def vg_check_pre_stats(): client = MongoClient("mongodb://localhost:27017") db = client.visual_genome_1_2 db_rel_train_all = db.relationships_all_train.find(no_cursor_timeout=True) db_rel_test_all = db.relationships_all_test.find(no_cursor_timeout=True) train_stats = {} test_stats = {} cnt = 0 zl.tick() for db_rel in db_rel_train_all: if cnt % 1000 == 0: print cnt, zl.tock() zl.tick() cnt += 1 for r in db_rel['relationships']: name = r['predicate'] if name not in train_stats: train_stats[name] = 0 train_stats[name] += 1 for db_rel in db_rel_test_all: if cnt % 1000 == 0: print cnt, zl.tock() zl.tick() cnt += 1 for r in db_rel['relationships']: name = r['predicate'] if name not in test_stats: test_stats[name] = 0 test_stats[name] += 1 zl.save('output/train_pre_stats.pkl', train_stats) zl.save('output/test_pre_stats.pkl', test_stats) print zl.sort_dict_by_val(train_stats) print zl.sort_dict_by_val(test_stats)
def setup(self, bottom, top): """Setup the RoIDataLayer.""" self._cur_idx = 0 self.gt_labels = {} self.meta = h5py.File('/home/zawlin/Dropbox/proj/sg_vrd_meta.h5', 'r') self.cache = h5py.File('output/sg_vrd_cache.h5','r') self.cache_mem = {} if os.path.exists('output/cache/sg_vrd_gt.pkl'): self.gt_labels = zl.load('output/cache/sg_vrd_gt.pkl') glog.info('loaded gt data from cache') else: glog.info( 'Preloading gt') zl.tic() for k in self.meta['gt/train'].keys(): rlp_labels = self.meta['gt/train/%s/rlp_labels'%k][...] sub_boxes = self.meta['gt/train/%s/sub_boxes'%k][...].astype(np.float) obj_boxes = self.meta['gt/train/%s/obj_boxes'%k][...].astype(np.float) if sub_boxes.shape[0]>0: zeros = np.zeros((sub_boxes.shape[0],1), dtype=np.float) # first index is always zero since we do one image by one image sub_boxes = np.concatenate((zeros, sub_boxes),axis=1) obj_boxes = np.concatenate((zeros, obj_boxes),axis=1) self.gt_labels[k] = {} self.gt_labels[k]['rlp_labels']=rlp_labels self.gt_labels[k]['sub_boxes']=sub_boxes self.gt_labels[k]['obj_boxes']=obj_boxes glog.info('done preloading gt %f'%zl.toc()) zl.save('output/cache/sg_vrd_gt.pkl',self.gt_labels) self.imids = [] for k in self.gt_labels.keys(): self.imids.append(k) self.imidx =0 random.shuffle(self.imids) # parse the layer parameter string, which must be valid YAML layer_params = yaml.load(self.param_str) self._num_classes = layer_params['num_classes'] self._name_to_top_map = {} # data blob: holds a batch of N images, each with 3 channels idx = 0 top[idx].reshape(cfg.TRAIN.IMS_PER_BATCH, 1024, 50,50) self._name_to_top_map['conv_new_1'] = idx idx += 1 top[idx].reshape(1, 5, 1, 1) self._name_to_top_map['sub_boxes'] = idx idx += 1 top[idx].reshape(1, 5, 1, 1) self._name_to_top_map['obj_boxes'] = idx idx += 1 # labels blob: R categorical labels in [0, ..., K] for K foreground # classes plus background top[idx].reshape(1, 1, 1, 1) self._name_to_top_map['labels'] = idx
def save_test_words_to_pickle(): ilsvrc_words = ['antelope','cup','ball','cooking'] model = gensim.models.Word2Vec.load_word2vec_format('./GoogleNews-vectors-negative300.bin', binary=True) word2vec = {} for w in ilsvrc_words: w = w.replace('_',' ') word2vec[w] = model[w] zl.save('test_word2vec',word2vec)
def save_ilsvrc_vectors_to_pickle(): ilsvrc_words = get_ilsvrc_word_list() model = gensim.models.Word2Vec.load_word2vec_format('./GoogleNews-vectors-negative300.bin', binary=True) word2vec = {} for w in ilsvrc_words: w = w.replace('_',' ') if w in model: word2vec[w] = model[w] else: print '%s not in model'%w zl.save('ilsvrc_word2vec',word2vec)
def vg_count_only_one_triplet(): client = MongoClient("mongodb://localhost:27017") db = client.visual_genome_1_2 db_results = db.relationships_all_train.find(no_cursor_timeout=True) cnt = 0 spo_info = {} spo_list = [] zl.tick() for doc in db_results: id = doc['image_id'] cnt += 1 if cnt % 1000 == 0: print cnt, zl.tock() zl.tick() rcnt = 0 for r in doc['relationships']: pre = r['predicate'] sub_name = r['subject']['name'] obj_name = r['object']['name'] spo = sub_name + '_' + pre + '_' + obj_name if spo not in spo_info: spo_info[spo] = 0 spo_info[spo] += 1 db_results_2 = db.relationships_all_test.find(no_cursor_timeout=True) for doc in db_results_2: id = doc['image_id'] cnt += 1 if cnt % 1000 == 0: print cnt, zl.tock() zl.tick() rcnt = 0 for r in doc['relationships']: pre = r['predicate'] sub_name = r['subject']['name'] obj_name = r['object']['name'] spo = sub_name + '_' + pre + '_' + obj_name if spo not in spo_info: spo_info[spo] = 0 spo_info[spo] += 1 zl.save('output/spo_info_vg.pkl', spo_info) #total_pairs = len(sub_obj_info.keys())+0.0 total_spo = len(spo_info.keys()) + 0.0 one_count = 0 for k in spo_info.keys(): if spo_info[k] >= 5: spo_list.append(k) one_count += 1 #print total_spo,one_count vg_total_annotation_count(spo_list)
def vg_make_meta_visual_phrase(): m = h5py.File('data/vg1_2_meta.h5', 'r', driver='core') h5f = h5py.File('data/vg1_2_vp_meta.h5') triplets = {} cnt = 0 zl.tick() for k in m['gt/train'].keys(): if cnt % 1000 == 0: print cnt, zl.tock() zl.tick() cnt += 1 # sub_boxes = m['gt/train/%s/sub_boxes'%k][...] # obj_boxes = m['gt/train/%s/obj_boxes'%k][...] rlp_labels = m['gt/train/%s/rlp_labels' % k][...] for i in xrange(rlp_labels.shape[0]): # sub_box = sub_boxes[i] # obj_box = obj_boxes[i] rlp_label = rlp_labels[i] # joint_bbox = [min(sub_bbox[0],obj_bbox[0]), min(sub_bbox[1],obj_bbox[1]),max(sub_bbox[2],obj_bbox[2]),max(sub_bbox[3],obj_bbox[3])] s_lbl = zl.idx2name_cls(m, rlp_label[0]) o_lbl = zl.idx2name_cls(m, rlp_label[2]) p_lbl = zl.idx2name_pre(m, rlp_label[1]) spo = '%s_%s_%s' % (s_lbl, p_lbl, o_lbl) # spo = '%d_%d_%d'%(rlp_label[0],rlp_label[1],rlp_label[2]) if spo not in triplets: triplets[spo] = 0 triplets[spo] += 1 zl.save('output/pkl/triplets_train_vp.pkl', triplets) triplets_sorted = zl.sort_dict_by_val(triplets) triplets_ok = [] for k, v in triplets_sorted: triplets_ok.append(k) print k, v triplets_ok = sorted(triplets_ok) triplets_ok = ['__background__'] + triplets_ok for i in xrange(len(triplets_ok)): h5f['meta/tri/idx2name/%d' % i] = triplets_ok[i] h5f['meta/tri/name2idx/%s' % triplets_ok[i]] = i print len(triplets_ok)
def vg_stats_predicate(): client = MongoClient("mongodb://localhost:27017") db = client.visual_genome_1_2 db_results = db.relationships_cannon.find(no_cursor_timeout=True) cnt = 0 mappings = make_mappings() mappings_p = make_p_mappings() wnl = WordNetLemmatizer() spl = SpellingReplacer() sub_obj_info = {} zl.tick() for doc in db_results: id = doc['image_id'] cnt += 1 if cnt % 1000 == 0: print cnt, zl.tock() zl.tick() rcnt = 0 for r in doc['relationships']: pre = r['predicate'] sub_name = r['subject']['name'] obj_name = r['object']['name'] so_pair = sub_name + '_' + obj_name if so_pair not in sub_obj_info: so_info = {'total': 0, 'predicates': []} else: so_info = sub_obj_info[so_pair] so_info['total'] += 1 if pre not in so_info['predicates']: so_info['predicates'].append(pre) zl.save('output/sub_obj_info.pkl', sub_obj_info) #total_pairs = len(sub_obj_info.keys())+0.0 total_pairs = 0.0 total_of_averages = 0.0 for k in sub_obj_info.keys(): so_info = sub_obj_info[k] total_predicates = len(so_info['predicates']) + 0.0 if so_info['total'] < 2: continue total_pairs += 1 total_annotated_pairs = so_info['total'] + 0.0 avg_predicates_for_this_pair = total_predicates / total_annotated_pairs total_of_averages += avg_predicates_for_this_pair total_of_averages /= total_pairs print 'total_pairs = %d' % total_pairs print 'total_of_averages = %d' % total_of_averages
def gen_meta_for_retrieval(): out_pkl = 'output/pkl/vg_retr_meta.pkl' m = h5py.File('/home/zawlin/Dropbox/proj/vg1_2_meta.h5', 'r') rlp_labels = [] files = [] counts = [] cnt = 0 zl.tick() for k in m['gt/test']: if cnt % 100 == 0: print cnt, zl.tock() zl.tick() cnt += 1 gt_rlp_labels = m['gt/test'][k]['rlp_labels'][...] for i in xrange(gt_rlp_labels.shape[0]): gt_rlp_label = gt_rlp_labels[i] if len(rlp_labels) == 0: rlp_labels.append(gt_rlp_label.tolist()) files.append([k]) counts.append(1) continue bInd = np.all(gt_rlp_label == rlp_labels, axis=1) ind = np.arange(len(rlp_labels))[bInd] if len(ind) == 0: rlp_labels.append(gt_rlp_label.tolist()) files.append([k]) counts.append(1) else: files[ind].append(k) counts[ind] = counts[ind] + 1 # rlp_labels.append(gt_rlp_label.tolist()) # files.append([k]) # counts.append(1) rlp_labels = np.array(rlp_labels) files = np.array(files) counts = np.array(counts) ind = np.argsort(counts)[::-1] counts = counts[ind] files = files[ind] rlp_labels = rlp_labels[ind] retr_meta = {'counts': counts, 'files': files, 'rlp_labels': rlp_labels} zl.save(out_pkl, retr_meta)
def vg_count_top(): client = MongoClient("mongodb://localhost:27017") db = client.visual_genome_1_2 db_results = db.relationships_cannon.find(no_cursor_timeout=True) cnt = 0 start = time.time() last = time.time() subjects = {} objects = {} objects_all = {} predicate = {} cnt = 0 for doc in db_results: imid = doc['image_id'] cnt += 1 if cnt % 10000 == 0: print cnt rcnt = 0 for r in doc['relationships']: if r['predicate'] == 'C_FAILED': continue if r['subject']['name'] == 'C_FAILED': continue if r['object']['name'] == 'C_FAILED': continue rcnt += 1 if r['predicate'] not in predicate: predicate[r['predicate']] = 1 else: predicate[r['predicate']] += 1 if r['subject']['name'] not in subjects: subjects[r['subject']['name']] = 1 else: subjects[r['subject']['name']] += 1 if r['object']['name'] not in objects: objects[r['object']['name']] = 1 else: objects[r['object']['name']] += 1 if r['subject']['name'] not in objects_all: objects_all[r['subject']['name']] = 1 else: objects_all[r['subject']['name']] += 1 if r['object']['name'] not in objects_all: objects_all[r['object']['name']] = 1 else: objects_all[r['object']['name']] += 1 zl.save('output/objects_all.pkl', objects_all) zl.save('output/objects.pkl', objects) zl.save('output/subjects.pkl', subjects) zl.save('output/predicate.pkl', predicate) objects_all_sorted = sorted(objects_all.items(), key=operator.itemgetter(1)) objects_sorted = sorted(objects.items(), key=operator.itemgetter(1)) subjects_sorted = sorted(subjects.items(), key=operator.itemgetter(1)) predicate_sorted = sorted(predicate.items(), key=operator.itemgetter(1))
def setup(self, bottom, top): self._cur_idx = 0 self.vgg_data = {} vgg_h5 = h5py.File( "output/precalc/vg1_2_2016_predicate_exp_train.hdf5", 'r') layer_params = yaml.load(self.param_str_) self.imids = [] for k in vgg_h5.keys(): self.imids.append(k) self.imidx = 0 if os.path.exists('output/cache/vg1_2_2016_pre_train_concat.pkl'): self.vgg_data = zl.load( 'output/cache/vg1_2_2016_pre_train_concat.pkl') print 'loaded train data from cache' else: print 'Preloading training data' zl.tick() for k in vgg_h5.keys(): sub_visual = vgg_h5[k]['sub_visual'][...] obj_visual = vgg_h5[k]['obj_visual'][...] pre_label = vgg_h5[k]['pre_label'][...] self.vgg_data[k] = {} self.vgg_data[k]['sub_visual'] = sub_visual self.vgg_data[k]['obj_visual'] = obj_visual self.vgg_data[k]['pre_label'] = pre_label print 'done preloading training data %f' % zl.tock() zl.save('output/cache/vg1_2_2016_pre_train_concat.pkl', self.vgg_data) vgg_h5.close() self._batch_size = layer_params['batch_size'] self.train_data = [] self._name_to_top_map = {} # data blob: holds a batch of N images, each with 3 channels # top[0].reshape(self._batch_size, 4096 * 2 ) top[0].reshape(self._batch_size, 2 * 4096) top[1].reshape(self._batch_size) self._name_to_top_map['visual'] = 0 self._name_to_top_map['label'] = 1
def setup(self, bottom, top): self._cur_idx = 0 self.vgg_data = {} self.gt_labels = {} vgg_h5 = h5py.File("output/precalc/vg1_2_2016_train.hdf5", 'r') if os.path.exists('output/cache/vg1_2_2016_train.pkl'): self.vgg_data = zl.load('output/cache/vg1_2_2016_train.pkl') print 'loaded train data from cache' else: print 'Preloading training data' zl.tick() for k in vgg_h5.keys(): classemes = vgg_h5[k]['classemes'][...] visuals = vgg_h5[k]['visuals'][...] locations = vgg_h5[k]['locations'][...] cls_confs = vgg_h5[k]['cls_confs'][...] self.vgg_data[k] = {} self.vgg_data[k]['classemes'] = classemes self.vgg_data[k]['visuals'] = visuals self.vgg_data[k]['cls_confs'] = cls_confs self.vgg_data[k]['locations'] = locations print 'done preloading training data %f' % zl.tock() zl.save('output/cache/vg1_2_2016_train.pkl', self.vgg_data) vgg_h5.close() self.meta = h5py.File('/home/zawlin/Dropbox/proj/vg1_2_meta.h5', 'r') if os.path.exists('output/cache/vg1_2_2016_gt.pkl'): self.gt_labels = zl.load('output/cache/vg1_2_2016_gt.pkl') print 'loaded gt data from cache' else: print 'Preloading gt' zl.tick() for k in self.meta['gt/train'].keys(): rlp_labels = self.meta['gt/train/%s/rlp_labels' % k][...] sub_boxes = self.meta['gt/train/%s/sub_boxes' % k][...].astype( np.float) obj_boxes = self.meta['gt/train/%s/obj_boxes' % k][...].astype( np.float) self.gt_labels[k] = {} self.gt_labels[k]['rlp_labels'] = rlp_labels self.gt_labels[k]['sub_boxes'] = sub_boxes self.gt_labels[k]['obj_boxes'] = obj_boxes print 'done preloading gt %f' % zl.tock() zl.save('output/cache/vg1_2_2016_gt.pkl', self.gt_labels) self.imids = [] for k in self.vgg_data.keys(): self.imids.append(k) self.imidx = 0 random.shuffle(self.imids) layer_params = yaml.load(self.param_str_) self._batch_size = layer_params['batch_size'] self.train_data = [] self._name_to_top_map = {} # top[0].reshape(self._batch_size, 201*2) top[0].reshape(self._batch_size, 4096 * 2) # top[0].reshape(self._batch_size, 4*2) top[1].reshape(self._batch_size) # self._name_to_top_map['classeme'] = 0 self._name_to_top_map['visual'] = 0 # self._name_to_top_map['location'] = 0 self._name_to_top_map['label'] = 1
def run_retrieval_zeroshot(): # h5_path = 'output/sg_vrd_2016_result_all_19500.hdf5' # h5_path = 'output/sg_vrd_2016_result_diff_all_5000.hdf5' # h5_path= 'output/results/lu_method_results.hdf5' # h5_path = 'output/sg_vrd_2016_result.hdf5.dd' # h5_path = 'output/results/lu_method_results_max.hdf5' h5_path = 'output/results/lu_visual_method_results.hdf5' data_root = 'data/sg_vrd_2016/Data/sg_test_images/' m = h5py.File('/home/zawlin/Dropbox/proj/sg_vrd_meta.h5', 'r', 'core') zeroshots = m['meta/zeroshots'][...] gt_cache_path = 'output/cache/sg_vrd_gt_cache.pkl' gt_h5f = {} np.random.seed(76) if os.path.exists(gt_cache_path): print 'load gt from cache' gt_h5f = zl.load(gt_cache_path) else: print 'cacheing gt' for k in m['gt/test']: gt_h5f[k] = {} sub_boxes = m['gt/test'][k]['sub_boxes'][...] rlp_labels = m['gt/test'][k]['rlp_labels'][...] obj_boxes = m['gt/test'][k]['obj_boxes'][...] gt_h5f[k]['sub_boxes'] = sub_boxes gt_h5f[k]['obj_boxes'] = obj_boxes gt_h5f[k]['rlp_labels'] = rlp_labels print 'caching gt done' zl.save(gt_cache_path, gt_h5f) cache_path = 'output/cache/%s.pkl' % h5_path.split('/')[-1] # h5f = h5py.File('output/sg_vrd_2016_result.classeme.hdf5') if os.path.exists(cache_path): print 'load from cache' h5f = zl.load(cache_path) else: h5_in = h5py.File(h5_path, 'r') h5f = {} print 'preloading data' for i in h5_in: h5f[i] = {} h5f[i]['rlp_labels'] = h5_in[i]['rlp_labels'][...] h5f[i]['rlp_confs'] = h5_in[i]['rlp_confs'][...] h5f[i]['sub_boxes'] = h5_in[i]['sub_boxes'][...] h5f[i]['obj_boxes'] = h5_in[i]['obj_boxes'][...] zl.save(cache_path, h5f) print 'preloading data done' #h5f = h5py.file('output/results/lu_method_results.hdf5') retr_meta = zl.load('output/pkl/vr_retr_meta.pkl') thresh = 0.0 images = {} imids = h5f.keys() results = {} cnt = 0 r_acc_100 = 0 r_acc_50 = 0 tp_total = 0 gt_total = 0 median = [] for k in xrange(len(retr_meta['rlp_labels'])): # if k>1000: # break rlp_label = retr_meta['rlp_labels'][k] if not np.any(np.all(zeroshots == rlp_label, axis=1)): continue gt_files = retr_meta['files'][k] cnt += 1 results = {} zl.tic() for imid in imids: rlp_labels = h5f[imid]['rlp_labels'] rlp_confs = h5f[imid]['rlp_confs'] if rlp_confs.shape[0] == 0: results[imid] = 0.0 continue indexor = np.arange(rlp_labels.shape[0]) ind = indexor[np.all(rlp_labels == rlp_label, axis=1)] if ind.shape[0] == 0: results[imid] = 0.0 continue confs = rlp_confs[ind] results[imid] = np.random.uniform(-1, 1) #np.average(confs) results_sorted = zl.sort_dict_by_val(results) total_gt = len(gt_files) + 0.0 gt_total += total_gt + 0.0 tp_50 = 0. tp_100 = 0. found = False delay = 0 s_lbl = zl.idx2name_cls(m, rlp_label[0]) p_lbl = zl.idx2name_pre(m, rlp_label[1]) o_lbl = zl.idx2name_cls(m, rlp_label[2]) lbl_str = '%s_%s_%s' % (s_lbl, p_lbl, o_lbl) r_at_k = 5 for i in xrange(len(results_sorted)): imid, v = results_sorted[i] if found and i >= r_at_k: break cor = imid in gt_files if cor: if not found: found = True median.append(i) if i < r_at_k: tp_100 += 1 tp_total += 1 if i < 50: tp_50 += 1 # if True: # cor_or_not = str(cor) # if cor :delay=0 # if delay ==0: # im = cv2.imread(data_root+imid+'.jpg') # cv2.putText(im, cor_or_not, (50, 50), cv2.FONT_HERSHEY_COMPLEX, 1, (0, 0, 255), 1) # cv2.putText(im, lbl_str, (50, 80), cv2.FONT_HERSHEY_COMPLEX, 1, (0, 0, 255), 1) # cv2.putText(im, str(i), (50, 100), cv2.FONT_HERSHEY_COMPLEX, 1, (0, 0, 255), 1) # cv2.imshow('im',im) # c = cv2.waitKey(delay)&0xFF # if c ==27: # exit(0) # if c == ord('s'): # delay = 1-delay # if c == ord('c'): # delay = 1 r_50 = tp_50 / r_at_k #total_gt r_100 = tp_100 / r_at_k #total_gt r_acc_50 += r_50 r_acc_100 += r_100 med = np.median(median) print '%d %f %f %f %f %d %f' % (cnt, r_50, r_100, r_acc_50 / cnt, r_acc_100 / cnt, med, zl.toc())
def run_retrieval_vp_v2(): # h5_path = 'output/sg_vrd_2016_result.classeme.hdf5' h5_path = 'output/precalc/vg1_2_vp2016_test_nms2_.4.hdf5' m = h5py.File('/home/zawlin/Dropbox/proj/vg1_2_meta.h5', 'r', 'core') m_vp = h5py.File('/home/zawlin/Dropbox/proj/vg1_2_vp_meta.h5', 'r', 'core') cache_path = 'output/cache/%s.pkl' % h5_path.split('/')[-1] data_root = '/home/zawlin/data/data_vrd/vg_1.2/' rlp_map = [] for i in xrange(1, len(m_vp['meta/tri/idx2name'].keys())): tri = str(m_vp['meta/tri/idx2name'][str(i)][...]) s_lbl = tri.split('_')[0] p_lbl = tri.split('_')[1] o_lbl = tri.split('_')[2] rlp_label = [ zl.name2idx_cls(m, s_lbl), zl.name2idx_pre(m, p_lbl), zl.name2idx_cls(m, o_lbl) ] rlp_map.append(rlp_label) rlp_map = np.array(rlp_map) if os.path.exists(cache_path): print 'load from cache' h5f = zl.load(cache_path) else: h5_in = h5py.File(h5_path, 'r') h5f = {} print 'preloading data' for i in h5_in: h5f[i] = {} h5f[i]['labels'] = h5_in[i]['labels'][...] h5f[i]['confs'] = h5_in[i]['confs'][...] h5f[i]['boxes'] = h5_in[i]['boxes'][...] zl.save(cache_path, h5f) print 'preloading data done' retr_meta = zl.load('output/pkl/vg_retr_meta.pkl') thresh = 0.0 images = {} imids = h5f.keys() results = {} cnt = 0 r_acc_100 = 0 r_acc_50 = 0 tp_total = 0 gt_total = 0 median = [] total = 0 retr_meta = zl.load('output/pkl/vg_retr_meta.pkl') for k in xrange(len(retr_meta['rlp_labels'])): total += retr_meta['counts'][k] if k > 1000: print total / 1000. exit(0) break continue cnt += 1 rlp_label = retr_meta['rlp_labels'][k] gt_files = retr_meta['files'][k] # print gt_files # exit(0) # for f in gt_files: # impath= zl.imid2path(m,f) # print impath # im= cv2.imread(data_root+impath) # cv2.imshow('im',im) # cv2.waitKey(0) results = {} zl.tick() ranks = [] for imid in imids: labels = h5f[imid]['labels'] - 1 rlp_confs = h5f[imid]['confs'] rlp_labels = rlp_map[labels] if rlp_labels.shape[0] == 0: results[imid] = 0.0 continue indexor = np.arange(rlp_labels.shape[0]) ind = indexor[np.all(rlp_labels == rlp_label, axis=1)] if ind.shape[0] == 0: results[imid] = 0.0 continue confs = rlp_confs[ind] results[imid] = np.average(confs) results_sorted = zl.sort_dict_by_val(results) total_gt = len(gt_files) + 0.0 gt_total += total_gt + 0.0 tp_50 = 0. tp_100 = 0. found = False s_lbl = zl.idx2name_cls(m, rlp_label[0]) p_lbl = zl.idx2name_pre(m, rlp_label[1]) o_lbl = zl.idx2name_cls(m, rlp_label[2]) lbl_str = '%s_%s_%s' % (s_lbl, p_lbl, o_lbl) delay = 0 for i in xrange(len(results_sorted)): imid, v = results_sorted[i] impath = zl.imid2path(m, imid) if found and i >= 5: break # print gt_files cor = imid in gt_files if cor: if not found: found = True median.append(i) if i < 5: tp_100 += 1 tp_total += 1 if i < 50: tp_50 += 1 # if True: # cor_or_not = str(cor) # if cor :delay=0 # if delay ==0: # im = cv2.imread(data_root+impath) # cv2.putText(im, cor_or_not, (50, 50), cv2.FONT_HERSHEY_COMPLEX, 1, (0, 0, 255), 1) # cv2.putText(im, lbl_str, (50, 80), cv2.FONT_HERSHEY_COMPLEX, 1, (0, 0, 255), 1) # cv2.putText(im, str(i), (50, 100), cv2.FONT_HERSHEY_COMPLEX, 1, (0, 0, 255), 1) # cv2.imshow('im',im) # c = cv2.waitKey(delay)&0xFF # if c ==27: # exit(0) # if c == ord('s'): # delay = 1-delay # if c == ord('c'): # delay = 1 r_50 = tp_50 / 5 #total_gt r_100 = tp_100 / 5 #total_gt r_acc_50 += r_50 r_acc_100 += r_100 med = np.median(median) print '%d %f %f %f %f %d %f' % (cnt, r_50, r_100, r_acc_50 / cnt, r_acc_100 / cnt, med, zl.tock())
def run_relation(model_type, iteration): cache_h5 = h5py.File('output/sg_vrd_cache.h5')['test'] result = h5py.File('output/sg_vrd_2016_result_' + model_type + '_' + iteration + '.hdf5') m = h5py.File('data/sg_vrd_meta.h5') #--------------cache boxes----------------- h5_boxes = h5py.File('output/precalc/sg_vrd_objs.hdf5') cache_boxes = {} if os.path.exists('output/cache/sg_vrd_objs_test.pkl'): cache_boxes = zl.load('output/cache/sg_vrd_objs_test.pkl') glog.info('loaded obj data from cache') else: glog.info('Preloading obj') zl.tic() for k in h5_boxes['test'].keys(): boxes = h5_boxes['test/%s/boxes' % k][...] cache_boxes[k] = boxes glog.info('done preloading obj %f' % zl.toc()) zl.save('output/cache/sg_vrd_obj_test.pkl', cache_boxes) #--------------cache boxes----------------- #--------------cache old boxes------------- h5_boxes = h5py.File('output/sg_vrd_2016_test.hdf5') cache_old_boxes = {} if os.path.exists('output/cache/sg_vrd_objs_test_vgg.pkl'): cache_old_boxes = zl.load('output/cache/sg_vrd_objs_test_vgg.pkl') glog.info('loaded obj data from cache') else: glog.info('Preloading obj') zl.tic() for k in h5_boxes.keys(): locations = h5_boxes['%s/locations' % k][...] cls_confs = h5_boxes['%s/cls_confs' % k][...] boxes = np.concatenate((locations, cls_confs[:, 1, np.newaxis], cls_confs[:, 0, np.newaxis]), axis=1) cache_old_boxes[k] = boxes glog.info('done preloading obj %f' % zl.toc()) zl.save('output/cache/sg_vrd_obj_test_vgg.pkl', cache_old_boxes) #--------------cache old boxes------------- data_root = 'data/sg_vrd_2016/Data/sg_test_images/' keep = 100 thresh = 0.0001 prototxt = 'models/sg_vrd/rel_pre_iccv/test_' + model_type + '.prototxt' model_path = 'output/rel_iccv/' + model_type + '_iter_' + iteration + '.caffemodel' net = caffe.Net(prototxt, model_path, caffe.TEST) # sio.savemat('output/'+model_type+'.mat',{'params_'+model_type:net.params['relation'][0].data}) # exit(0) #net = caffe.Net('models/sg_vrd/relation/test.prototxt','output/models/sg_vrd_relation_vgg16_iter_264000.caffemodel',caffe.TEST) cnt = 0 for imid in cache_h5.keys(): cnt += 1 if cnt % 10 == 0: glog.info(cnt) obj_boxes_gt = m['gt/test'][imid]['obj_boxes'][...] sub_boxes_gt = m['gt/test'][imid]['sub_boxes'][...] rlp_labels_gt = m['gt/test'][imid]['rlp_labels'][...] rlp_labels = [] rlp_confs = [] sub_boxes = [] obj_boxes = [] #boxes = cache_boxes[imid] boxes = cache_old_boxes[imid] if boxes.shape[0] >= 2: for s in xrange(boxes[:20].shape[0]): for o in xrange(boxes[:20].shape[0]): if s == o: continue if boxes[s][4] < 0.01: continue if boxes[o][4] < 0.01: continue sbox = boxes[s][:4] obox = boxes[o][:4] rlp_labels.append([boxes[s, 5], -1, boxes[o, 5]]) rlp_confs.append(boxes[s, 4] + boxes[o, 4]) sub_boxes.append(sbox) obj_boxes.append(obox) if len(sub_boxes) <= 0: continue #sub_box = np.array(sub_box) #obj_box = np.array(obj_box) im_path = C.get_sg_vrd_path_test(imid) im_detect(net, im_path, sub_boxes, obj_boxes) relation_prob = net.blobs['relation_prob'].data[...] for r in xrange(relation_prob.shape[0]): argmax = np.argmax(relation_prob[r, ...]) rs = relation_prob[r, argmax].squeeze() rlp_labels[r][1] = argmax rlp_confs[r] = rlp_confs[r] + rs im = cv2.imread(data_root + imid + '.jpg') for i in xrange(len(sub_boxes)): sb = sub_boxes[i] cv2.rectangle(im, (int(sb[0]), int(sb[1])), (int(sb[2]), int(sb[3])), (255, 0, 0), 2) pass cv2.imshow('im', im) #if cv2.waitKey(0)==27: # exit(0) #rlp_confs.append(rlp_conf) #rlp_labels.append(rlp_label) #sub_boxes.append(sub_box) #obj_boxes.append(obj_box) # for i in xrange(70): # rs = relation_score[i] # if rs>0.0: # predicate =i # #print relation_score[predicate] # rlp_label = np.array([sub_cls[s],predicate,obj_cls[s]]).astype(np.int32) # #print '%s %s %s %f'%(m['meta/cls/idx2name/'+str(rlp_label[0])][...],m['meta/pre/idx2name/'+str(rlp_label[1])][...],m['meta/cls/idx2name/'+str(rlp_label[2])][...],rs) # rlp_conf = rs#+sub_score+obj_score#relation_score[predicate] # rlp_confs.append(rlp_conf) # rlp_labels.append(rlp_label) # sub_boxes.append(sub_boxes_gt[s]) # obj_boxes.append(obj_boxes_gt[s]) result.create_dataset(imid + '/rlp_confs', dtype='float16', data=np.array(rlp_confs).astype(np.float16)) result.create_dataset(imid + '/sub_boxes', dtype='float16', data=np.array(sub_boxes).astype(np.float16)) result.create_dataset(imid + '/obj_boxes', dtype='float16', data=np.array(obj_boxes).astype(np.float16)) result.create_dataset(imid + '/rlp_labels', dtype='float16', data=np.array(rlp_labels).astype(np.float16))
def setup(self, bottom, top): self._cur_idx = 0 self.rdata = sio.loadmat('data/meta/vrd/annotation_train.mat', struct_as_record=False, squeeze_me=True) self.meta = h5py.File('data/sg_vrd_meta.h5', 'r') self.gt_labels = {} if os.path.exists('output/cache/sg_vrd_gt.pkl'): self.gt_labels = zl.load('output/cache/sg_vrd_gt.pkl') glog.info('loaded gt data from cache') else: glog.info('Preloading gt') zl.tic() for k in self.meta['gt/train'].keys(): rlp_labels = self.meta['gt/train/%s/rlp_labels' % k][...] sub_boxes = self.meta['gt/train/%s/sub_boxes' % k][...].astype( np.float) obj_boxes = self.meta['gt/train/%s/obj_boxes' % k][...].astype( np.float) if sub_boxes.shape[0] > 0: zeros = np.zeros((sub_boxes.shape[0], 1), dtype=np.float) # first index is always zero since we do one image by one image sub_boxes = np.concatenate((zeros, sub_boxes), axis=1) obj_boxes = np.concatenate((zeros, obj_boxes), axis=1) self.gt_labels[k] = {} self.gt_labels[k]['rlp_labels'] = rlp_labels self.gt_labels[k]['sub_boxes'] = sub_boxes self.gt_labels[k]['obj_boxes'] = obj_boxes glog.info('done preloading gt %f' % zl.toc()) vgg_h5 = h5py.File("output/sg_vrd_2016_train_predicate_exp_train.hdf5", 'r', 'core') if os.path.exists('output/cache/sg_vrd_2016_train.pkl'): self.vgg_data = zl.load('output/cache/sg_vrd_2016_train.pkl') glog.info('loaded train data from cache') else: glog.info('Preloading training data') zl.tic() self.vgg_data = {} for k in vgg_h5.keys(): sub_visual = vgg_h5[k]['sub_classeme'][...] obj_visual = vgg_h5[k]['obj_classeme'][...] pre_label = vgg_h5[k]['pre_label'][...] self.vgg_data[k] = {} self.vgg_data[k]['sub_visual'] = sub_visual self.vgg_data[k]['obj_visual'] = obj_visual self.vgg_data[k]['pre_label'] = pre_label glog.info('done preloading training data %f' % zl.toc()) zl.save('output/cache/sg_vrd_2016_train.pkl', self.vgg_data) vgg_h5.close() self.meta = h5py.File('data/sg_vrd_meta.h5', 'r', 'core') layer_params = yaml.load(self.param_str) self._batch_size = layer_params['batch_size'] self.train_data = [] self._name_to_top_map = {} # data blob: holds a batch of N images, each with 3 channels # top[0].reshape(self._batch_size, 4096 * 2 ) top[0].reshape(self._batch_size, 2 * 101) top[1].reshape(self._batch_size) # self._name_to_top_map['visual'] = 0 # self._name_to_top_map['classeme'] = 0 self._name_to_top_map['visual'] = 0 # self._name_to_top_map['location'] = 1 self._name_to_top_map['label'] = 1
def run_relation_batch(model_type, iteration): vgg_h5 = h5py.File('output/precalc/vg1_2_2016_test.hdf5') vgg_data = {} if os.path.exists('output/cache/vg1_2_2016_test.pkl'): vgg_data = zl.load('output/cache/vg1_2_2016_test.pkl') print 'loaded test data from cache' else: print 'Preloading testing data' zl.tick() for k in vgg_h5.keys(): classemes = vgg_h5[k]['classemes'][...] visuals = vgg_h5[k]['visuals'][...] locations = vgg_h5[k]['locations'][...] cls_confs = vgg_h5[k]['cls_confs'][...] vgg_data[k] = {} vgg_data[k]['classemes'] = classemes vgg_data[k]['visuals'] = visuals vgg_data[k]['cls_confs'] = cls_confs vgg_data[k]['locations'] = locations print 'done preloading testing data %f' % zl.tock() zl.save('output/cache/vg1_2_2016_test.pkl', vgg_data) vgg_h5.close() result = h5py.File('output/vg_results/vg1_2_2016_result_' + model_type + '_' + iteration + '.hdf5') m = h5py.File('data/vg1_2_meta.h5') data_root = 'data/vg1_2_2016/Data/test/' keep = 100 thresh = 0.0001 net = caffe.Net( 'models/vg1_2/relation/test_' + model_type + '.prototxt', 'output/relation/vg/relation_vgg16_' + model_type + '_iter_' + iteration + '.caffemodel', caffe.TEST) #net = caffe.Net('models/sg_vrd/relation/test.prototxt','output/models/sg_vrd_relation_vgg16_iter_264000.caffemodel',caffe.TEST) cnt = 1 zl.tick() imids = sorted(vgg_data.keys()) for imid in imids: if cnt % 100 == 0: print cnt, zl.tock() zl.tick() cnt += 1 if imid in result: continue classemes = vgg_data[imid]['classemes'] visuals = vgg_data[imid]['visuals'] locations = vgg_data[imid]['locations'] cls_confs = vgg_data[imid]['cls_confs'] # im = cv2.imread(data_root+imid+'.jpg') # #print cls_confs # # for box in locations: # # b=box[:4].astype(np.int32) # # cv2.rectangle(im,(b[0],b[1]),(b[2],b[3]),(255,0,0)) # w,h = im.shape[2],im.shape[1] rlp_labels = [] rlp_confs = [] sub_boxes = [] obj_boxes = [] relation_vectors = [] classemes_in = [] visuals_in = [] locations_in = [] cls_confs_in = [] sub_cls_in = [] obj_cls_in = [] sub_score_in = [] obj_score_in = [] sub_boxes = [] obj_boxes = [] for s in xrange(len(locations)): for o in xrange(len(locations)): if s == o: continue sub = locations[s] obj = locations[o] sub_visual = visuals[s] obj_visual = visuals[o] sub_cls = cls_confs[s, 0] obj_cls = cls_confs[o, 0] sub_score = cls_confs[s, 1] obj_score = cls_confs[o, 1] sub_classme = classemes[s] obj_classme = classemes[o] sub_loc_encoded = bbox_transform(np.array([obj[:4]]), np.array([sub[:4]]))[0] obj_loc_encoded = bbox_transform(np.array([sub[:4]]), np.array([obj[:4]]))[0] visual = np.hstack((sub_visual, obj_visual)).reshape(8192) classeme = np.hstack((sub_classme, obj_classme)).reshape(402) loc = np.hstack((sub_loc_encoded, obj_loc_encoded)).reshape(8) classemes_in.append(classeme) visuals_in.append(visual) locations_in.append(loc) sub_cls_in.append(sub_cls) obj_cls_in.append(obj_cls) sub_score_in.append(sub_score) obj_score_in.append(obj_score) sub_boxes.append(sub[:4]) obj_boxes.append(obj[:4]) if 'all' in model_type: blob = { 'classeme': np.array(classemes_in), 'visual': np.array(visuals_in), 'location': np.array(locations_in) } net.blobs['classeme'].reshape(*blob['classeme'].shape) net.blobs['visual'].reshape(*blob['visual'].shape) net.blobs['location'].reshape(*blob['location'].shape) elif 'visual' in model_type: blob = { 'visual': np.array(visuals_in), } net.blobs['visual'].reshape(*blob['visual'].shape) elif 'classeme' in model_type: blob = { 'classeme': np.array(classemes_in), } net.blobs['classeme'].reshape(*blob['classeme'].shape) elif 'location' in model_type: blob = {'location': np.array(locations_in)} #batch this net.blobs['location'].reshape(*blob['location'].shape) if len(locations_in) == 0: rlp_confs = [] sub_boxes = [] obj_boxes = [] rlp_labels = [] else: net.forward_all(**blob) relation_score = net.blobs['relation_prob'].data.copy() argmax = np.argmax(relation_score, axis=1) rs = relation_score[np.arange(relation_score.shape[0]), argmax] rlp_labels = np.vstack((sub_cls_in, argmax, obj_cls_in)).T rlp_confs = np.array(sub_score_in) + np.array(rs) + np.array( obj_score_in) result.create_dataset(imid + '/rlp_confs', dtype='float16', data=np.array(rlp_confs).astype(np.float16)) result.create_dataset(imid + '/sub_boxes', dtype='float16', data=np.array(sub_boxes).astype(np.float16)) result.create_dataset(imid + '/obj_boxes', dtype='float16', data=np.array(obj_boxes).astype(np.float16)) result.create_dataset(imid + '/rlp_labels', dtype='float16', data=np.array(rlp_labels).astype(np.float16))
def gen_meta_for_retrieval(): out_pkl = 'output/pkl/vr_retr_meta.pkl' m = h5py.File('/home/zawlin/Dropbox/proj/sg_vrd_meta.h5', 'r') data_root = 'data/sg_vrd_2016/Data/sg_test_images/' rlp_labels = [] files = [] counts = [] sub_boxes = [] obj_boxes = [] for k in m['gt/test']: gt_rlp_labels = m['gt/test'][k]['rlp_labels'][...] gt_sub_boxes = m['gt/test'][k]['sub_boxes'][...] gt_obj_boxes = m['gt/test'][k]['obj_boxes'][...] for i in xrange(gt_rlp_labels.shape[0]): gt_rlp_label = gt_rlp_labels[i] gt_sub_box = gt_sub_boxes[i] gt_obj_box = gt_obj_boxes[i] if len(rlp_labels) == 0: rlp_labels.append(gt_rlp_label.tolist()) files.append([k]) sub_boxes.append([gt_sub_box.tolist()]) obj_boxes.append([gt_obj_box.tolist()]) counts.append(1) continue bInd = np.all(gt_rlp_label == rlp_labels, axis=1) ind = np.arange(len(rlp_labels))[bInd] if len(ind) == 0: rlp_labels.append(gt_rlp_label.tolist()) files.append([k]) counts.append(1) sub_boxes.append([gt_sub_box.tolist()]) obj_boxes.append([gt_obj_box.tolist()]) else: files[ind].append(k) counts[ind] = counts[ind] + 1 sub_boxes[ind].append(gt_sub_box.tolist()) obj_boxes[ind].append(gt_obj_box.tolist()) # rlp_labels.append(gt_rlp_label.tolist()) # files.append([k]) # counts.append(1) rlp_labels = np.array(rlp_labels) files = np.array(files) counts = np.array(counts) sub_boxes = np.array(sub_boxes) obj_boxes = np.array(obj_boxes) ind = np.argsort(counts)[::-1] counts = counts[ind] files = files[ind] rlp_labels = rlp_labels[ind] sub_boxes = sub_boxes[ind] obj_boxes = obj_boxes[ind] print sub_boxes[:4] for i in xrange(20): rlp_label = rlp_labels[i] print files[i] s_lbl = zl.idx2name_cls(m, rlp_label[0]) p_lbl = zl.idx2name_pre(m, rlp_label[1]) o_lbl = zl.idx2name_cls(m, rlp_label[2]) print s_lbl, p_lbl, o_lbl for j in xrange(len(files[i])): s_box = sub_boxes[i][j] o_box = obj_boxes[i][j] fpath = files[i][j] impath = data_root + fpath + '.jpg' im = cv2.imread(impath) cv2.rectangle(im, (s_box[0], s_box[1]), (s_box[2], s_box[3]), (0, 255, 0), 1) cv2.rectangle(im, (o_box[0], o_box[1]), (o_box[2], o_box[3]), (255, 0, 0), 1) cv2.imshow('im', im) cv2.waitKey(0) retr_meta = { 'counts': counts, 'files': files, 'rlp_labels': rlp_labels, 'sub_boxes': sub_boxes, 'obj_boxes': obj_boxes } zl.save(out_pkl, retr_meta)