def test(test_loader, target_net):
box_num = 0
correct_cnt, total_cnt = 0., 0.
print '========== Testing ======='
results = []
batch_time = network.AverageMeter()
end = time.time()
im_counter = 0
for i, sample in enumerate(test_loader):
correct_cnt_t, total_cnt_t = 0., 0.
# Forward pass
im_data = Variable(sample['visual'].cuda(), volatile=True)
im_counter += im_data.size(0)
im_info = sample['image_info']
gt_objects = sample['objects']
object_rois = target_net(im_data, im_info)[1]
results.append(object_rois.cpu().data.numpy())
box_num += object_rois.size(0)
correct_cnt_t, total_cnt_t = check_recall(object_rois, gt_objects, 200)
correct_cnt += correct_cnt_t
total_cnt += total_cnt_t
batch_time.update(time.time() - end)
end = time.time()
if (i + 1) % 100 == 0 and i > 0:
print('[{0}/{6}] Time: {1:2.3f}s/img).'
'\t[object] Avg: {2:2.2f} Boxes/im, Top-200 recall: {3:2.3f} ({4:.0f}/{5:.0f})'.format(
i + 1, batch_time.avg,
box_num / float(im_counter), correct_cnt / float(total_cnt)* 100, correct_cnt, total_cnt,
len(test_loader)))
recall = correct_cnt / float(total_cnt)
print '====== Done Testing ===='
return recall, results
def test(loader, model, top_Ns, nms=-1., triplet_nms=-1., use_gt_boxes=False):
print '========== Testing ======='
model.eval()
rel_cnt = 0.
rel_cnt_correct = np.zeros(2)
phrase_cnt_correct = np.zeros(2)
pred_cnt_correct = np.zeros(2)
total_region_rois_num = 0
max_region_rois_num = 0
result = []
batch_time = network.AverageMeter()
end = time.time()
for i, sample in enumerate(loader): # (im_data, im_info, gt_objects, gt_relationships)
assert len(sample['visual']) == 1
input_visual = sample['visual'][0].cuda()
gt_objects = sample['objects']
gt_relationships = sample['relations']
image_info = sample['image_info']
# Forward pass
total_cnt_t, cnt_correct_t, eval_result_t = model.module.evaluate(
input_visual, image_info, gt_objects, gt_relationships,
top_Ns = top_Ns, nms=nms, triplet_nms=triplet_nms,
use_gt_boxes=use_gt_boxes)
eval_result_t['path'] = sample['path'][0] # for visualization
rel_cnt += total_cnt_t
result.append(eval_result_t)
rel_cnt_correct += cnt_correct_t[0]
phrase_cnt_correct += cnt_correct_t[1]
pred_cnt_correct += cnt_correct_t[2]
total_region_rois_num += cnt_correct_t[3]
max_region_rois_num = cnt_correct_t[3] if cnt_correct_t[3] > max_region_rois_num else max_region_rois_num
batch_time.update(time.time() - end)
end = time.time()
if (i + 1) % 500 == 0 and i > 0:
print('[Evaluation][%d/%d][%.2fs/img][avg: %d subgraphs, max: %d subgraphs]' %(i+1, len(loader), batch_time.avg, total_region_rois_num / float(i+1), max_region_rois_num))
for idx, top_N in enumerate(top_Ns):
print('\tTop-%d Recall:\t[Pred] %2.3f%%\t[Phr] %2.3f%%\t[Rel] %2.3f%%' % (
top_N, pred_cnt_correct[idx] / float(rel_cnt) * 100,
phrase_cnt_correct[idx] / float(rel_cnt) * 100,
rel_cnt_correct[idx] / float(rel_cnt) * 100))
recall = [rel_cnt_correct / float(rel_cnt),
phrase_cnt_correct / float(rel_cnt),
pred_cnt_correct / float(rel_cnt)]
print('\n====== Done Testing ====')
return recall, result
def test(test_loader, target_net):
box_num = np.array([0, 0])
correct_cnt, total_cnt = np.array([0, 0]), np.array([0, 0])
print '========== Testing ======='
results_obj = []
results_region = []
batch_time = network.AverageMeter()
end = time.time()
im_counter = 0
for i, sample in enumerate(test_loader):
correct_cnt_t, total_cnt_t = np.array([0, 0]), np.array([0, 0])
# Forward pass
# measure the data loading time
im_data = Variable(sample['visual'].cuda(), volatile=True)
im_counter += im_data.size(0)
im_info = sample['image_info']
gt_objects = sample['objects']
gt_regions = sample['regions']
object_rois, region_rois = target_net(im_data, im_info)[1:]
results_obj.append(object_rois.cpu().data.numpy())
results_region.append(region_rois.cpu().data.numpy())
box_num[0] += object_rois.size(0)
box_num[1] += region_rois.size(0)
correct_cnt_t[0], total_cnt_t[0] = check_recall(
object_rois, gt_objects, 50)
correct_cnt_t[1], total_cnt_t[1] = check_recall(
region_rois, gt_regions, 50)
correct_cnt += correct_cnt_t
total_cnt += total_cnt_t
batch_time.update(time.time() - end)
end = time.time()
if (i + 1) % 100 == 0 and i > 0:
print(
'[{0}/{10}] Time: {1:2.3f}s/img).'
'\t[object] Avg: {2:2.2f} Boxes/im, Top-50 recall: {3:2.3f} ({4:d}/{5:d})'
'\t[region] Avg: {6:2.2f} Boxes/im, Top-50 recall: {7:2.3f} ({8:d}/{9:d})'
.format(i + 1, batch_time.avg, box_num[0] / float(im_counter),
correct_cnt[0] / float(total_cnt[0]) * 100,
correct_cnt[0], total_cnt[0],
box_num[1] / float(im_counter),
correct_cnt[1] / float(total_cnt[1]) * 100,
correct_cnt[1], total_cnt[1], len(test_loader)))
recall = correct_cnt / total_cnt.astype(np.float)
print '====== Done Testing ===='
return recall, results_obj, results_region
default=-1.,
help='NMS threshold for post object NMS (negative means not NMS)')
parser.add_argument(
'--triplet_nms',
type=float,
default=0.4,
help='Triplet NMS threshold for post object NMS (negative means not NMS)')
# parser.add_argument('--rerank', action='store_true', help='Whether to rerankt the object score')
# testing settings
parser.add_argument('--use_gt_boxes',
action='store_true',
help='Use ground truth bounding boxes for evaluation')
args = parser.parse_args()
# Overall loss logger
overall_train_loss = network.AverageMeter()
overall_train_rpn_loss = network.AverageMeter()
overall_gradients_norm_logger = network.LoggerMeter()
is_best = False
best_recall = [0., 0.]
best_recall_phrase = [0., 0.]
best_recall_pred = [0., 0.]
def main():
global args, is_best, best_recall, best_recall_pred, best_recall_phrase
# To set the model name automatically
# Set options
options = {
def train(train_loader, target_net, optimizer, epoch):
batch_time = network.AverageMeter()
data_time = network.AverageMeter()
train_loss = network.AverageMeter()
train_loss_obj_box = network.AverageMeter()
train_loss_obj_entropy = network.AverageMeter()
accuracy_obj = network.AccuracyMeter()
target_net.train()
end = time.time()
for i, sample in enumerate(train_loader):
# measure the data loading time
data_time.update(time.time() - end)
im_data = sample['visual'][0].cuda()
im_info = sample['image_info']
gt_objects = sample['objects']
anchor_targets = [
np_to_variable(sample['rpn_targets']['object'][0][0],
is_cuda=True,
dtype=torch.LongTensor),
np_to_variable(sample['rpn_targets']['object'][0][1],
is_cuda=True),
np_to_variable(sample['rpn_targets']['object'][0][2],
is_cuda=True),
np_to_variable(sample['rpn_targets']['object'][0][3], is_cuda=True)
]
# Forward pass
target_net(im_data, im_info, rpn_data=anchor_targets)
# record loss
loss = target_net.loss
# total loss
train_loss.update(loss.data[0], im_data.size(0))
# object bbox reg
train_loss_obj_box.update(target_net.loss_box.data[0], im_data.size(0))
# object score
train_loss_obj_entropy.update(target_net.loss_cls.data[0],
im_data.size(0))
# accuracy
accuracy_obj.update(target_net.tp, target_net.tf, target_net.fg_cnt,
target_net.bg_cnt)
# backward
optimizer.zero_grad()
torch.cuda.synchronize()
loss.backward()
if not args.disable_clip_gradient:
network.clip_gradient(target_net, 10.)
torch.cuda.synchronize()
optimizer.step()
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if (i + 1) % args.log_interval == 0:
print(
'Epoch: [{0}][{1}/{2}]\t'
'Batch_Time: {batch_time.avg:.3f}s\t'
'lr: {lr: f}\t'
'Loss: {loss.avg:.4f}\n'
'\t[object]:\t'
'tp: {accuracy_obj.true_pos:.3f}, \t'
'tf: {accuracy_obj.true_neg:.3f}, \t'
'fg/bg=({accuracy_obj.foreground:.1f}/{accuracy_obj.background:.1f})\t'
'cls_loss: {cls_loss_object.avg:.3f}\t'
'reg_loss: {reg_loss_object.avg:.3f}'.format(
epoch,
i + 1,
len(train_loader),
batch_time=batch_time,
lr=args.lr,
data_time=data_time,
loss=train_loss,
cls_loss_object=train_loss_obj_entropy,
reg_loss_object=train_loss_obj_box,
accuracy_obj=accuracy_obj))
def test(loader, model, top_Ns, nms=-1., triplet_nms=-1., use_gt_boxes=False):
print '========== Testing ======='
model.eval() # pytorch：将模式设置为evaluation模式
print 'set eval Success!'
rel_cnt = 0. # 已测试的batch个数
rel_cnt_correct = np.zeros(2) # 正确的batch个数
phrase_cnt_correct = np.zeros(2) # 正确的phrase个数
pred_cnt_correct = np.zeros(2) # 正确的predicate（relationship）个数
total_region_rois_num = 0
max_region_rois_num = 0
result = []
batch_time = network.AverageMeter()
end = time.time()
print 'before loop!'
# enumerate: (index, element)
for i, sample in enumerate(
loader): # (im_data, im_info, gt_objects, gt_relationships)
print 'in loop 1'
assert len(sample['visual']
) == 1 # 检查batch_size:author规定evaluation的batch_size必须为1
input_visual = sample['visual'][0].cuda() # 将图片放到GPU上
gt_objects = sample['objects']
gt_relationships = sample['relations']
image_info = sample['image_info']
# Forward pass
print 'in loop 2'
# 测试函数(对单个batch)： model.module.evaluate->models/HDN_v2/factorizable_network_v4.py
# cnt_correct_t（四元，正确数量）
total_cnt_t, cnt_correct_t, eval_result_t = model.module.evaluate(
input_visual,
image_info,
gt_objects,
gt_relationships,
top_Ns=top_Ns,
nms=nms,
triplet_nms=triplet_nms,
use_gt_boxes=use_gt_boxes)
eval_result_t['path'] = sample['path'][0] # for visualization
rel_cnt += total_cnt_t
result.append(eval_result_t)
rel_cnt_correct += cnt_correct_t[0]
phrase_cnt_correct += cnt_correct_t[1]
pred_cnt_correct += cnt_correct_t[2]
total_region_rois_num += cnt_correct_t[3]
max_region_rois_num = cnt_correct_t[3] if cnt_correct_t[
3] > max_region_rois_num else max_region_rois_num
batch_time.update(time.time() - end)
end = time.time()
if (i + 1) % 500 == 0 and i > 0:
print(
'[Evaluation][%d/%d][%.2fs/img][avg: %d subgraphs, max: %d subgraphs]'
% (i + 1, len(loader), batch_time.avg,
total_region_rois_num / float(i + 1), max_region_rois_num))
for idx, top_N in enumerate(top_Ns):
print(
'\tTop-%d Recall:\t[Pred] %2.3f%%\t[Phr] %2.3f%%\t[Rel] %2.3f%%'
% (top_N, pred_cnt_correct[idx] / float(rel_cnt) * 100,
phrase_cnt_correct[idx] / float(rel_cnt) * 100,
rel_cnt_correct[idx] / float(rel_cnt) * 100))
recall = [
rel_cnt_correct / float(rel_cnt), phrase_cnt_correct / float(rel_cnt),
pred_cnt_correct / float(rel_cnt)
]
print('\n====== Done Testing ====')
return recall, result
