def preprocessAnswers(answers, vocab, fAppend):
ret = []
for e in answers:
e = e['answer'].lower().split()
x = nlp.applyVocab(e, vocab).tolist()
ret += [x]
ret = np.asarray(ret)
with h5py.File('data/preprocessed/clevr.h5', 'a') as f:
data = f.create_dataset(fAppend+'Answers', data=ret)
def preprocessQuestions(questions, vocab, fAppend, maxLen=45):
ret = []
retImgIdx= []
for e in questions:
retImgIdx += [e['image_index']]
e = (e['question'].lower()[:-1] + ' ?').split()
x = nlp.applyVocab(e, vocab).tolist()
x += [0]*(maxLen - len(x))
ret += [x]
ret = np.asarray(ret)
retImgIdx = np.asarray(retImgIdx)
with h5py.File('data/preprocessed/clevr.h5', 'a') as f:
data = f.create_dataset(fAppend+'Questions', data=ret)
data = f.create_dataset(fAppend+'ImageIdx', data=retImgIdx)
def preprocessPrograms(programs, vocab, fAppend, maxLen=45):
#'_' quick vocab patch
for k in list(vocab.keys()):
vocab[k[2:]] = vocab[k]
ret = []
retMask = []
for p in programs:
p = p['program']
p = BTree(p).flat()
p = nlp.applyVocab(p, vocab).tolist()
retMask += [len(p)]
p = [p + (45-len(p))*[0]]
ret += p
ret = np.asarray(ret).astype(np.int)
retMask = np.asarray(retMask)
with h5py.File('data/preprocessed/clevr.h5', 'a') as f:
data = f.create_dataset(fAppend+'Programs', data=ret)
data = f.create_dataset(fAppend+'ProgramMask', data=retMask)