loss = opt.w_cls * loss_cls + \
opt.w_geo * loss_geo + \
opt.w_sem * loss_sem + \
opt.w_ort * loss_ort + \
opt.w_eqv * loss_eqv
# Visualization
with torch.no_grad():
if epoch % 5 == 0:
pams_viz = pams.view(-1, nc * k, h, w)
pams_tf_viz = pams_tf.view(-1, nc * k, h, w)
pams_viz = pams_viz / pams_viz.max(
2, keepdim=True)[0].max(3, keepdim=True)[0]
pams_tf_viz = pams_tf_viz / pams_tf_viz.max(
2, keepdim=True)[0].max(3, keepdim=True)[0]
viz.vis_inputs(epoch, inputs, prefix='')
viz.vis_inputs(epoch, inputs_tf, prefix='_tf')
viz.vis_inputs(epoch, saliency, prefix='saliency')
viz.vis_DFF_heatmaps(epoch,
pams_viz,
threshold=0.1,
prefix='pams')
viz.vis_DFF_heatmaps(epoch,
pams_tf_viz,
threshold=0.1,
prefix='pams_tf')
if phase == 'train':
viz.vis_losses(epoch, [
loss, loss_cls, loss_geo, loss_sem, loss_ort,
loss_eqv
