print(
" x=%.4f y=%.4f z=%.4f l=%.4f h=%.4f w=%.4f y=%.4f" %
(results['loc_loss_cls_%d_box_%d' % (class_idx, 0)],
results['loc_loss_cls_%d_box_%d' %
(class_idx, 1)], results['loc_loss_cls_%d_box_%d' %
(class_idx, 2)],
results['loc_loss_cls_%d_box_%d' %
(class_idx, 3)], results['loc_loss_cls_%d_box_%d' %
(class_idx, 4)],
results['loc_loss_cls_%d_box_%d' %
(class_idx, 5)], results['loc_loss_cls_%d_box_%d' %
(class_idx, 6)]), )
# add summaries ====================================================
for key in metrics_update_ops:
write_summary_scale(key, results[key], results['step'],
train_config['train_dir'])
write_summary_scale('learning rate', results['learning_rate'],
results['step'], train_config['train_dir'])
# save checkpoint ==================================================
if (epoch_idx + 1) % train_config['save_every_epoch'] == 0:
checkpoint_path = os.path.join(train_config['train_dir'],
train_config['checkpoint_path'])
config_path = os.path.join(train_config['train_dir'],
train_config['config_path'])
train_config_path = os.path.join(train_config['train_dir'],
'train_config')
print('save checkpoint at step %d to %s' %
(epoch_idx, checkpoint_path))
saver.save(sess,
checkpoint_path,
latest_filename='checkpoint',
def eval_once(graph, gpu_options, saver, checkpoint_path):
"""Evaluate the model once. """
with tf.Session(graph=graph,
config=tf.ConfigProto(gpu_options=gpu_options)) as sess:
sess.run(tf.variables_initializer(tf.global_variables()))
sess.run(tf.variables_initializer(tf.local_variables()))
print('Restore from checkpoint %s' % checkpoint_path)
saver.restore(sess, checkpoint_path)
previous_step = sess.run(global_step)
print('Global step = %d' % previous_step)
start_time = time.time()
print("NUM_TEST_SAMPLE", NUM_TEST_SAMPLE)
for frame_idx in range(NUM_TEST_SAMPLE):
(input_v, vertex_coord_list, keypoint_indices_list, edges_list,
cls_labels, encoded_boxes, valid_boxes)\
= data_provider.provide(frame_idx)
feed_dict = {
t_initial_vertex_features: input_v,
t_class_labels: cls_labels,
t_encoded_gt_boxes: encoded_boxes,
t_valid_gt_boxes: valid_boxes,
t_is_training: config['eval_is_training'],
}
feed_dict.update(dict(zip(t_edges_list, edges_list)))
feed_dict.update(
dict(zip(t_keypoint_indices_list, keypoint_indices_list)))
feed_dict.update(dict(zip(t_vertex_coord_list, vertex_coord_list)))
results = sess.run(fetches, feed_dict=feed_dict)
if NUM_TEST_SAMPLE >= 10:
if (frame_idx + 1) % (NUM_TEST_SAMPLE // 10) == 0:
print('@frame %d' % frame_idx)
print('cls:%f, loc:%f, reg:%f, loss: %f' %
(results['cls_loss'], results['loc_loss'],
results['reg_loss'], results['total_loss']))
for class_idx in range(NUM_CLASSES):
print('Class_%d: recall=%f, prec=%f, mAP=%f, loc=%f' %
(class_idx, results['recall_%d' % class_idx],
results['precision_%d' % class_idx],
results['mAP_%d' % class_idx],
results['loc_loss_cls_%d' % class_idx]))
print(' '+\
'x=%.4f y=%.4f z=%.4f l=%.4f h=%.4f w=%.4f y=%.4f'
%(
results['loc_loss_cls_%d_box_%d'%(class_idx, 0)],
results['loc_loss_cls_%d_box_%d'%(class_idx, 1)],
results['loc_loss_cls_%d_box_%d'%(class_idx, 2)],
results['loc_loss_cls_%d_box_%d'%(class_idx, 3)],
results['loc_loss_cls_%d_box_%d'%(class_idx, 4)],
results['loc_loss_cls_%d_box_%d'%(class_idx, 5)],
results['loc_loss_cls_%d_box_%d'%(class_idx, 6)]),
)
print('STEP: %d, time cost: %f' %
(results['step'], time.time() - start_time))
print('cls:%f, loc:%f, reg:%f, loss: %f' %
(results['cls_loss'], results['loc_loss'], results['reg_loss'],
results['total_loss']))
for class_idx in range(NUM_CLASSES):
print('Class_%d: recall=%f, prec=%f, mAP=%f, loc=%f' %
(class_idx, results['recall_%d' % class_idx],
results['precision_%d' % class_idx],
results['mAP_%d' % class_idx],
results['loc_loss_cls_%d' % class_idx]))
print(
" x=%.4f y=%.4f z=%.4f l=%.4f h=%.4f w=%.4f y=%.4f" %
(results['loc_loss_cls_%d_box_%d' % (class_idx, 0)],
results['loc_loss_cls_%d_box_%d' %
(class_idx, 1)], results['loc_loss_cls_%d_box_%d' %
(class_idx, 2)],
results['loc_loss_cls_%d_box_%d' %
(class_idx, 3)], results['loc_loss_cls_%d_box_%d' %
(class_idx, 4)],
results['loc_loss_cls_%d_box_%d' %
(class_idx, 5)], results['loc_loss_cls_%d_box_%d' %
(class_idx, 6)]), )
# add summaries ====================================================
for key in metrics_update_ops:
write_summary_scale(key, results[key], results['step'],
eval_config['eval_dir'])
return results['step']