def train_pnet(model_store_path,
end_epoch,
imdb,
batch_size,
frequent=10,
base_lr=0.01,
use_cuda=True):
if not os.path.exists(model_store_path):
os.makedirs(model_store_path)
lossfn = LossFn()
net = PNet(is_train=True, use_cuda=use_cuda)
net.train()
if use_cuda:
net.cuda()
optimizer = torch.optim.Adam(net.parameters(), lr=base_lr)
train_data = TrainImageReader(imdb, 12, batch_size, shuffle=True)
frequent = 10
for cur_epoch in range(1, end_epoch + 1):
train_data.reset() # shuffle
for batch_idx, (image, (gt_label, gt_bbox,
gt_landmark)) in enumerate(train_data):
im_tensor = [
image_tools.convert_image_to_tensor(image[i, :, :, :])
for i in range(image.shape[0])
]
im_tensor = torch.stack(im_tensor)
im_tensor = Variable(im_tensor)
gt_label = Variable(torch.from_numpy(gt_label).float())
gt_bbox = Variable(torch.from_numpy(gt_bbox).float())
# gt_landmark = Variable(torch.from_numpy(gt_landmark).float())
if use_cuda:
im_tensor = im_tensor.cuda()
gt_label = gt_label.cuda()
gt_bbox = gt_bbox.cuda()
# gt_landmark = gt_landmark.cuda()
cls_pred, box_offset_pred = net(im_tensor)
# all_loss, cls_loss, offset_loss = lossfn.loss(gt_label=label_y,gt_offset=bbox_y, pred_label=cls_pred, pred_offset=box_offset_pred)
cls_loss = lossfn.cls_loss(gt_label, cls_pred)
box_offset_loss = lossfn.box_loss(gt_label, gt_bbox,
box_offset_pred)
# landmark_loss = lossfn.landmark_loss(gt_label,gt_landmark,landmark_offset_pred)
all_loss = cls_loss * 1.0 + box_offset_loss * 0.5
if batch_idx % frequent == 0:
accuracy = compute_accuracy(cls_pred, gt_label)
show1 = accuracy.data.cpu().numpy()
show2 = cls_loss.data.cpu().numpy()
show3 = box_offset_loss.data.cpu().numpy()
# show4 = landmark_loss.data.cpu().numpy()
show5 = all_loss.data.cpu().numpy()
print(
"%s : Epoch: %d, Step: %d, accuracy: %s, det loss: %s, bbox loss: %s, all_loss: %s, lr:%s "
% (datetime.datetime.now(), cur_epoch, batch_idx, show1,
show2, show3, show5, base_lr))
optimizer.zero_grad()
all_loss.backward()
optimizer.step()
torch.save(
net.state_dict(),
os.path.join(model_store_path, "pnet_epoch_%d.pt" % cur_epoch))
torch.save(
net,
os.path.join(model_store_path,
"pnet_epoch_model_%d.pkl" % cur_epoch))
def trainer(ops):
set_seed(ops.seed)
use_cuda = torch.cuda.is_available()
device = torch.device('cuda:0' if use_cuda else 'cpu')
if ops.pattern == 'P-Net':
m_XNet = PNet()
mtcnn_detector = None
elif ops.pattern == 'R-Net':
m_XNet = RNet()
elif ops.pattern == 'O-Net':
m_XNet = ONet()
# datasets
dataset = LoadImagesAndLabels(pattern=ops.pattern,
path_img=ops.path_img,
path_anno=ops.path_anno,
batch_size=ops.batch_size)
print('dataset len : ', dataset.__len__())
dataloader = DataLoader(dataset,
batch_size=1,
num_workers=ops.num_workers,
shuffle=True,
pin_memory=False,
drop_last=True)
print('{} : \n'.format(ops.pattern), m_XNet)
m_XNet = m_XNet.to(device)
m_loss = LossFn()
if ops.Optimizer_X == 'Adam':
optimizer = torch.optim.Adam(m_XNet.parameters(),
lr=ops.init_lr,
betas=(0.9, 0.99),
weight_decay=1e-6)
elif ops.Optimizer_X == 'SGD':
optimizer = torch.optim.SGD(m_XNet.parameters(),
lr=ops.init_lr,
momentum=0.9,
weight_decay=1e-6)
elif ops.Optimizer_X == 'RMSprop':
optimizer = torch.optim.RMSprop(m_XNet.parameters(),
lr=ops.init_lr,
alpha=0.9,
weight_decay=1e-6)
else:
print('------>>> Optimizer init error : ', ops.Optimizer_X)
# load finetune model
if os.access(ops.ft_model, os.F_OK):
chkpt = torch.load(ops.ft_model, map_location=device)
print('chkpt:\n', ops.ft_model)
m_XNet.load_state_dict(chkpt)
# train
print(' epoch : ', ops.epochs)
best_loss = np.inf
loss_mean = 0.
loss_cls_mean = 0.
loss_idx = 0.
init_lr = ops.init_lr
loss_cnt = 0
loss_cnt = 0
for epoch in range(0, ops.epochs):
if loss_idx != 0:
if best_loss > (loss_mean / loss_idx):
best_loss = loss_mean / loss_idx
loss_cnt = 0
else:
if loss_cnt > 3:
init_lr = init_lr * 0.5
set_learning_rate(optimizer, init_lr)
loss_cnt = 0
else:
loss_cnt += 1
loss_mean = 0.
loss_cls_mean = 0.
loss_idx = 0.
print('\nepoch %d ' % epoch)
m_XNet = m_XNet.train()
random.shuffle(dataset.annotations) # shuffle 图片组合
for i, (imgs, gt_labels, gt_offsets, pos_num, part_num,
neg_num) in enumerate(dataloader):
imgs = imgs.squeeze(0)
gt_labels = gt_labels.squeeze(0)
gt_offsets = gt_offsets.squeeze(0)
# print('imgs size {}, labels size {}, offsets size {}'.format(imgs.size(),gt_labels.size(),gt_offsets.size()))
if use_cuda:
imgs = imgs.cuda() # (bs, 3, h, w)
gt_labels = gt_labels.cuda()
gt_offsets = gt_offsets.cuda()
cls_pred, box_offset_pred = m_XNet(imgs)
cls_loss = m_loss.focal_Loss(gt_labels, cls_pred)
box_offset_loss = m_loss.box_loss(gt_labels, gt_offsets,
box_offset_pred)
if ops.pattern == 'O-Net':
all_loss = cls_loss * 1.0 + box_offset_loss * 0.4
elif ops.pattern == 'R-Net':
all_loss = cls_loss * 1.0 + box_offset_loss * 0.6
else:
all_loss = cls_loss * 1.0 + box_offset_loss * 1.0
loss_mean += all_loss.item()
loss_cls_mean += cls_loss.item()
loss_idx += 1.
optimizer.zero_grad()
all_loss.backward()
optimizer.step()
if i % 5 == 0:
loc_time = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())
print("[%s]-<%s>Epoch: %d, [%d/%d],lr:%.6f,Loss:%.5f - Mean Loss:%.5f - Mean cls loss:%5f,cls_loss:%.5f ,bbox_loss:%.5f,imgs_batch: %4d,best_loss: %.5f" \
% (loc_time,ops.pattern,epoch, i,dataset.__len__(), optimizer.param_groups[0]['lr'], \
all_loss.item(),loss_mean/loss_idx,loss_cls_mean/loss_idx,cls_loss.item(),box_offset_loss.item(),imgs.size()[0],best_loss), ' ->pos:{},part:{},neg:{}'.format(pos_num.item(),part_num.item(),neg_num.item()))
if i % 50 == 0 and i > 1:
accuracy = compute_accuracy(cls_pred, gt_labels)
print("\n ------------- >>> accuracy: %f\n" %
(accuracy.item()))
accuracy = compute_accuracy(cls_pred, gt_labels)
torch.save(m_XNet.state_dict(),
ops.ckpt + '{}_latest.pth'.format(ops.pattern))
if i % 80 == 0 and i > 1:
torch.save(
m_XNet.state_dict(),
ops.ckpt + '{}_epoch-{}.pth'.format(ops.pattern, epoch))
def train_pnet(self, train_data_path):
device = torch.device('cuda')
lossfn = LossFn()
net = PNet()
# 返回 一样的 net = net.to(device)
net.to(device)
# 切换到train 状态 net.eval() 测试状态
net.train()
# print(net)
optimizer = torch.optim.Adam(net.parameters(), lr=1e-3)
self.viz.line(Y=torch.FloatTensor([0.]),
X=torch.FloatTensor([0.]),
win='pnet_train_loss',
opts=dict(title='train loss'))
# 加载数据 ratios : pos:part:neg:landmark
trian_datasets = DataReader(train_data_path,
im_size=12,
transform=self.trainTransform,
batch_size=4096,
ratios=(2, 1, 1, 2))
for epoch in range(1):
print("epoch:", epoch)
for step, (imgs, cls_labels, rois,
landmarks) in enumerate(trian_datasets):
# [b, 3, 12, 12],[b],[4],[10]
im_tensor = imgs.to(device)
cls_labels = cls_labels.to(device)
rois = rois.to(device)
landmarks = landmarks.to(device)
cls_pred, box_offset_pred, landmarks_pred = net(im_tensor)
# 貌似这里打印最后一个的loss,对于整体来说不怎么准确
cls_loss = lossfn.cls_loss(cls_labels, cls_pred)
box_offset_loss = lossfn.box_loss(cls_labels, rois,
box_offset_pred)
landmark_loss = lossfn.landmark_loss(cls_labels, landmarks,
landmarks_pred)
print("cls_loss:", cls_loss)
print("box_offset_loss:", box_offset_loss)
print("landmark_loss:", landmark_loss)
all_loss = cls_loss * 1.0 + box_offset_loss * 0.5 + landmark_loss * 0.5
self.viz.line(Y=torch.FloatTensor([all_loss]),
X=torch.FloatTensor([step]),
win='pnet_train_loss',
update='append')
optimizer.zero_grad()
all_loss.backward()
optimizer.step()
print("all_loss:", all_loss)
print("-" * 40, "step:", step, "-" * 40)
if step % 1000 == 0:
accuracy = compute_accuracy(cls_pred, cls_labels)
recoll = compute_recoll(cls_pred, cls_labels)
print(
"=" * 80, "\n\n=> acc:{}\n=> recoll:{}\n\n".format(
accuracy, recoll), "=" * 80)
if step % 1000 == 0:
torch.save(
net.state_dict(),
os.path.join("../data/models/",
"pnet_epoch_%d.pt" % epoch))
torch.save(
net,
os.path.join("../data/models/",
"pnet_epoch_model_%d.pkl" % epoch))
epoch += 1