def get_feat(dataset, is_train=False):
ret = []
for i in xrange(1501):
cur = []
for j in xrange(1, 7):
images = glob.glob('%s/recs/%s/id_%d_%d*' %
(args.data_dir, dataset, i, j))
if len(images) == 0:
continue
cam = []
for k in images:
bs, flst = 0, open(k)
for line in flst:
bs += 1
org_iter = get_imRecordIter(args,
k[len(args.data_dir) + 1:-4],
(3, 224, 112),
1,
shuffle=is_train,
aug=is_train,
even_iter=True)
cam.append(org_iter)
if len(cam) > 0:
cur.append(cam)
if len(cur) > 0:
ret.append(cur)
return ret
'%s/%s_%d' % (model_path, args.base_model_load_prefix, sets),
args.base_model_load_epoch)
base_mod = mx.mod.Module(symbol=build_base_net(args),
data_names=('data', ),
label_names=None,
context=devices)
base_mod.bind(data_shapes=[('data', (args.batch_size, 3, 224, 112))],
for_training=False)
base_mod.init_params(initializer=None,
arg_params=arg_params,
aux_params=aux_params,
force_init=True)
dataiter = get_imRecordIter(args,
'%s%d' % (args.dataset, sets), (3, 224, 112),
args.batch_size,
shuffle=False,
aug=False,
even_iter=False)
dataiter.reset()
output = base_mod.predict(dataiter)
F = output
F2 = F
print(F.shape)
cnt_lst = np.loadtxt(args.data_dir + '/' + 'image_test' + str(sets) +
'.txt').astype(int)
N = cnt_lst.shape[0] / 2
avp = []
arg_params, aux_params = load_checkpoint(
'../baseline/%s/%s' % (model_path, args.base_model_load_prefix),
args.base_model_load_epoch)
base_mod = mx.mod.Module(symbol=sym_base_net(args.network, is_test=True),
data_names=('data', ),
label_names=None,
context=devices)
base_mod.bind(data_shapes=[('data', (1024, 3, 224, 112))], for_training=False)
base_mod.init_params(initializer=None,
arg_params=arg_params,
aux_params=aux_params,
force_init=True)
dataiter = get_imRecordIter(args,
'recs/eval_test', (3, 224, 112),
1024,
shuffle=False,
aug=False,
even_iter=True)
dataiter.reset()
F = base_mod.predict(dataiter)
del dataiter
print 'l2'
print F
print 'base feat predicted'
query = np.loadtxt('/data3/matt/MARS/MARS-evaluation/info/query.csv',
delimiter=',').astype(int)
gallery = np.loadtxt('/data3/matt/MARS/MARS-evaluation/info/gallery.csv',
delimiter=',').astype(int)
cnts, cnt = [0, 0, 0, 0], [0]
batch_size = args.batch_size
num_epoch = args.num_epoches
arg_params, aux_params = load_checkpoint('models/%s' % test_args.model_load_prefix, test_args.model_load_epoch)
data1, data2 = sym_base_net(args.network, is_train=args.e2e, global_stats=True)
Q = create_moduleQ(data1, data2, devices, args.sample_size, args.num_sim, args.num_hidden, args.num_acts, args.min_states, args.min_imgs, fusion=args.fusion, is_train=True, nh=not args.history, is_e2e=args.e2e, bn=args.q_bn)
Q.init_params(initializer=None,
arg_params=arg_params,
aux_params=aux_params,
allow_missing=False,
force_init=True)
valid_iter = get_imRecordIter(
args, 'recs/%s'%args.valid_set, (3, 224, 112), 1,
shuffle=False, aug=False, even_iter=True)
train_iter = get_imRecordIter(
args, 'recs/%s'%args.train_set, (3, 224, 112), 1,
shuffle=False, aug=True, even_iter=True)
valid_lst = np.loadtxt('%s/recs/%s.txt'%(args.data_dir, args.valid_set)).astype(int)
train_lst = np.loadtxt('%s/recs/%s.txt'%(args.data_dir, args.train_set)).astype(int)
valid = BatchProvider(valid_iter, valid_lst, False, args.sample_size, sample_ratio=0.5, is_valid=True, need_feat=args.history)
train = BatchProvider(train_iter, train_lst, True, args.sample_size, sample_ratio=0.5, need_feat=args.history)
N = args.num_id
cmcs, ap, cmcn, vscores, vturns = [[], [], [], []], [], [1, 5, 10, 20], [], []
max_penalty=1