def evaluate(create_input_dict_fn,
create_model_fn,
eval_config,
categories,
checkpoint_dir,
eval_dir,
graph_hook_fn=None,
evaluator_list=None,
intra_op=0,
inter_op=0):
"""Evaluation function for detection models.
Args:
create_input_dict_fn: a function to create a tensor input dictionary.
create_model_fn: a function that creates a DetectionModel.
eval_config: a eval_pb2.EvalConfig protobuf.
categories: a list of category dictionaries. Each dict in the list should
have an integer 'id' field and string 'name' field.
checkpoint_dir: directory to load the checkpoints to evaluate from.
eval_dir: directory to write evaluation metrics summary to.
graph_hook_fn: Optional function that is called after the training graph is
completely built. This is helpful to perform additional changes to the
training graph such as optimizing batchnorm. The function should modify
the default graph.
evaluator_list: Optional list of instances of DetectionEvaluator. If not
given, this list of metrics is created according to the eval_config.
Returns:
metrics: A dictionary containing metric names and values from the latest
run.
"""
model = create_model_fn()
if eval_config.ignore_groundtruth and not eval_config.export_path:
logging.fatal('If ignore_groundtruth=True then an export_path is '
'required. Aborting!!!')
tensor_dict, losses_dict = _extract_predictions_and_losses(
model=model,
create_input_dict_fn=create_input_dict_fn,
ignore_groundtruth=eval_config.ignore_groundtruth)
def _process_batch(tensor_dict,
sess,
batch_index,
counters,
losses_dict=None):
"""Evaluates tensors in tensor_dict, losses_dict and visualizes examples.
This function calls sess.run on tensor_dict, evaluating the original_image
tensor only on the first K examples and visualizing detections overlaid
on this original_image.
Args:
tensor_dict: a dictionary of tensors
sess: tensorflow session
batch_index: the index of the batch amongst all batches in the run.
counters: a dictionary holding 'success' and 'skipped' fields which can
be updated to keep track of number of successful and failed runs,
respectively. If these fields are not updated, then the success/skipped
counter values shown at the end of evaluation will be incorrect.
losses_dict: Optional dictonary of scalar loss tensors.
Returns:
result_dict: a dictionary of numpy arrays
result_losses_dict: a dictionary of scalar losses. This is empty if input
losses_dict is None.
"""
try:
if not losses_dict:
losses_dict = {}
trace = False
if batch_index == 0 and trace:
run_options = tf.RunOptions(
trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
else:
run_options = None
run_metadata = None
start_time = time.time()
result_dict, result_losses_dict = sess.run(
[tensor_dict, losses_dict],
options=run_options,
run_metadata=run_metadata)
if (batch_index % 100 == 0):
logging.info('Step %d: %.3f sec', batch_index,
time.time() - start_time)
if batch_index == 0 and trace:
trace = timeline.Timeline(step_stats=run_metadata.step_stats)
dir = 'logs'
if not os.path.exists(dir):
os.makedirs(dir)
with open(
dir + '/rfcn-timeline-' +
time.strftime("%Y%m%d-%H%M%S") + '.json', 'w') as file:
file.write(
trace.generate_chrome_trace_format(show_memory=False))
counters['success'] += 1
except tf.errors.InvalidArgumentError:
logging.info('Skipping image')
counters['skipped'] += 1
return {}, {}
global_step = tf.train.global_step(sess, tf.train.get_global_step())
if batch_index < eval_config.num_visualizations:
tag = 'image-{}'.format(batch_index)
eval_util.visualize_detection_results(
result_dict,
tag,
global_step,
categories=categories,
summary_dir=eval_dir,
export_dir=eval_config.visualization_export_dir,
show_groundtruth=eval_config.visualize_groundtruth_boxes,
groundtruth_box_visualization_color=eval_config.
groundtruth_box_visualization_color,
min_score_thresh=eval_config.min_score_threshold,
max_num_predictions=eval_config.max_num_boxes_to_visualize,
skip_scores=eval_config.skip_scores,
skip_labels=eval_config.skip_labels,
keep_image_id_for_visualization_export=eval_config.
keep_image_id_for_visualization_export)
return result_dict, result_losses_dict
variables_to_restore = tf.global_variables()
global_step = tf.train.get_or_create_global_step()
variables_to_restore.append(global_step)
if graph_hook_fn:
graph_hook_fn()
if eval_config.use_moving_averages:
variable_averages = tf.train.ExponentialMovingAverage(0.0)
variables_to_restore = variable_averages.variables_to_restore()
saver = tf.train.Saver(variables_to_restore)
def _restore_latest_checkpoint(sess):
latest_checkpoint = tf.train.latest_checkpoint(checkpoint_dir)
saver.restore(sess, latest_checkpoint)
if not evaluator_list:
evaluator_list = get_evaluators(eval_config, categories)
metrics = eval_util.repeated_checkpoint_run(
tensor_dict=tensor_dict,
summary_dir=eval_dir,
evaluators=evaluator_list,
batch_processor=_process_batch,
checkpoint_dirs=[checkpoint_dir],
variables_to_restore=None,
restore_fn=_restore_latest_checkpoint,
num_batches=eval_config.num_examples,
eval_interval_secs=eval_config.eval_interval_secs,
max_number_of_evaluations=(
1 if eval_config.ignore_groundtruth else
eval_config.max_evals if eval_config.max_evals else None),
master=eval_config.eval_master,
save_graph=eval_config.save_graph,
save_graph_dir=(eval_dir if eval_config.save_graph else ''),
losses_dict=losses_dict)
return metrics
def evaluate(create_input_dict_fn, create_model_fn, eval_config, categories,
checkpoint_dir, eval_dir):
"""Evaluation function for detection models.
Args:
create_input_dict_fn: a function to create a tensor input dictionary.
create_model_fn: a function that creates a DetectionModel.
eval_config: a eval_pb2.EvalConfig protobuf.
categories: a list of category dictionaries. Each dict in the list should
have an integer 'id' field and string 'name' field.
checkpoint_dir: directory to load the checkpoints to evaluate from.
eval_dir: directory to write evaluation metrics summary to.
"""
model = create_model_fn()
if eval_config.ignore_groundtruth and not eval_config.export_path:
logging.fatal('If ignore_groundtruth=True then an export_path is '
'required. Aborting!!!')
tensor_dict = _extract_prediction_tensors(
model=model,
create_input_dict_fn=create_input_dict_fn,
ignore_groundtruth=eval_config.ignore_groundtruth)
def _process_batch(tensor_dict, sess, batch_index, counters, update_op):
"""Evaluates tensors in tensor_dict, visualizing the first K examples.
This function calls sess.run on tensor_dict, evaluating the original_image
tensor only on the first K examples and visualizing detections overlaid
on this original_image.
Args:
tensor_dict: a dictionary of tensors
sess: tensorflow session
batch_index: the index of the batch amongst all batches in the run.
counters: a dictionary holding 'success' and 'skipped' fields which can
be updated to keep track of number of successful and failed runs,
respectively. If these fields are not updated, then the success/skipped
counter values shown at the end of evaluation will be incorrect.
update_op: An update op that has to be run along with output tensors. For
example this could be an op to compute statistics for slim metrics.
Returns:
result_dict: a dictionary of numpy arrays
"""
if batch_index >= eval_config.num_visualizations:
if 'original_image' in tensor_dict:
tensor_dict = {
k: v
for (k, v) in tensor_dict.items() if k != 'original_image'
}
try:
(result_dict, _) = sess.run([tensor_dict, update_op])
counters['success'] += 1
except tf.errors.InvalidArgumentError:
logging.info('Skipping image')
counters['skipped'] += 1
return {}
global_step = tf.train.global_step(sess, slim.get_global_step())
if batch_index < eval_config.num_visualizations:
tag = 'image-{}'.format(batch_index)
eval_util.visualize_detection_results(
result_dict,
tag,
global_step,
categories=categories,
summary_dir=eval_dir,
export_dir=eval_config.visualization_export_dir,
show_groundtruth=eval_config.visualization_export_dir)
return result_dict
def _process_aggregated_results(result_lists):
eval_metric_fn_key = eval_config.metrics_set
if eval_metric_fn_key not in EVAL_METRICS_FN_DICT:
raise ValueError('Metric not found: {}'.format(eval_metric_fn_key))
return EVAL_METRICS_FN_DICT[eval_metric_fn_key](result_lists,
categories=categories)
variables_to_restore = tf.global_variables()
global_step = slim.get_or_create_global_step()
variables_to_restore.append(global_step)
if eval_config.use_moving_averages:
variable_averages = tf.train.ExponentialMovingAverage(0.0)
variables_to_restore = variable_averages.variables_to_restore()
saver = tf.train.Saver(variables_to_restore)
def _restore_latest_checkpoint(sess):
latest_checkpoint = tf.train.latest_checkpoint(checkpoint_dir)
saver.restore(sess, latest_checkpoint)
eval_util.repeated_checkpoint_run(
tensor_dict=tensor_dict,
update_op=tf.no_op(),
summary_dir=eval_dir,
aggregated_result_processor=_process_aggregated_results,
batch_processor=_process_batch,
checkpoint_dirs=[checkpoint_dir],
variables_to_restore=None,
restore_fn=_restore_latest_checkpoint,
num_batches=eval_config.num_examples,
eval_interval_secs=eval_config.eval_interval_secs,
max_number_of_evaluations=(
1 if eval_config.ignore_groundtruth else
eval_config.max_evals if eval_config.max_evals else None),
master=eval_config.eval_master,
save_graph=eval_config.save_graph,
save_graph_dir=(eval_dir if eval_config.save_graph else ''))
def evaluate(create_input_dict_fn, create_model_fn, eval_config, input_config, categories,
checkpoint_dir, eval_dir, run_mode='latest', is_save_detection_results=False, detection_results_name=''):
"""Evaluation function for detection models.
Args:
create_input_dict_fn: a function to create a tensor input dictionary.
create_model_fn: a function that creates a DetectionModel.
eval_config: a eval_pb2.EvalConfig protobuf.
input_config: a input_reader.InputReader protobuf.
categories: a list of category dictionaries. Each dict in the list should
have an integer 'id' field and string 'name' field.
checkpoint_dir: directory to load the checkpoints to evaluate from.
eval_dir: directory to write evaluation metrics summary to.
run_mode: when run_mode is latest, it run infinite and the latest checkpoint is evaluated.
when run_mode is all, all checkpoints are evaluated and finish evaluation)
is_save_detection_results: whether or not to save detection results.
detection_results_name: a filename to a detection_results pickle file.
"""
model = create_model_fn()
if eval_config.ignore_groundtruth and not eval_config.export_path:
logging.fatal('If ignore_groundtruth=True then an export_path is required. Aborting!!!')
preprocess_input_options = [preprocessor_input_builder.build(step)
for step in input_config.preprocess_input_options]
tensor_dict = _extract_prediction_tensors(model=model,
create_input_dict_fn=create_input_dict_fn,
ignore_groundtruth=eval_config.ignore_groundtruth,
preprocess_input_options=preprocess_input_options)
def _process_batch(tensor_dict, sess, batch_index, counters, update_op):
"""Evaluates tensors in tensor_dict, visualizing the first K examples.
This function calls sess.run on tensor_dict, evaluating the original_image
tensor only on the first K examples and visualizing detections overlaid
on this original_image.
Args:
tensor_dict: a dictionary of tensors
sess: tensorflow session
batch_index: the index of the batch amongst all batches in the run.
counters: a dictionary holding 'success' and 'skipped' fields which can
be updated to keep track of number of successful and failed runs,
respectively. If these fields are not updated, then the success/skipped
counter values shown at the end of evaluation will be incorrect.
update_op: An update op that has to be run along with output tensors. For
example this could be an op to compute statistics for slim metrics.
Returns:
result_dict: a dictionary of numpy arrays
"""
if batch_index >= eval_config.num_visualizations and not is_save_detection_results:
if 'original_image' in tensor_dict:
tensor_dict = {k: v for (k, v) in tensor_dict.items()
if k != 'original_image'}
try:
(result_dict, _) = sess.run([tensor_dict, update_op])
counters['success'] += 1
except tf.errors.InvalidArgumentError:
logging.info('Skipping image')
counters['skipped'] += 1
return {}
global_step = tf.train.global_step(sess, slim.get_global_step())
if batch_index < eval_config.num_visualizations:
tag = 'image-{}'.format(batch_index)
eval_util.visualize_detection_results(
result_dict, tag, global_step, categories=categories,
summary_dir=eval_dir,
export_dir=eval_config.visualization_export_dir)
return result_dict
def _process_aggregated_results(result_lists):
eval_metric_fn_key = eval_config.metrics_set
if eval_metric_fn_key not in EVAL_METRICS_FN_DICT:
raise ValueError('Metric not found: {}'.format(eval_metric_fn_key))
return EVAL_METRICS_FN_DICT[eval_metric_fn_key](result_lists,
categories=categories,
box_type='rbox' if model.is_rbbox else 'box',
is_save_detection_results=is_save_detection_results,
detection_results_name=detection_results_name)
variables_to_restore = tf.global_variables()
global_step = slim.get_or_create_global_step()
variables_to_restore.append(global_step)
if eval_config.use_moving_averages:
variable_averages = tf.train.ExponentialMovingAverage(0.0)
variables_to_restore = variable_averages.variables_to_restore()
saver = tf.train.Saver(variables_to_restore)
def _restore_latest_checkpoint(sess):
latest_checkpoint = tf.train.latest_checkpoint(checkpoint_dir)
saver.restore(sess, latest_checkpoint)
class _Restorer(object):
def __init__(self, checkpoint_dir, run_mode):
ckpt_list = glob(checkpoint_dir + '/*.index')
ckpt_list = [ckpt[:-6] for ckpt in ckpt_list]
sort_nicely(ckpt_list)
if run_mode == 'latest_once':
self.ckpt_list = ckpt_list[-1:]
else: # run_mode == 'all':
self.ckpt_list = ckpt_list[1:]
self.n_ckpt = len(self.ckpt_list)
self.cnt = 0
def __call__(self, sess):
saver.restore(sess, self.ckpt_list[self.cnt])
self.cnt += 1
def all_checked(self):
if self.cnt == self.n_ckpt:
return True
else:
return False
if run_mode == 'all' or run_mode == 'latest_once':
eval_util.all_checkpoint_run(
tensor_dict=tensor_dict,
update_op=tf.no_op(),
summary_dir=eval_dir,
aggregated_result_processor=_process_aggregated_results,
batch_processor=_process_batch,
checkpoint_dirs=[checkpoint_dir],
variables_to_restore=None,
restorer=_Restorer(checkpoint_dir, run_mode),
num_batches=eval_config.num_examples,
master=eval_config.eval_master,
save_graph=eval_config.save_graph,
save_graph_dir=(eval_dir if eval_config.save_graph else ''))
else: # run_mode == 'latest'
eval_util.repeated_checkpoint_run(
tensor_dict=tensor_dict,
update_op=tf.no_op(),
summary_dir=eval_dir,
aggregated_result_processor=_process_aggregated_results,
batch_processor=_process_batch,
checkpoint_dirs=[checkpoint_dir],
variables_to_restore=None,
restore_fn=_restore_latest_checkpoint,
num_batches=eval_config.num_examples,
eval_interval_secs=eval_config.eval_interval_secs,
max_number_of_evaluations=(
1 if eval_config.ignore_groundtruth else
eval_config.max_evals if eval_config.max_evals else
None),
master=eval_config.eval_master,
save_graph=eval_config.save_graph,
save_graph_dir=(eval_dir if eval_config.save_graph else ''))
def evaluate(create_input_dict_fn, create_model_fn, eval_config, categories,
checkpoint_dir, eval_dir, graph_hook_fn=None):
"""Evaluation function for detection models.
Args:
create_input_dict_fn: a function to create a tensor input dictionary.
create_model_fn: a function that creates a DetectionModel.
eval_config: a eval_pb2.EvalConfig protobuf.
categories: a list of category dictionaries. Each dict in the list should
have an integer 'id' field and string 'name' field.
checkpoint_dir: directory to load the checkpoints to evaluate from.
eval_dir: directory to write evaluation metrics summary to.
graph_hook_fn: Optional function that is called after the training graph is
completely built. This is helpful to perform additional changes to the
training graph such as optimizing batchnorm. The function should modify
the default graph.
Returns:
metrics: A dictionary containing metric names and values from the latest
run.
"""
model = create_model_fn()
if eval_config.ignore_groundtruth and not eval_config.export_path:
logging.fatal('If ignore_groundtruth=True then an export_path is '
'required. Aborting!!!')
tensor_dict = _extract_prediction_tensors(
model=model,
create_input_dict_fn=create_input_dict_fn,
ignore_groundtruth=eval_config.ignore_groundtruth)
def _process_batch(tensor_dict, sess, batch_index, counters):
"""Evaluates tensors in tensor_dict, visualizing the first K examples.
This function calls sess.run on tensor_dict, evaluating the original_image
tensor only on the first K examples and visualizing detections overlaid
on this original_image.
Args:
tensor_dict: a dictionary of tensors
sess: tensorflow session
batch_index: the index of the batch amongst all batches in the run.
counters: a dictionary holding 'success' and 'skipped' fields which can
be updated to keep track of number of successful and failed runs,
respectively. If these fields are not updated, then the success/skipped
counter values shown at the end of evaluation will be incorrect.
Returns:
result_dict: a dictionary of numpy arrays
"""
try:
result_dict = sess.run(tensor_dict)
counters['success'] += 1
except tf.errors.InvalidArgumentError:
logging.info('Skipping image')
counters['skipped'] += 1
return {}
global_step = tf.train.global_step(sess, tf.train.get_global_step())
if batch_index < eval_config.num_visualizations:
tag = 'image-{}'.format(batch_index)
eval_util.visualize_detection_results(
result_dict,
tag,
global_step,
categories=categories,
summary_dir=eval_dir,
export_dir=eval_config.visualization_export_dir,
show_groundtruth=eval_config.visualize_groundtruth_boxes,
groundtruth_box_visualization_color=eval_config.
groundtruth_box_visualization_color,
min_score_thresh=eval_config.min_score_threshold,
max_num_predictions=eval_config.max_num_boxes_to_visualize,
skip_scores=eval_config.skip_scores,
skip_labels=eval_config.skip_labels,
keep_image_id_for_visualization_export=eval_config.
keep_image_id_for_visualization_export)
return result_dict
variables_to_restore = tf.global_variables()
global_step = tf.train.get_or_create_global_step()
variables_to_restore.append(global_step)
if graph_hook_fn: graph_hook_fn()
if eval_config.use_moving_averages:
variable_averages = tf.train.ExponentialMovingAverage(0.0)
variables_to_restore = variable_averages.variables_to_restore()
saver = tf.train.Saver(variables_to_restore)
def _restore_latest_checkpoint(sess):
latest_checkpoint = tf.train.latest_checkpoint(checkpoint_dir)
saver.restore(sess, latest_checkpoint)
metrics = eval_util.repeated_checkpoint_run(
tensor_dict=tensor_dict,
summary_dir=eval_dir,
evaluators=get_evaluators(eval_config, categories),
batch_processor=_process_batch,
checkpoint_dirs=[checkpoint_dir],
variables_to_restore=None,
restore_fn=_restore_latest_checkpoint,
num_batches=eval_config.num_examples,
eval_interval_secs=eval_config.eval_interval_secs,
max_number_of_evaluations=(1 if eval_config.ignore_groundtruth else
eval_config.max_evals
if eval_config.max_evals else None),
master=eval_config.eval_master,
save_graph=eval_config.save_graph,
save_graph_dir=(eval_dir if eval_config.save_graph else ''))
return metrics
def evaluate(create_input_dict_fn, create_model_fn, eval_config,
checkpoint_dir, eval_dir,
repeat_evaluation=True):
model = create_model_fn()
data_preprocessing_steps = [
preprocessor_builder.build(step)
for step in eval_config.data_preprocessing_steps]
tensor_dict = _extract_prediction_tensors(
model=model,
create_input_dict_fn=create_input_dict_fn,
data_preprocessing_steps=data_preprocessing_steps,
ignore_groundtruth=eval_config.ignore_groundtruth,
evaluate_with_lexicon=eval_config.eval_with_lexicon)
summary_writer = tf.summary.FileWriter(eval_dir)
def _process_batch(tensor_dict, sess, batch_index, counters, update_op):
if batch_index >= eval_config.num_visualizations:
if 'original_image' in tensor_dict:
tensor_dict = {k: v for (k, v) in tensor_dict.items()
if k != 'original_image'}
try:
(result_dict, _, glyphs) = sess.run([tensor_dict, update_op, tf.get_collection('glyph')])
counters['success'] += 1
except tf.errors.InvalidArgumentError:
logging.info('Skipping image')
counters['skipped'] += 1
return {}
global_step = tf.train.global_step(sess, tf.train.get_global_step())
if batch_index < eval_config.num_visualizations:
eval_util.visualize_recognition_results(
result_dict,
'Recognition_{}'.format(batch_index),
global_step,
summary_dir=eval_dir,
export_dir=os.path.join(eval_dir, 'vis'),
summary_writer=summary_writer,
only_visualize_incorrect=eval_config.only_visualize_incorrect)
return result_dict
def _process_aggregated_results(result_lists):
eval_metric_fn_key = eval_config.metrics_set
if eval_metric_fn_key not in EVAL_METRICS_FN_DICT:
raise ValueError('Metric not found: {}'.format(eval_metric_fn_key))
return EVAL_METRICS_FN_DICT[eval_metric_fn_key](result_lists)
variables_to_restore = tf.global_variables()
#variables_to_restore = tf.trainable_variables()
global_step = tf.train.get_or_create_global_step()
variables_to_restore.append(global_step)
if eval_config.use_moving_averages:
variable_averages = tf.train.ExponentialMovingAverage(0.0)
variables_to_restore = variable_averages.variables_to_restore()
saver = tf.train.Saver(variables_to_restore)
def _restore_latest_checkpoint(sess):
latest_checkpoint = tf.train.latest_checkpoint(checkpoint_dir)
saver.restore(sess, latest_checkpoint)
eval_util.repeated_checkpoint_run(
tensor_dict=tensor_dict,
update_op=tf.no_op(),
summary_dir=eval_dir,
aggregated_result_processor=_process_aggregated_results,
batch_processor=_process_batch,
checkpoint_dirs=[checkpoint_dir],
variables_to_restore=None,
restore_fn=_restore_latest_checkpoint,
num_batches=eval_config.num_examples,
eval_interval_secs=eval_config.eval_interval_secs,
max_number_of_evaluations=(
1 if eval_config.ignore_groundtruth else
eval_config.max_evals if eval_config.max_evals else
None if repeat_evaluation else 1),
master=eval_config.eval_master,
save_graph=eval_config.save_graph,
save_graph_dir=(eval_dir if eval_config.save_graph else ''))
summary_writer.close()