def save_prediction_losses_results(self, eval_avod_losses,
num_valid_samples, global_step,
predictions_base_dir, box_rep):
"""Helper function to save the AVOD loss evaluation results.
Args:
eval_avod_losses: A dictionary containing the loss sums
num_valid_samples: An int, number of valid evaluated samples
i.e. samples with valid ground-truth.
global_step: Global step at which the metrics are computed.
predictions_base_dir: Base directory for storing the results.
box_rep: A string, the format of the 3D bounding box
one of 'box_3d', 'box_8c' etc.
"""
sum_avod_cls_loss = eval_avod_losses[KEY_SUM_AVOD_CLS_LOSS]
sum_avod_reg_loss = eval_avod_losses[KEY_SUM_AVOD_REG_LOSS]
sum_avod_total_loss = eval_avod_losses[KEY_SUM_AVOD_TOTAL_LOSS]
sum_avod_loc_loss = eval_avod_losses[KEY_SUM_AVOD_LOC_LOSS]
sum_avod_ang_loss = eval_avod_losses[KEY_SUM_AVOD_ANG_LOSS]
sum_avod_cls_accuracy = \
eval_avod_losses[KEY_SUM_AVOD_CLS_ACC]
num_valid_regression_samples = \
eval_avod_losses[KEY_NUM_VALID_REG_SAMPLES]
avg_avod_cls_loss = sum_avod_cls_loss / num_valid_samples
avg_avod_reg_loss = sum_avod_reg_loss / num_valid_samples
avg_avod_total_loss = sum_avod_total_loss / num_valid_samples
if num_valid_regression_samples > 0:
avg_avod_loc_loss = \
sum_avod_loc_loss / num_valid_regression_samples
if box_rep in ['box_3d', 'box_4ca']:
avg_avod_ang_loss = \
sum_avod_ang_loss / num_valid_regression_samples
else:
avg_avod_loc_loss = 0
avg_avod_ang_loss = 0
avg_avod_cls_accuracy = sum_avod_cls_accuracy / num_valid_samples
# Write summaries
summary_utils.add_scalar_summary(
'avod_losses/classification/classification', avg_avod_cls_loss,
self.summary_writer, global_step)
summary_utils.add_scalar_summary(
'avod_losses/regression/regression_total', avg_avod_reg_loss,
self.summary_writer, global_step)
summary_utils.add_scalar_summary('avod_losses/regression/localization',
avg_avod_loc_loss,
self.summary_writer, global_step)
if box_rep in ['box_3d', 'box_4ca']:
summary_utils.add_scalar_summary(
'avod_losses/regression/orientation', avg_avod_ang_loss,
self.summary_writer, global_step)
print("Step {}: Average AVOD Losses: "
"cls {:.5f}, "
"reg {:.5f}, "
"total {:.5f} ".format(
global_step,
avg_avod_cls_loss,
avg_avod_reg_loss,
avg_avod_total_loss,
))
if box_rep in ['box_3d', 'box_4ca']:
print("Step {} Average AVOD Losses: "
"loc {:.5f} "
"ang {:.5f}".format(
global_step,
avg_avod_loc_loss,
avg_avod_ang_loss,
))
elif box_rep in ['box_8c', 'box_8co', 'box_4c']:
print("Step {} Average AVOD Losses: "
"loc {:.5f} ".format(global_step, avg_avod_loc_loss))
else:
raise NotImplementedError('Print average loss not implemented')
print("Step {}: Average Classification Accuracy: {} ".format(
global_step, avg_avod_cls_accuracy))
# Append to end of file
avg_loss_file_path = predictions_base_dir + '/avod_avg_losses.csv'
if box_rep in ['box_3d', 'box_4ca']:
with open(avg_loss_file_path, 'ba') as fp:
np.savetxt(fp, [
np.hstack([
global_step,
avg_avod_cls_loss,
avg_avod_reg_loss,
avg_avod_total_loss,
avg_avod_loc_loss,
avg_avod_ang_loss,
])
],
fmt='%d, %.5f, %.5f, %.5f, %.5f, %.5f')
elif box_rep in ['box_8c', 'box_8co', 'box_4c']:
with open(avg_loss_file_path, 'ba') as fp:
np.savetxt(fp, [
np.hstack([
global_step,
avg_avod_cls_loss,
avg_avod_reg_loss,
avg_avod_total_loss,
avg_avod_loc_loss,
])
],
fmt='%d, %.5f, %.5f, %.5f, %.5f')
else:
raise NotImplementedError('Saving losses not implemented')
avg_acc_file_path = predictions_base_dir + '/avod_avg_cls_acc.csv'
with open(avg_acc_file_path, 'ba') as fp:
np.savetxt(fp,
np.reshape([global_step, avg_avod_cls_accuracy],
(1, 2)),
fmt='%d, %.5f')
def save_prediction_losses_results(self, sum_losses, num_valid_samples,
global_step, predictions_base_dir):
"""Helper function to save the AVOD loss evaluation results.
Args:
eval_avod_losses: A dictionary containing the loss sums
num_valid_samples: An int, number of valid evaluated samples
i.e. samples with valid ground-truth.
global_step: Global step at which the metrics are computed.
predictions_base_dir: Base directory for storing the results.
box_rep: A string, the format of the 3D bounding box
one of 'box_3d', 'box_8c' etc.
"""
num_valid_reg_samples = max(1, sum_losses[KEY_NUM_VALID_REG_SAMPLES])
avg_losses_key_str = ''
avg_losses_print_str = ''
avg_losses_data = []
for sk, has_h, has_ang in zip(\
self.model_losses_keys, self.model_has_h_flags,\
self.model_has_angle_flags):
stage_losses = sum_losses[sk]
sum_cls_loss = stage_losses[KEY_SUM_CLS_LOSS]
avg_cls_loss = sum_cls_loss / num_valid_samples
# Write summaries
#f'{sk}_losses/classification/cls',
summary_utils.add_scalar_summary(
f'{sk}_loss/retinanet_losses/cls/norm/val', avg_cls_loss,
self.summary_writer, global_step)
sum_reg_loss = stage_losses[KEY_SUM_REG_LOSS]
avg_reg_loss = sum_reg_loss / num_valid_reg_samples
#f'{sk}_losses/regression/reg',
summary_utils.add_scalar_summary(
f'{sk}_loss/retinanet_losses/reg/norm/val', avg_reg_loss,
self.summary_writer, global_step)
avg_losses_print_str += f'[{sk}]cls:{avg_cls_loss}, reg:{avg_reg_loss}, '
avg_losses_key_str += f'{sk}_cls {sk}_reg '
avg_losses_data.extend([avg_cls_loss, avg_reg_loss])
if KEY_SUM_REG_H_LOSS in stage_losses:
sum_h_loss = stage_losses[KEY_SUM_REG_H_LOSS]
avg_h_loss = sum_h_loss / num_valid_reg_samples
#f'{sk}_losses/regression/h',
summary_utils.add_scalar_summary(
f'{sk}_loss/retinanet_losses/reg/norm/h_reg', avg_h_loss,
self.summary_writer, global_step)
avg_losses_print_str += f'h:{avg_h_loss}, '
avg_losses_key_str += f'{sk}_h '
avg_losses_data.append(avg_h_loss)
if KEY_SUM_REG_ANGLE_CLS_LOSS in stage_losses:
sum_angle_cls_loss = stage_losses[KEY_SUM_REG_ANGLE_CLS_LOSS]
avg_angle_cls_loss = sum_angle_cls_loss / num_valid_reg_samples
#f'{sk}_losses/regression/angle_cls',
summary_utils.add_scalar_summary(
f'{sk}_loss/retinanet_losses/reg/norm/angle_cls',
avg_angle_cls_loss, self.summary_writer, global_step)
avg_losses_print_str += f'angle:{avg_angle_cls_loss}, '
avg_losses_key_str += f'{sk}_angle '
avg_losses_data.append(avg_angle_cls_loss)
sum_cls_acc = sum_losses[KEY_SUM_CLS_ACC]
avg_cls_acc = sum_cls_acc / num_valid_samples
summary_utils.add_scalar_summary(f'output/cls_accuracy', avg_cls_acc,
self.summary_writer, global_step)
self.summary_writer.flush()
sum_total_loss = sum_losses[KEY_SUM_TOTAL_LOSS]
avg_total_loss = sum_total_loss / num_valid_samples
avg_losses_print_str += f'\ntotal:{avg_total_loss}'
avg_losses_key_str += 'total '
avg_losses_data.append(avg_total_loss)
# Append to end of file
avg_loss_file_dir = predictions_base_dir + '/avg_losses/'\
+ self.dataset_config.data_split
if not os.path.exists(avg_loss_file_dir):
os.makedirs(avg_loss_file_dir)
avg_loss_file_path = avg_loss_file_dir + '/avg_losses.csv'
if not os.path.exists(avg_loss_file_path):
with open(avg_loss_file_path, 'w') as f:
f.write(f'Step {avg_losses_key_str}\n')
save = True
if save:
print(
f"Step {global_step}: Average Losses: \n{avg_losses_print_str}"
)
with open(avg_loss_file_path, 'ba') as fp:
avg_losses_data = np.array(avg_losses_data)
np.savetxt(fp, [np.hstack([global_step, avg_losses_data])],
fmt='%d' + ', %.5f' * (len(avg_losses_data)))