test_loader = DataLoader(test_dataset,
batch_size=config_map['batch_size'],
shuffle=False) #, num_workers=4)
data_len = int(len(test_dataset) / config_map['batch_size'])
logger.info('the dataset has %d images' % (len(train_dataset)))
logger.info('the batch_size is %d' % (config_map['batch_size']))
criterion = nn.BCEWithLogitsLoss()
num_iter = 0
best_mAP = 0.0
train_len = len(train_dataset)
train_iter = config_map['resume_epoch'] * len(train_loader)
last_little_mAP = 0.0
my_vis.img('label colors', get_class_color_img())
for epoch in range(config_map['resume_epoch'], config_map['epoch_num']):
backbone_net_p.train()
logger.info('\n\nStarting epoch %d / %d' %
(epoch + 1, config_map['epoch_num']))
logger.info('Learning Rate for this epoch: {}'.format(
optimizer.param_groups[0]['lr']))
epoch_start_time = time.clock()
total_loss = 0.
avg_loss = 0.
for i, (images, mask_label) in enumerate(train_loader):
if my_vis and i % train_config['show_img_iter_during_train'] == 0:
yolo_p.eval()
# img, label_bboxes = next(test_iter)
pred, _ = yolo_p(img[0].unsqueeze(0))
detect_tensor = non_max_suppression(pred.to('cpu'),
conf_thres=0.5,
nms_thres=0.1)
detect_tensor = detect_tensor[0]
show_img = unorm(img[0])
if detect_tensor is not None:
# print(show_img.shape)
show_img = draw_debug_rect(show_img.permute(1, 2, 0),
detect_tensor[..., 1:5],
detect_tensor[..., -1],
detect_tensor[..., -2], name_list)
my_vis.img('detect bboxes show', show_img)
yolo_p.train()
optimizer.zero_grad()
now_loss.backward()
optimizer.step()
it_ed_time = time.clock()
it_cost_time = it_ed_time - it_st_time
if (i + 1) % 5 == 0:
avg_loss = total_loss / (i + 1)
logger.info(
'Epoch [%d/%d], Iter [%d/%d] expect end in %.2f min. Loss: %.4f, average_loss: %.4f'
% (epoch + 1, train_config['epoch_num'], i + 1,
len(train_loader), it_cost_time *
(len(train_loader) - i + 1) // 60,