save_QA.py

#!/usr/bin/env python
# coding: utf-8

# In[1]:


import json
import h5py
import numpy as np
import copy
from random import shuffle, seed
import sys
import os.path
import argparse
import glob
import numpy as np
import scipy.io
import pdb
import string
import h5py
import nltk
nltk.download('punkt')
from nltk.tokenize import word_tokenize
import gensim
import json
import re
import cv2
import matplotlib.pyplot as plt


# In[2]:



def extract_feat(doc):
    feat = []
    for word in doc:
        try:
            feat.append(model_w2v[word])
        except:
            pass
    return feat

def tokenize(sentence):
    return [i for i in re.split(r"([-.\"',:? !\$#@~()*&\^%;\[\]/\\\+<>\n=])", sentence) if i!='' and i!=' ' and i!='\n'];

def prepro_question(imgs, method):
  
    # preprocess all the question
    print('example processed tokens:')
    for i,img in enumerate(imgs):
        s = img['question']
        if method == 'nltk':
            txt = word_tokenize(str(s).lower())
        else:
            txt = tokenize(s)
        img['processed_tokens'] = txt
        if i < 10: print(txt)
        if i % 1000 == 0:
            sys.stdout.write("processing %d/%d (%.2f%% done)   \r" %  (i, len(img), i*100.0/len(imgs)) )
            sys.stdout.flush()   
    return imgs

def get_top_answers(imgs, num_ans):
    counts = {}
    for img in imgs:
        ans = img['ans'] 
        counts[ans] = counts.get(ans, 0) + 1

    cw = sorted([(count,w) for w,count in counts.items()], reverse=True)
    print('top answer and their counts:') 
    print('\n'.join(map(str,cw[:20])))
    
    vocab = []
    for i in range(num_ans):
        vocab.append(cw[i][1])

    return vocab[:num_ans]


def filter_question(imgs, atoi):
    new_imgs = []
    for i, img in enumerate(imgs):
        if atoi.get(img['ans'],len(atoi)+1) != len(atoi)+1:
            new_imgs.append(img)

    print('question number reduce from %d to %d '%(len(imgs), len(new_imgs)))
    return new_imgs


# In[3]:

model_path = 'Raw_Data/WordVecs/GoogleNews-vectors-negative300.bin'
model_w2v = gensim.models.KeyedVectors.load_word2vec_format(model_path, binary=True)

imgs_train = json.load(open('Raw_Data/vqa_raw_train.json' , 'r'))
num_ans = 1000
top_ans = get_top_answers(imgs_train, num_ans )
atoi = {w:i for i,w in enumerate(top_ans)}
itoa = {i:w for i,w in enumerate(top_ans)}
feat_dim = 300
imgs_data_train = json.load(open('vqa_final_train.json' , 'r'))
num_ans = 1000
method = 'nltk'
max_length = 26
dir_path = "Final_Data/QA/"
N = len(imgs_data_train)



def save_data():
    
 
        for i,img in enumerate(imgs_data_train):
            #print('X' , img['ques_id'])
            img_path = img['img_path']  

            s = img['question']
            print(i,s)
            if method == 'nltk':
                    txt = word_tokenize(str(s).lower())
            else:
                    txt = tokenize(s)

            img['processed_tokens'] = txt            
            question_id = img['ques_id']
            feat = np.array(extract_feat(img['processed_tokens']))
            label_arrays = np.zeros((1, max_length, feat_dim), dtype='float32')
            label_length = min(max_length, len(feat)) # record the length of this sequence
            label_arrays[0, :label_length, :] = feat
            ans_arrays = atoi[img['ans']]

            f = h5py.File(os.path.join( dir_path , str(question_id) + '.h5'), "w")
            f.create_dataset("ques_train", dtype='float32', data=label_arrays)
            f.create_dataset("answers", dtype='uint32', data=ans_arrays)
            f.close()
        return    
        
 
 
 
data = save_data()