class Trainer(object):
"""
The class for Pose Estimation. Include train, val, val & predict.
"""
def __init__(self, configer):
self.configer = configer
self.batch_time = AverageMeter()
self.data_time = AverageMeter()
self.train_losses = AverageMeter()
self.val_losses = AverageMeter()
self.seg_running_score = SegRunningScore(configer)
self.seg_visualizer = SegVisualizer(configer)
self.seg_loss_manager = LossManager(configer)
self.module_runner = ModuleRunner(configer)
self.seg_model_manager = ModelManager(configer)
self.seg_data_loader = DataLoader(configer)
self.optim_scheduler = OptimScheduler(configer)
self.seg_net = None
self.train_loader = None
self.val_loader = None
self.optimizer = None
self.scheduler = None
self._init_model()
def _init_model(self):
self.seg_net = self.seg_model_manager.semantic_segmentor()
self.seg_net = self.module_runner.load_net(self.seg_net)
Log.info('Params Group Method: {}'.format(
self.configer.get('optim', 'group_method')))
if self.configer.get('optim', 'group_method') == 'decay':
params_group = self.group_weight(self.seg_net)
else:
assert self.configer.get('optim', 'group_method') is None
params_group = self._get_parameters()
self.optimizer, self.scheduler = self.optim_scheduler.init_optimizer(
params_group)
self.train_loader = self.seg_data_loader.get_trainloader()
self.val_loader = self.seg_data_loader.get_valloader()
self.pixel_loss = self.seg_loss_manager.get_seg_loss()
@staticmethod
def group_weight(module):
group_decay = []
group_no_decay = []
for m in module.modules():
if isinstance(m, nn.Linear):
group_decay.append(m.weight)
if m.bias is not None:
group_no_decay.append(m.bias)
elif isinstance(m, nn.modules.conv._ConvNd):
group_decay.append(m.weight)
if m.bias is not None:
group_no_decay.append(m.bias)
else:
if hasattr(m, 'weight'):
group_no_decay.append(m.weight)
if hasattr(m, 'bias'):
group_no_decay.append(m.bias)
assert len(list(
module.parameters())) == len(group_decay) + len(group_no_decay)
groups = [
dict(params=group_decay),
dict(params=group_no_decay, weight_decay=.0)
]
return groups
def _get_parameters(self):
bb_lr = []
nbb_lr = []
params_dict = dict(self.seg_net.named_parameters())
for key, value in params_dict.items():
if 'backbone' not in key:
nbb_lr.append(value)
else:
bb_lr.append(value)
params = [{
'params': bb_lr,
'lr': self.configer.get('lr', 'base_lr')
}, {
'params':
nbb_lr,
'lr':
self.configer.get('lr', 'base_lr') *
self.configer.get('lr', 'nbb_mult')
}]
return params
def __train(self):
"""
Train function of every epoch during train phase.
"""
self.seg_net.train()
start_time = time.time()
for i, data_dict in enumerate(self.train_loader):
if self.configer.get('lr', 'metric') == 'iters':
self.scheduler.step(self.configer.get('iters'))
else:
self.scheduler.step(self.configer.get('epoch'))
if self.configer.get('lr', 'is_warm'):
self.module_runner.warm_lr(self.configer.get('iters'),
self.scheduler,
self.optimizer,
backbone_list=[
0,
])
inputs = data_dict['img']
targets = data_dict['labelmap']
self.data_time.update(time.time() - start_time)
# Change the data type.
# inputs, targets = self.module_runner.to_device(inputs, targets)
# Forward pass.
outputs = self.seg_net(inputs)
# outputs = self.module_utilizer.gather(outputs)
# Compute the loss of the train batch & backward.
loss = self.pixel_loss(outputs,
targets,
gathered=self.configer.get(
'network', 'gathered'))
if self.configer.exists('train', 'loader') and self.configer.get(
'train', 'loader') == 'rs':
batch_size = self.configer.get(
'train', 'batch_size') * self.configer.get(
'train', 'batch_per_gpu')
self.train_losses.update(loss.item(), batch_size)
else:
self.train_losses.update(loss.item(), inputs.size(0))
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
# Update the vars of the train phase.
self.batch_time.update(time.time() - start_time)
start_time = time.time()
self.configer.plus_one('iters')
# Print the log info & reset the states.
if self.configer.get('iters') % self.configer.get(
'solver', 'display_iter') == 0:
Log.info(
'Train Epoch: {0}\tTrain Iteration: {1}\t'
'Time {batch_time.sum:.3f}s / {2}iters, ({batch_time.avg:.3f})\t'
'Data load {data_time.sum:.3f}s / {2}iters, ({data_time.avg:3f})\n'
'Learning rate = {3}\tLoss = {loss.val:.8f} (ave = {loss.avg:.8f})\n'
.format(self.configer.get('epoch'),
self.configer.get('iters'),
self.configer.get('solver', 'display_iter'),
self.module_runner.get_lr(self.optimizer),
batch_time=self.batch_time,
data_time=self.data_time,
loss=self.train_losses))
self.batch_time.reset()
self.data_time.reset()
self.train_losses.reset()
if self.configer.get('iters') == self.configer.get(
'solver', 'max_iters'):
break
# Check to val the current model.
if self.configer.get('iters') % self.configer.get(
'solver', 'test_interval') == 0:
self.__val()
self.configer.plus_one('epoch')
def __val(self, data_loader=None):
"""
Validation function during the train phase.
"""
self.seg_net.eval()
start_time = time.time()
data_loader = self.val_loader if data_loader is None else data_loader
for j, data_dict in enumerate(data_loader):
inputs = data_dict['img']
targets = data_dict['labelmap']
with torch.no_grad():
# Change the data type.
inputs, targets = self.module_runner.to_device(inputs, targets)
# Forward pass.
outputs = self.seg_net(inputs)
# Compute the loss of the val batch.
loss = self.pixel_loss(outputs,
targets,
gathered=self.configer.get(
'network', 'gathered'))
outputs = self.module_runner.gather(outputs)
self.val_losses.update(loss.item(), inputs.size(0))
self._update_running_score(outputs[-1], data_dict['meta'])
# self.seg_running_score.update(pred.max(1)[1].cpu().numpy(), targets.cpu().numpy())
# Update the vars of the val phase.
self.batch_time.update(time.time() - start_time)
start_time = time.time()
self.configer.update(['performance'],
self.seg_running_score.get_mean_iou())
self.configer.update(['val_loss'], self.val_losses.avg)
self.module_runner.save_net(self.seg_net, save_mode='performance')
self.module_runner.save_net(self.seg_net, save_mode='val_loss')
# Print the log info & reset the states.
Log.info('Test Time {batch_time.sum:.3f}s, ({batch_time.avg:.3f})\t'
'Loss {loss.avg:.8f}\n'.format(batch_time=self.batch_time,
loss=self.val_losses))
Log.info('Mean IOU: {}\n'.format(
self.seg_running_score.get_mean_iou()))
Log.info('Pixel ACC: {}\n'.format(
self.seg_running_score.get_pixel_acc()))
self.batch_time.reset()
self.val_losses.reset()
self.seg_running_score.reset()
self.seg_net.train()
def _update_running_score(self, pred, metas):
pred = pred.permute(0, 2, 3, 1)
for i in range(pred.size(0)):
ori_img_size = metas[i]['ori_img_size']
border_size = metas[i]['border_size']
ori_target = metas[i]['ori_target']
total_logits = cv2.resize(
pred[i, :border_size[1], :border_size[0]].cpu().numpy(),
tuple(ori_img_size),
interpolation=cv2.INTER_CUBIC)
labelmap = np.argmax(total_logits, axis=-1)
self.seg_running_score.update(labelmap[None], ori_target[None])
def train(self):
# cudnn.benchmark = True
if self.configer.get('network',
'resume') is not None and self.configer.get(
'network', 'resume_val'):
self.__val()
while self.configer.get('iters') < self.configer.get(
'solver', 'max_iters'):
self.__train()
self.__val(data_loader=self.seg_data_loader.get_valloader(
dataset='val'))
self.__val(data_loader=self.seg_data_loader.get_valloader(
dataset='train'))
class Tester(object):
def __init__(self, configer):
self.configer = configer
self.blob_helper = BlobHelper(configer)
self.seg_visualizer = SegVisualizer(configer)
self.seg_parser = SegParser(configer)
self.seg_model_manager = ModelManager(configer)
self.seg_data_loader = DataLoader(configer)
self.module_runner = ModuleRunner(configer)
self.device = torch.device(
'cpu' if self.configer.get('gpu') is None else 'cuda')
self.seg_net = None
self._init_model()
def _init_model(self):
self.seg_net = self.seg_model_manager.semantic_segmentor()
self.seg_net = self.module_runner.load_net(self.seg_net)
self.seg_net.eval()
def _get_blob(self, ori_image, scale=None):
assert scale is not None
image = None
if self.configer.exists('test', 'input_size'):
image = self.blob_helper.make_input(image=ori_image,
input_size=self.configer.get(
'test', 'input_size'),
scale=scale)
elif self.configer.exists(
'test', 'min_side_length') and not self.configer.exists(
'test', 'max_side_length'):
image = self.blob_helper.make_input(
image=ori_image,
min_side_length=self.configer.get('test', 'min_side_length'),
scale=scale)
elif not self.configer.exists(
'test', 'min_side_length') and self.configer.exists(
'test', 'max_side_length'):
image = self.blob_helper.make_input(
image=ori_image,
max_side_length=self.configer.get('test', 'max_side_length'),
scale=scale)
elif self.configer.exists('test',
'min_side_length') and self.configer.exists(
'test', 'max_side_length'):
image = self.blob_helper.make_input(
image=ori_image,
min_side_length=self.configer.get('test', 'min_side_length'),
max_side_length=self.configer.get('test', 'max_side_length'),
scale=scale)
else:
Log.error('Test setting error')
exit(1)
b, c, h, w = image.size()
border_hw = [h, w]
if self.configer.exists('test', 'fit_stride'):
stride = self.configer.get('test', 'fit_stride')
pad_w = 0 if (w % stride == 0) else stride - (w % stride) # right
pad_h = 0 if (h % stride == 0) else stride - (h % stride) # down
expand_image = torch.zeros(
(b, c, h + pad_h, w + pad_w)).to(image.device)
expand_image[:, :, 0:h, 0:w] = image
image = expand_image
return image, border_hw
def __test_img(self, image_path, label_path, vis_path, raw_path):
Log.info('Image Path: {}'.format(image_path))
ori_image = ImageHelper.read_image(
image_path,
tool=self.configer.get('data', 'image_tool'),
mode=self.configer.get('data', 'input_mode'))
total_logits = None
if self.configer.get('test', 'mode') == 'ss_test':
total_logits = self.ss_test(ori_image)
elif self.configer.get('test', 'mode') == 'sscrop_test':
total_logits = self.sscrop_test(ori_image)
elif self.configer.get('test', 'mode') == 'ms_test':
total_logits = self.ms_test(ori_image)
elif self.configer.get('test', 'mode') == 'mscrop_test':
total_logits = self.mscrop_test(ori_image)
else:
Log.error('Invalid test mode:{}'.format(
self.configer.get('test', 'mode')))
exit(1)
label_map = np.argmax(total_logits, axis=-1)
label_img = np.array(label_map, dtype=np.uint8)
ori_img_bgr = ImageHelper.get_cv2_bgr(ori_image,
mode=self.configer.get(
'data', 'input_mode'))
image_canvas = self.seg_parser.colorize(label_img,
image_canvas=ori_img_bgr)
ImageHelper.save(image_canvas, save_path=vis_path)
ImageHelper.save(ori_image, save_path=raw_path)
if self.configer.exists('data', 'label_list'):
label_img = self.__relabel(label_img)
if self.configer.exists('data',
'reduce_zero_label') and self.configer.get(
'data', 'reduce_zero_label'):
label_img = label_img + 1
label_img = label_img.astype(np.uint8)
label_img = Image.fromarray(label_img, 'P')
Log.info('Label Path: {}'.format(label_path))
ImageHelper.save(label_img, label_path)
def ss_test(self, ori_image):
ori_width, ori_height = ImageHelper.get_size(ori_image)
total_logits = np.zeros(
(ori_height, ori_width, self.configer.get('data', 'num_classes')),
np.float32)
image, border_hw = self._get_blob(ori_image, scale=1.0)
results = self._predict(image)
results = cv2.resize(results[:border_hw[0], :border_hw[1]],
(ori_width, ori_height),
interpolation=cv2.INTER_CUBIC)
total_logits += results
return total_logits
def sscrop_test(self, ori_image):
ori_width, ori_height = ImageHelper.get_size(ori_image)
total_logits = np.zeros(
(ori_height, ori_width, self.configer.get('data', 'num_classes')),
np.float32)
image, border_hw = self._get_blob(ori_image, scale=1.0)
crop_size = self.configer.get('test', 'crop_size')
if image.size()[3] > crop_size[0] and image.size()[2] > crop_size[1]:
results = self._crop_predict(image, crop_size)
else:
results = self._predict(image)
results = cv2.resize(results[:border_hw[0], :border_hw[1]],
(ori_width, ori_height),
interpolation=cv2.INTER_CUBIC)
total_logits += results
return total_logits
def mscrop_test(self, ori_image):
ori_width, ori_height = ImageHelper.get_size(ori_image)
crop_size = self.configer.get('test', 'crop_size')
total_logits = np.zeros(
(ori_height, ori_width, self.configer.get('data', 'num_classes')),
np.float32)
for scale in self.configer.get('test', 'scale_search'):
image, border_hw = self._get_blob(ori_image, scale=scale)
if image.size()[3] > crop_size[0] and image.size(
)[2] > crop_size[1]:
results = self._crop_predict(image, crop_size)
else:
results = self._predict(image)
results = cv2.resize(results[:border_hw[0], :border_hw[1]],
(ori_width, ori_height),
interpolation=cv2.INTER_CUBIC)
total_logits += results
if self.configer.get('data', 'image_tool') == 'cv2':
mirror_image = cv2.flip(ori_image, 1)
else:
mirror_image = ori_image.transpose(Image.FLIP_LEFT_RIGHT)
image, border_hw = self._get_blob(mirror_image, scale=1.0)
if image.size()[3] > crop_size[0] and image.size(
)[2] > crop_size[1]:
results = self._crop_predict(image, crop_size)
else:
results = self._predict(image)
results = results[:border_hw[0], :border_hw[1]]
results = cv2.resize(results[:, ::-1], (ori_width, ori_height),
interpolation=cv2.INTER_CUBIC)
total_logits += results
return total_logits
def ms_test(self, ori_image):
ori_width, ori_height = ImageHelper.get_size(ori_image)
total_logits = np.zeros(
(ori_height, ori_width, self.configer.get('data', 'num_classes')),
np.float32)
for scale in self.configer.get('test', 'scale_search'):
image, border_hw = self._get_blob(ori_image, scale=scale)
results = self._predict(image)
results = cv2.resize(results[:border_hw[0], :border_hw[1]],
(ori_width, ori_height),
interpolation=cv2.INTER_CUBIC)
total_logits += results
if self.configer.get('data', 'image_tool') == 'cv2':
mirror_image = cv2.flip(ori_image, 1)
else:
mirror_image = ori_image.transpose(Image.FLIP_LEFT_RIGHT)
image, border_hw = self._get_blob(mirror_image, scale=scale)
results = self._predict(image)
results = results[:border_hw[0], :border_hw[1]]
results = cv2.resize(results[:, ::-1], (ori_width, ori_height),
interpolation=cv2.INTER_CUBIC)
total_logits += results
return total_logits
def _crop_predict(self, image, crop_size):
height, width = image.size()[2:]
np_image = image.squeeze(0).permute(1, 2, 0).cpu().numpy()
height_starts = self._decide_intersection(height, crop_size[1])
width_starts = self._decide_intersection(width, crop_size[0])
split_crops = []
for height in height_starts:
for width in width_starts:
image_crop = np_image[height:height + crop_size[1],
width:width + crop_size[0]]
split_crops.append(image_crop[np.newaxis, :])
split_crops = np.concatenate(
split_crops, axis=0) # (n, crop_image_size, crop_image_size, 3)
inputs = torch.from_numpy(split_crops).permute(0, 3, 1,
2).to(self.device)
with torch.no_grad():
results = self.seg_net.forward(inputs)
results = results[-1].permute(0, 2, 3, 1).cpu().numpy()
reassemble = np.zeros(
(np_image.shape[0], np_image.shape[1], results.shape[-1]),
np.float32)
index = 0
for height in height_starts:
for width in width_starts:
reassemble[height:height + crop_size[1],
width:width + crop_size[0]] += results[index]
index += 1
return reassemble
def _decide_intersection(self, total_length, crop_length):
stride = int(crop_length * self.configer.get(
'test', 'crop_stride_ratio')) # set the stride as the paper do
times = (total_length - crop_length) // stride + 1
cropped_starting = []
for i in range(times):
cropped_starting.append(stride * i)
if total_length - cropped_starting[-1] > crop_length:
cropped_starting.append(total_length -
crop_length) # must cover the total image
return cropped_starting
def _predict(self, inputs):
with torch.no_grad():
results = self.seg_net.forward(inputs)
results = results[-1].squeeze(0).permute(1, 2, 0).cpu().numpy()
return results
def __relabel(self, label_map):
height, width = label_map.shape
label_dst = np.zeros((height, width), dtype=np.uint8)
for i in range(self.configer.get('data', 'num_classes')):
label_dst[label_map == i] = self.configer.get(
'data', 'label_list')[i]
label_dst = np.array(label_dst, dtype=np.uint8)
return label_dst
def test(self):
base_dir = os.path.join(self.configer.get('project_dir'),
'val/results', self.configer.get('dataset'))
test_img = self.configer.get('test', 'test_img')
test_dir = self.configer.get('test', 'test_dir')
if test_img is None and test_dir is None:
Log.error('test_img & test_dir not exists.')
exit(1)
if test_img is not None and test_dir is not None:
Log.error('Either test_img or test_dir.')
exit(1)
if test_img is not None:
base_dir = os.path.join(base_dir, 'test_img')
filename = test_img.rstrip().split('/')[-1]
label_path = os.path.join(
base_dir, 'label',
'{}.png'.format('.'.join(filename.split('.')[:-1])))
raw_path = os.path.join(base_dir, 'raw', filename)
vis_path = os.path.join(
base_dir, 'vis',
'{}_vis.png'.format('.'.join(filename.split('.')[:-1])))
FileHelper.make_dirs(label_path, is_file=True)
FileHelper.make_dirs(raw_path, is_file=True)
FileHelper.make_dirs(vis_path, is_file=True)
self.__test_img(test_img, label_path, vis_path, raw_path)
else:
base_dir = os.path.join(
base_dir, 'test_dir',
self.configer.get('checkpoints', 'checkpoints_name'),
self.configer.get('test', 'out_dir'))
FileHelper.make_dirs(base_dir)
for filename in FileHelper.list_dir(test_dir):
image_path = os.path.join(test_dir, filename)
label_path = os.path.join(
base_dir, 'label',
'{}.png'.format('.'.join(filename.split('.')[:-1])))
raw_path = os.path.join(base_dir, 'raw', filename)
vis_path = os.path.join(
base_dir, 'vis',
'{}_vis.png'.format('.'.join(filename.split('.')[:-1])))
FileHelper.make_dirs(label_path, is_file=True)
FileHelper.make_dirs(raw_path, is_file=True)
FileHelper.make_dirs(vis_path, is_file=True)
self.__test_img(image_path, label_path, vis_path, raw_path)
def debug(self):
base_dir = os.path.join(self.configer.get('project_dir'),
'vis/results', self.configer.get('dataset'),
'debug')
if not os.path.exists(base_dir):
os.makedirs(base_dir)
count = 0
for i, data_dict in enumerate(self.seg_data_loader.get_trainloader()):
inputs = data_dict['img']
targets = data_dict['labelmap']
for j in range(inputs.size(0)):
count = count + 1
if count > 20:
exit(1)
image_bgr = self.blob_helper.tensor2bgr(inputs[j])
label_map = targets[j].numpy()
image_canvas = self.seg_parser.colorize(label_map,
image_canvas=image_bgr)
cv2.imwrite(
os.path.join(base_dir, '{}_{}_vis.png'.format(i, j)),
image_canvas)
cv2.imshow('main', image_canvas)
cv2.waitKey()