def compare_classification(pred, gt):
inside_class = sdf_nn.post_process_sign(pred)
eval_dict = evaluation.compare_predictions_binary_tensors(
ground_truth=gt.squeeze(),
predicted=inside_class,
prediction_name='training_metrics')
return eval_dict
def post_process(batch_pred, train_opt, output_ids, output_pred_ind,
patch_radius, fixed_radius):
# post-processing of the prediction
if 'imp_surf' in train_opt.outputs:
oi_imp = output_ids['imp'][0]
imp_surf_pred = batch_pred[:, output_pred_ind[oi_imp]:
output_pred_ind[oi_imp] + 1]
imp_surf_pred = sdf_nn.post_process_distance(pred=imp_surf_pred)
if not fixed_radius:
imp_surf_pred *= patch_radius.unsqueeze(dim=1)
batch_pred[:, output_pred_ind[oi_imp]:output_pred_ind[oi_imp] + 1] = \
imp_surf_pred
if 'imp_surf_magnitude' in train_opt.outputs:
oi_ism = output_ids['ism'][0]
imp_surf_mag_pred = batch_pred[:, output_pred_ind[oi_ism]:
output_pred_ind[oi_ism] + 1]
imp_surf_mag_pred = sdf_nn.post_process_magnitude(
pred=imp_surf_mag_pred)
if not fixed_radius:
imp_surf_mag_pred *= patch_radius.unsqueeze(dim=1)
batch_pred[:, output_pred_ind[oi_ism]:output_pred_ind[oi_ism] + 1] = \
imp_surf_mag_pred
if 'imp_surf_sign' in train_opt.outputs:
oi_iss = output_ids['iss'][0]
imp_surf_sig_pred = batch_pred[:, output_pred_ind[oi_iss]:
output_pred_ind[oi_iss] + 1]
imp_surf_sig_pred = sdf_nn.post_process_sign(pred=imp_surf_sig_pred)
batch_pred[:, output_pred_ind[oi_iss]:output_pred_ind[oi_iss] +
1] = imp_surf_sig_pred
def post_process(batch_pred, train_opt, output_ids, output_pred_ind,
patch_radius, fixed_radius, trans):
# post-processing of the prediction
if 'imp_surf' in train_opt.outputs:
oi_imp = output_ids['imp'][0]
imp_surf_pred = batch_pred[:, output_pred_ind[oi_imp]:
output_pred_ind[oi_imp] + 1]
imp_surf_pred = sdf_nn.post_process_distance(pred=imp_surf_pred)
if not fixed_radius:
imp_surf_pred *= patch_radius.unsqueeze(dim=1)
batch_pred[:, output_pred_ind[oi_imp]:output_pred_ind[oi_imp] + 1] = \
imp_surf_pred
if 'imp_surf_magnitude' in train_opt.outputs:
oi_ism = output_ids['ism'][0]
imp_surf_mag_pred = batch_pred[:, output_pred_ind[oi_ism]:
output_pred_ind[oi_ism] + 1]
imp_surf_mag_pred = sdf_nn.post_process_magnitude(
pred=imp_surf_mag_pred)
if not fixed_radius:
imp_surf_mag_pred *= patch_radius.unsqueeze(dim=1)
batch_pred[:, output_pred_ind[oi_ism]:output_pred_ind[oi_ism] + 1] = \
imp_surf_mag_pred
if 'imp_surf_sign' in train_opt.outputs:
oi_iss = output_ids['iss'][0]
imp_surf_sig_pred = batch_pred[:, output_pred_ind[oi_iss]:
output_pred_ind[oi_iss] + 1]
imp_surf_sig_pred = sdf_nn.post_process_sign(pred=imp_surf_sig_pred)
batch_pred[:, output_pred_ind[oi_iss]:output_pred_ind[oi_iss] +
1] = imp_surf_sig_pred
if 'normals' in train_opt.outputs:
oi_n = output_ids['normal'][0]
n_pred = batch_pred[:, output_pred_ind[oi_n]:output_pred_ind[oi_n] + 3]
if trans is not None:
n_pred[:, :] = torch.bmm(n_pred.unsqueeze(1),
trans.transpose(2, 1)).squeeze(dim=1)
n_pred_len = torch.max(
n_pred.new_tensor([sys.float_info.epsilon * 100]),
n_pred.norm(p=2, dim=1, keepdim=True))
n_pred = n_pred / n_pred_len
batch_pred[:, output_pred_ind[oi_n]:output_pred_ind[oi_n] + 3] = n_pred