def test_phase(opt, net, testloader, cuda=False, log_save_path=None):
with torch.no_grad():
net.eval()
start = time()
avg_frame_rate = 0
mae = 0.0
rmse = 0.0
me = 0.0
for j, data in enumerate(testloader):
inputs, labels = data['image'], data['all_num']
inputs, labels = inputs.type(
torch.float32), labels.unsqueeze(1).type(torch.float32)
if cuda:
inputs, labels = inputs.cuda(), labels.cuda()
# process with SSDCNet
features = net(inputs)
div_res = net.resample(features)
merge_res = net.parse_merge(div_res)
outputs = merge_res['div' + str(net.div_times)]
del merge_res
# compute the count value ( gt & pre)
pre = (outputs).sum()
gt = labels.sum()
# compute the metrics
mae += abs(pre - gt)
rmse += (pre - gt) * (pre - gt)
me += (pre - gt)
end = time()
running_frame_rate = opt['test_batch_size'] * float(1 /
(end - start))
avg_frame_rate = (avg_frame_rate * j + running_frame_rate) / (j +
1)
if j % 1 == 0: # print every 2000 mini-batches
print(
'Test:[%5d/%5d] pre: %.3f gt:%.3f err:%.3f frame: %.2fHz/%.2fHz'
% (j + 1, len(testloader), pre, gt, pre - gt,
running_frame_rate, avg_frame_rate))
start = time()
im_num = len(testloader)
log_str = '%10s\t %8s\t &%8s\t &%8s\t\\\\' % (' ', 'mae', 'rmse',
'me') + '\n'
log_str += '%-10s\t %8.3f\t %8.3f\t %8.3f\t' % (
'test', mae / im_num, math.sqrt(rmse / im_num), me / im_num) + '\n'
print(log_str)
if log_save_path:
txt_write(log_save_path, log_str, mode='w')
# return log
im_num = len(testloader)
test_dict = dict()
test_dict['mae'] = mae / im_num
test_dict['mse'] = math.sqrt(rmse / (im_num))
test_dict['me'] = me / (im_num)
return test_dict
def test_phase(opt, net, testloader, log_save_path=None):
with torch.no_grad():
net.eval()
start = time()
avg_frame_rate = 0
mae = 0.0
rmse = 0.0
me = 0.0
for j, data in enumerate(testloader):
inputs, labels = data['image'], data['target']
inputs, labels = inputs.type(
torch.float32), labels.unsqueeze(1).type(torch.float32)
inputs, labels = inputs.cuda(), labels.cuda()
features = net(inputs)
div_res = net.resample(features)
merge_res = net.parse_merge(div_res)
outputs = merge_res['div' + str(net.args['div_times'])]
del merge_res
pre = (outputs).sum()
gt = labels.sum()
mae += abs(pre - gt)
rmse += (pre - gt) * (pre - gt)
me += (pre - gt)
end = time()
running_frame_rate = opt['test_batch_size'] * float(1 /
(end - start))
avg_frame_rate = (avg_frame_rate * j + running_frame_rate) / (j +
1)
if j % 1 == 0: # print every 2000 mini-batches
print(
'Test:[%5d/%5d] pre: %.3f gt:%.3f err:%.3f frame: %.2fHz/%.2fHz'
% (j + 1, len(testloader), pre, gt, pre - gt,
running_frame_rate, avg_frame_rate))
start = time()
log_str = '%10s\t %8s\t &%8s\t &%8s\t\\\\' % (' ', 'mae', 'rmse',
'me') + '\n'
log_str += '%-10s\t %8.3f\t %8.3f\t %8.3f\t' % (
'test', mae / (j + 1), math.sqrt(rmse /
(j + 1)), me / (j + 1)) + '\n'
if log_save_path:
txt_write(log_save_path, log_str, mode='w')
im_num = len(testloader)
return mae / (im_num), math.sqrt(rmse / (im_num)), me / (im_num)