def load_loss(self, config_dict):
# normal
if config_dict.get('MAE', False):
pairwise_loss = torch.nn.modules.loss.L1Loss()
else:
pairwise_loss = torch.nn.modules.loss.MSELoss()
image_pixel_loss = losses_generic.LossOnDict(key='img_crop',
loss=pairwise_loss)
image_imgNet_bare = losses_images.ImageNetCriterium(
criterion=pairwise_loss,
weight=config_dict['loss_weight_imageNet'],
do_maxpooling=config_dict.get('do_maxpooling', True))
image_imgNet_loss = losses_generic.LossOnDict(key='img_crop',
loss=image_imgNet_bare)
losses_train = []
losses_test = []
if 'img_crop' in config_dict['output_types']:
if config_dict['loss_weight_rgb'] > 0:
losses_train.append(image_pixel_loss)
losses_test.append(image_pixel_loss)
if config_dict['loss_weight_imageNet'] > 0:
losses_train.append(image_imgNet_loss)
losses_test.append(image_imgNet_loss)
loss_train = losses_generic.PreApplyCriterionListDict(losses_train,
sum_losses=True)
loss_test = losses_generic.PreApplyCriterionListDict(losses_test,
sum_losses=True)
# annotation and pred is organized as a list, to facilitate multiple output types (e.g. heatmap and 3d loss)
return loss_train, loss_test
def load_loss(self, config_dict):
# normal
if config_dict.get('MAE', False):
pairwise_loss = torch.nn.modules.loss.L1Loss()
else:
pairwise_loss = torch.nn.modules.loss.MSELoss()
if 1: #"box" in config_dict['training_set'] or "walk_full" in config_dict['training_set']:
pairwise_loss = losses_generic.LossInstanceMeanStdFromLabel(
pairwise_loss)
img_key = 'img'
image_pixel_loss = losses_generic.LossOnDict(key=img_key,
loss=pairwise_loss)
image_imgNet_bare = losses_images.ImageNetCriterium(
criterion=pairwise_loss,
weight=config_dict['loss_weight_imageNet'],
do_maxpooling=config_dict.get('do_maxpooling', True))
image_imgNet_loss = losses_generic.LossOnDict(key=img_key,
loss=image_imgNet_bare)
losses_train = []
losses_test = []
if img_key in config_dict['output_types']:
if config_dict['loss_weight_rgb'] > 0:
losses_train.append(image_pixel_loss)
losses_test.append(image_pixel_loss)
if config_dict['loss_weight_imageNet'] > 0:
losses_train.append(image_imgNet_loss)
losses_test.append(image_imgNet_loss)
# priors on crop
if config_dict['spatial_transformer']:
losses_train.append(
losses_generic.AffineCropPositionPrior(
config_dict['fullFrameResolution'], weight=0.1))
loss_train = losses_generic.PreApplyCriterionListDict(losses_train,
sum_losses=True)
loss_test = losses_generic.PreApplyCriterionListDict(losses_test,
sum_losses=True)
# annotation and pred is organized as a list, to facilitate multiple output types (e.g. heatmap and 3d loss)
return loss_train, loss_test
def load_loss(self, config_dict):
weight = 1
if config_dict['training_set'] in ['h36m', 'h36m_mpii']:
weight = 17 / 16 # becasue spine is set to root = 0
print(
"MPJPE test weight = {}, to normalize different number of joints".
format(weight))
# normal
if config_dict.get('MAE', False):
pairwise_loss = torch.nn.modules.loss.L1Loss()
else:
pairwise_loss = torch.nn.modules.loss.MSELoss()
image_pixel_loss = losses_generic.LossOnDict(key='img_crop',
loss=pairwise_loss)
image_imgNet_bare = losses_images.ImageNetCriterium(
criterion=pairwise_loss,
weight=config_dict['loss_weight_imageNet'],
do_maxpooling=config_dict.get('do_maxpooling', True))
image_imgNet_loss = losses_generic.LossOnDict(key='img_crop',
loss=image_imgNet_bare)
losses_train = []
losses_test = []
if 'img_crop' in config_dict['output_types']:
if config_dict['loss_weight_rgb'] > 0:
losses_train.append(image_pixel_loss)
losses_test.append(image_pixel_loss)
if config_dict['loss_weight_imageNet'] > 0:
losses_train.append(image_imgNet_loss)
losses_test.append(image_imgNet_loss)
loss_train = losses_generic.PreApplyCriterionListDict(losses_train,
sum_losses=True)
loss_test = losses_generic.PreApplyCriterionListDict(losses_test,
sum_losses=True)
# annotation and pred is organized as a list, to facilitate multiple output types (e.g. heatmap and 3d loss)
return loss_train, loss_test