def get_blob(self, data_dict, scale=None, flip=False):
assert scale is not None
img_list, meta_list = [], []
for image, meta in zip(DCHelper.tolist(data_dict['img']), DCHelper.tolist(data_dict['meta'])):
c, h, w = image.size()
border_hw = [int(h*scale), int(w*scale)]
meta['border_hw'] = border_hw
image = TensorHelper.resize(image, border_hw, mode='bilinear', align_corners=True)
if flip:
image = image.flip([2])
if self.configer.exists('test', 'fit_stride'):
stride = self.configer.get('test', 'fit_stride')
pad_w = 0 if (border_hw[1] % stride == 0) else stride - (border_hw[1] % stride) # right
pad_h = 0 if (border_hw[0] % stride == 0) else stride - (border_hw[0] % stride) # down
expand_image = torch.zeros((c, border_hw[0] + pad_h, border_hw[1] + pad_w)).to(image.device)
expand_image[:, 0:border_hw[0], 0:border_hw[1]] = image
image = expand_image
img_list.append(image)
meta_list.append(meta)
new_data_dict = dict(
img=DCHelper.todc(img_list, stack=True, samples_per_gpu=True),
meta=DCHelper.todc(meta_list, samples_per_gpu=True, cpu_only=True)
)
return new_data_dict
def val(self):
"""
Validation function during the train phase.
"""
self.det_net.eval()
start_time = time.time()
with torch.no_grad():
for j, data_dict in enumerate(self.val_loader):
inputs = data_dict['img']
batch_gt_bboxes = data_dict['bboxes']
batch_gt_labels = data_dict['labels']
metas = data_dict['meta']
data_dict['bboxes'] = DCHelper.todc(
batch_gt_bboxes,
gpu_list=self.configer.get('gpu'),
cpu_only=True)
data_dict['labels'] = DCHelper.todc(
batch_gt_labels,
gpu_list=self.configer.get('gpu'),
cpu_only=True)
data_dict['meta'] = DCHelper.todc(
metas, gpu_list=self.configer.get('gpu'), cpu_only=True)
# Forward pass.
inputs = RunnerHelper.to_device(self, inputs)
loss, test_group = self.det_net(data_dict)
# Compute the loss of the train batch & backward.
loss = loss.mean()
self.val_losses.update(loss.item(), inputs.size(0))
test_indices_and_rois, test_roi_locs, test_roi_scores, test_rois_num = test_group
batch_detections = FastRCNNTest.decode(test_roi_locs,
test_roi_scores,
test_indices_and_rois,
test_rois_num,
self.configer, metas)
batch_pred_bboxes = self.__get_object_list(batch_detections)
self.det_running_score.update(batch_pred_bboxes,
batch_gt_bboxes, batch_gt_labels)
# Update the vars of the val phase.
self.batch_time.update(time.time() - start_time)
start_time = time.time()
RunnerHelper.save_net(self,
self.det_net,
iters=self.runner_state['iters'])
# 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('Val mAP: {}\n'.format(self.det_running_score.get_mAP()))
self.det_running_score.reset()
self.batch_time.reset()
self.val_losses.reset()
self.det_net.train()
def _crop_predict(self, data_dict, crop_size, crop_stride_ratio):
split_batch = list()
height_starts_list = list()
width_starts_list = list()
hw_list = list()
for image in DCHelper.tolist(data_dict['img']):
height, width = image.size()[2:]
hw_list.append([height, width])
np_image = image.squeeze(0).permute(1, 2, 0).cpu().numpy()
height_starts = self._decide_intersection(height, crop_size[1], crop_stride_ratio)
width_starts = self._decide_intersection(width, crop_size[0], crop_stride_ratio)
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, :])
height_starts_list.append(height_starts)
width_starts_list.append(width_starts)
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)
split_batch.append(inputs)
out_list = list()
with torch.no_grad():
results = self.seg_net.forward(DCHelper.todc(split_batch, stack=True, samples_per_gpu=1))
for res in results:
out_list.append(res[-1].permute(0, 2, 3, 1).cpu().numpy())
total_logits = [np.zeros((hw[0], hw[1],
self.configer.get('data', 'num_classes')), np.float32) for hw in hw_list]
count_predictions = [np.zeros((hw[0], hw[1],
self.configer.get('data', 'num_classes')), np.float32) for hw in hw_list]
for i in range(len(height_starts_list)):
index = 0
for height in height_starts_list[i]:
for width in width_starts_list[i]:
total_logits[i][height:height+crop_size[1], width:width+crop_size[0]] += out_list[i][index]
count_predictions[i][height:height+crop_size[1], width:width+crop_size[0]] += 1
index += 1
for i in range(len(total_logits)):
total_logits[i] /= count_predictions[i]
for i, meta in enumerate(DCHelper.tolist(data_dict['meta'])):
total_logits[i] = cv2.resize(total_logits[i][:meta['border_hw'][0], :meta['border_hw'][1]],
(meta['ori_img_size'][0], meta['ori_img_size'][1]),
interpolation=cv2.INTER_CUBIC)
return total_logits
def train(self):
"""
Train function of every epoch during train phase.
"""
self.det_net.train()
start_time = time.time()
# Adjust the learning rate after every epoch.
self.runner_state['epoch'] += 1
for i, data_dict in enumerate(self.train_loader):
Trainer.update(self)
batch_gt_bboxes = data_dict['bboxes']
batch_gt_labels = data_dict['labels']
metas = data_dict['meta']
data_dict['bboxes'] = DCHelper.todc(
batch_gt_bboxes,
gpu_list=self.configer.get('gpu'),
cpu_only=True)
data_dict['labels'] = DCHelper.todc(
batch_gt_labels,
gpu_list=self.configer.get('gpu'),
cpu_only=True)
data_dict['meta'] = DCHelper.todc(
metas, gpu_list=self.configer.get('gpu'), cpu_only=True)
self.data_time.update(time.time() - start_time)
# Forward pass.
loss = self.det_net(data_dict)
loss = loss.mean()
self.train_losses.update(loss.item(), data_dict['img'].size(0))
self.optimizer.zero_grad()
loss.backward()
RunnerHelper.clip_grad(self.det_net, 10.)
self.optimizer.step()
# Update the vars of the train phase.
self.batch_time.update(time.time() - start_time)
start_time = time.time()
self.runner_state['iters'] += 1
# Print the log info & reset the states.
if self.runner_state['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.runner_state['epoch'],
self.runner_state['iters'],
self.configer.get('solver', 'display_iter'),
RunnerHelper.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('lr', 'metric') == 'iters' \
and self.runner_state['iters'] == self.configer.get('solver', 'max_iters'):
break
# Check to val the current model.
if self.runner_state['iters'] % self.configer.get(
'solver', 'test_interval') == 0:
self.val()