj):
        return {name: obj.__getattribute__(name) for name in dir(obj) if not name.startswith('_')}
  
    # Check all config
    print(json.dumps(getattr(config), indent=4))
    # start training
    train.execute()
  

It is build and tested on python=3.7.4

• pytroch=1.6
• tqdm
• h5py

Nikhil Karnwal nikhilkarnwal93@gmail.com