def visualise():
mc_ctx = MultiChoiceQuestionManger()
to_sentence = SentenceGenerator(trainset='trainval')
# writer = ExperimentWriter('latex/examples_replay_buffer_rescore')
writer = ExperimentWriter('latex/examples_replay_buffer_rescore_prior')
# d = load_json('vqa_replay_buffer/vqa_replay_low_rescore.json')
d = load_json('vqa_replay_buffer/vqa_replay_low_rescore_prior_05_04.json')
memory = d['memory']
# show random 100
keys = deepcopy(memory.keys())
np.random.seed(123)
np.random.shuffle(keys)
vis_keys = keys[:100]
for i, quest_key in enumerate(vis_keys):
pathes = memory[quest_key]
if len(pathes) == 0:
continue
# if it has valid questions
quest_id = int(quest_key)
image_id = mc_ctx.get_image_id(quest_id)
gt_question = mc_ctx.get_question(quest_id)
answer = mc_ctx.get_gt_answer(quest_id)
head = 'Q: %s A: %s' % (gt_question, answer)
im_file = '%s2014/COCO_%s2014_%012d.jpg' % ('val', 'val', image_id)
im_path = os.path.join(IM_ROOT, im_file)
questions = []
for p in pathes.keys():
conf1, conf2 = pathes[p]
_tokens = [int(t) for t in p.split(' ')]
sentence = to_sentence.index_to_question(_tokens)
descr = '%s (%0.2f-%0.2f)' % (sentence, conf1, conf2)
questions.append(descr)
writer.add_result(image_id, quest_id, im_path, head, questions)
writer.render()
def test():
import json
import numpy as np
from w2v_answer_encoder import MultiChoiceQuestionManger
model = StateClassifier(input_dim=512, phase='test')
model.build()
prob = model.prob
# Load vocabulary
# to_sentence = SentenceGenerator(trainset='trainval')
# create multiple choice question manger
mc_manager = MultiChoiceQuestionManger(subset='val',
answer_coding='sequence')
sess = tf.Session()
# Load model
ckpt = tf.train.get_checkpoint_state(FLAGS.train_dir)
checkpoint_path = ckpt.model_checkpoint_path
saver = tf.train.Saver()
saver.restore(sess, checkpoint_path)
# get data
result = []
reader = StateDataPetcher(batch_size=18,
subset='dev',
shuffle=False,
max_epoch=1)
num = reader.num_samples
for itr in range(num):
feat, label, quest_id = reader.pop_batch()
feed_dict = model.fill_feed_dict([feat])
scores = sess.run(prob, feed_dict=feed_dict)
idx = scores.argmax()
# parse question and answer
assert (np.unique(quest_id).size == 1)
quest_id = quest_id[0]
question = mc_manager.get_question(quest_id)
mc_ans = mc_manager.get_candidate_answers(quest_id)
vaq_answer = mc_ans[idx]
real_answer = mc_ans[label.argmax()]
# add result
result.append({u'answer': vaq_answer, u'question_id': quest_id})
# show results
if itr % 100 == 0:
print('============== %d ============' % itr)
print('question id: %d' % quest_id)
print('question\t: %s' % question)
print('answer\t: %s' % real_answer)
print('VAQ answer\t: %s (%0.2f)' % (vaq_answer, scores[idx]))
quest_ids = [res[u'question_id'] for res in result]
# save results
tf.logging.info('Saving results')
res_file = 'result/rescore_state_dev_dev.json'
json.dump(result, open(res_file, 'w'))
from vqa_eval import evaluate_model
acc = evaluate_model(res_file, quest_ids)
print('Over all accuarcy: %0.2f' % acc)
return acc
def process(method, inf_type='rand'):
if inf_type == 'rand':
res_file = 'result/tmp_bs_RL2_final_%s.json' % method
else:
res_file = 'result/tmp_bs_RL2_final_%s_BEAM.json' % method
if os.path.exists(res_file):
print('File %s already exist, skipped' % res_file)
return
# cands = load_results()
model = _TYPE2Model[method]()
mc_ctx = MultiChoiceQuestionManger(subset='val')
task_data = load_lm_outputs(method, inf_type)
belief_sets = {}
t = time()
num = len(task_data)
for i, ans_key in enumerate(task_data.keys()):
# time it
avg_time = (time() - t)
print('%d/%d (%0.2f sec/sample)' % (i, num, avg_time))
t = time()
# extract basis info
cands = task_data[ans_key]
quest_id = cands[0]['question_id']
# gt_answer = mc_ctx.get_gt_answer(quest_id)
image_id = mc_ctx.get_image_id(quest_id)
image = mc_ctx.get_image_file(quest_id)
# process
gt_question = mc_ctx.get_question(quest_id)
i_scores, i_questions = [], []
for item in cands:
target = item['question']
pred_ans, vqa_score = model.get_score(image_id, target)
# inset check
is_valid = compare_answer(pred_ans, ans_key)
if not is_valid:
continue
i_questions.append(target)
i_scores.append([float(vqa_score), item['score']])
print('%d/%d' % (len(i_questions), len(cands)))
bs_i = {
'image': image,
'image_id': image_id,
'question': gt_question,
'answer': ans_key,
'belief_sets': i_questions,
'belief_strength': i_scores
}
belief_sets[ans_key] = bs_i
save_json(res_file, belief_sets)
def process(delta=0.2):
# w2v_ncoder = SentenceEncoder()
# load gt and answer manager
ctx = MultiChoiceQuestionManger(subset='val')
# load candidates
candidates = load_json('result/var_vaq_beam_VAQ-VARDSDC_full.json')
# load candidate scores
score_list = load_json(
'result/var_vaq_beam_VAQ-VARDSDC_full_oracle_dump.json')
score_d = {item['aug_quest_id']: item['CIDEr'] for item in score_list}
# loop over questions
dataset = {}
unk_image_ids = []
question_id2image_id = {}
for item in candidates:
aug_id = item['question_id']
question = item['question']
image_id = item['image_id']
unk_image_ids.append(image_id)
question_id = int(aug_id / 1000)
score = score_d[aug_id]
question_id2image_id[question_id] = image_id
if question_id in dataset:
assert (question not in dataset[question_id])
dataset[question_id][question] = score
else:
dataset[question_id] = {question: score}
# get stat
unk_image_ids = set(unk_image_ids)
num_images = len(unk_image_ids)
print('Find %d unique keys from %d images' % (len(dataset), num_images))
print('%0.3f questions on average' % (len(dataset) / float(num_images)))
# visualise
vis_keys = dataset.keys()
np.random.shuffle(vis_keys)
for quest_id in vis_keys[:20]:
ans = ctx.get_gt_answer(quest_id)
image_id = ctx.get_image_id(quest_id)
gt = ctx.get_question(quest_id).lower()
print('\ngt: %s' % gt)
for quest, sc in dataset[quest_id].items():
print('%s (%0.3f)' % (quest, sc))
def process(delta = 0.2):
w2v_ncoder = SentenceEncoder()
# load gt and answer manager
ctx = MultiChoiceQuestionManger(subset='train')
# load candidates
candidates = load_json('result/var_vaq_beam_VAQ-VAR_full_kptrain.json')
# load candidate scores
score_list = load_json('result/var_vaq_beam_VAQ-VAR_full_kptrain_oracle_dump.json')
score_d = {item['aug_quest_id']: item['CIDEr'] for item in score_list}
# loop over questions
dataset = {}
unk_image_ids = []
question_id2image_id = {}
for item in candidates:
aug_id = item['question_id']
question = item['question']
image_id = item['image_id']
unk_image_ids.append(image_id)
question_id = int(aug_id / 1000)
score = score_d[aug_id]
question_id2image_id[question_id] = image_id
if question_id in dataset:
assert (question not in dataset[question_id])
dataset[question_id][question] = score
else:
dataset[question_id] = {question: score}
# get stat
unk_image_ids = set(unk_image_ids)
num_images = len(unk_image_ids)
print('Find %d unique keys from %d images' % (len(dataset), num_images))
print('%0.3f questions on average' % (len(dataset) / float(num_images)))
# build tuple
num_pairs = 0
offset = 0
cst_pairs = []
image_ids, quest_ids, question_w2v, answer_w2v = [], [], [], []
num_task = len(dataset)
t = time()
for _i, (quest_id, item) in enumerate(dataset.items()):
if _i % 1000 == 0:
print('processed: %d/%d (%0.2f sec./batch)' % (_i, num_task, time()-t))
t = time()
ans = ctx.get_gt_answer(quest_id)
image_id = ctx.get_image_id(quest_id)
assert(image_id == question_id2image_id[quest_id])
gt = ctx.get_question(quest_id).lower()
gt = ' '.join(word_tokenize(gt))
include_gt = np.any(np.array(item.values()) == 10.)
sc, ps = [], []
if gt not in item and not include_gt:
item[gt] = 10.
for q, s in item.items():
sc.append(s)
ps.append(q)
sc = np.array(sc, dtype=np.float32)
_this_n = len(ps)
path_ind = np.arange(_this_n) + offset
# data checking and assertion
try:
assert (np.sum(sc == 10.) <= 1) # only one gt
except Exception as e:
ind = np.where(sc == 10.)[0]
for _idx in ind:
print('%s' % (ps[_idx]))
raise e
# find contrastive pairs
diff = sc[np.newaxis, :] - sc[:, np.newaxis]
valid_entries = diff >= delta
neg, pos = np.where(valid_entries)
assert (np.all(np.greater_equal(sc[pos] - sc[neg], delta)))
pos_q_ind = path_ind[pos]
neg_q_ind = path_ind[neg]
# save
_this_pairs = [[p, n] for p, n in zip(pos_q_ind, neg_q_ind)]
cst_pairs += _this_pairs
# encode answer
_ans_w2v = w2v_ncoder.encode(ans)
ans_w2v = np.tile(_ans_w2v, [_this_n, 1])
answer_w2v.append(ans_w2v)
# encode questions
for p in ps:
_q_w2v = w2v_ncoder.encode(p)
question_w2v.append(_q_w2v)
image_ids.append(image_id)
quest_ids.append(quest_id)
# update pointer
offset += _this_n
num_pairs += _this_n
print('Total pairs: %d' % num_pairs)
# merge
cst_pairs = np.array(cst_pairs, dtype=np.int32)
image_ids = np.array(image_ids, dtype=np.int32)
quest_ids = np.array(quest_ids, dtype=np.int32)
answer_w2v = np.concatenate(answer_w2v, axis=0).astype(np.float32)
question_w2v = np.concatenate(question_w2v, axis=0).astype(np.float32)
from util import save_hdf5
sv_file = 'result/cst_ranking_kptrain_delta%g.data' % delta
save_hdf5(sv_file, {'cst_pairs': cst_pairs,
'image_ids': image_ids,
'quest_ids': quest_ids,
'answer_w2v': answer_w2v,
'question_w2v': question_w2v})
class PredictionVisualiser(object):
def __init__(self, model_name, K=3, do_plot=True):
self._gt_mgr = MultiChoiceQuestionManger(subset='trainval',
load_ans=True)
self._rev_map = SentenceGenerator(trainset='trainval')
self._top_k = K
self._do_plot = do_plot
self._model_name = model_name
self._cache_dir = 'att_maps/%s' % self._model_name
mkdir_if_missing(self._cache_dir)
def plot(self, quest_id, scores, att_map):
if type(quest_id) != int:
quest_id = int(quest_id)
scores = scores.flatten()
if scores.size == 2001:
scores[-1] = 0
# show question and gt answer
question = self._gt_mgr.get_question(quest_id)
gt_ans = self._gt_mgr.get_gt_answer(quest_id)
print('\n====================================')
print('Q: %s' % question)
print('A: %s' % gt_ans)
# show top k prediction
index = (-scores).argsort()[:self._top_k]
for idx in index:
pred_ans = self._rev_map.index_to_top_answer(idx)
print('P: %-20s\t(%0.2f)' % (pred_ans, scores[idx]))
print('\n')
# show image
im_file = self._gt_mgr.get_image_file(quest_id)
im = imread(im_file)
if np.rank(im) == 2:
im = np.tile(im[::, np.newaxis], [1, 1, 3])
if self._do_plot:
imshow(im)
plt.show()
else:
self.save_cache_file(quest_id, im, att_map, question)
return
# show attention map
tokens = _tokenize_sentence(question)
self._show_attention_map(im, att_map, tokens)
def save_cache_file(self, quest_id, im, att_map, question):
from scipy.io import savemat
sv_path = os.path.join(self._cache_dir, '%d.mat' % quest_id)
savemat(sv_path, {'im': im, 'att_map': att_map, 'quest': question})
def _show_attention_map(self, im, att_map, tokens):
att_map = att_map.reshape([-1, 14, 14])
num = att_map.shape[0]
if num == 1:
tokens = [' '.join(tokens)] # merge to a sentence
else:
tokens = [' '.join(tokens)] # merge to a sentence
# mean_map = att_map.mean(axis=0)[np.newaxis, ::]
# att_map = np.concatenate([att_map, mean_map], axis=0)
# tokens.append('average')
# render and plot
for i, am in enumerate(att_map):
am = resize(am, im.shape[:2], preserve_range=True)
am = am / am.max()
v = im * am[:, :, np.newaxis]
v = np.minimum(np.round(v).astype(np.uint8), 255)
if self._do_plot:
imshow(v)
plt.title('%s <%d>' % (tokens[0], i))
plt.show()
def main():
split = 'test'
data_file = 'data/ivqa_multiple_choices_%s_questions.pkl' % split
print(data_file)
mc_ctx = MultiChoiceQuestionManger(subset='val', load_ans=True)
# find unique questions
# questions = load_questions()
question_ids = mc_ctx.get_question_ids()
question_dict = {}
answer_dict = {}
question_id2question_key = {}
# set question and answer keys
unique_question_idx = 0
answer_idx = 0
for i, quest_id in enumerate(question_ids):
if i % 1000 == 0:
print('Metric Maker: parsed %d/%d questions' % (i, len(question_ids)))
question = mc_ctx.get_question(quest_id)
quest_key = _generate_key(question)
question_id2question_key[quest_id] = quest_key
if quest_key in question_dict:
question_dict[quest_key]['counts'] += 1
else:
question_dict[quest_key] = {'counts': 1, 'key_idx': unique_question_idx}
unique_question_idx += 1
# parse answers
mc_answer = mc_ctx.get_gt_answer(quest_id)
answer_key = _generate_key(mc_answer)
if answer_key in answer_dict:
answer_dict[answer_key]['quest_id'].append(quest_id)
else:
answer_dict[answer_key] = {'quest_id': [quest_id],
'answer_idx': answer_idx}
answer_idx += 1
# sort questions by answer type
quest_vocab = QuestionVocab(question_dict, question_id2question_key)
quest_index_by_answer_type = sort_questions_by_answer_type(mc_ctx, quest_vocab)
# build basic data structure for iVQA
dataset = build_candidate_answers(answer_dict, split=split)
# add popular questions
dataset = add_popular_questions(dataset, mc_ctx, quest_vocab,
quest_index_by_answer_type)
# add contrastive questions
dataset = add_contrastive_questions(dataset, mc_ctx, quest_vocab, num=100)
# add plausible questions
dataset = add_plausible_questions(dataset, quest_vocab, num=100)
# add random questions
dataset = add_random_questions(dataset, mc_ctx, quest_vocab,
answer_dict, quest_index_by_answer_type, num=200)
# save data
data_file = 'data/ivqa_multiple_choices_%s_questions.pkl' % split
pickle(data_file, {'dataset': dataset, 'quest_vocab': quest_vocab})