def evaluate(create_input_dict_fn,
create_model_fn,
eval_config,
categories,
checkpoint_dir,
eval_dir,
graph_hook_fn=None,
evaluator_list=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.
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 = {}
result_dict, result_losses_dict = sess.run(
[tensor_dict, losses_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, result_losses_dict
if graph_hook_fn: graph_hook_fn()
variables_to_restore = tf.global_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)
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,
eval_export_path=eval_config.export_path)
return metrics
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_dicts = _extract_prediction_tensors(
model=model,
create_input_dict_fn=create_input_dict_fn,
ignore_groundtruth=eval_config.ignore_groundtruth)
def _process_batch(tensor_dicts,
sess,
batch_index,
counters,
losses_dict=None):
"""Evaluates tensors in tensor_dicts, visualizing the first K examples.
This function calls sess.run on tensor_dicts, evaluating the original_image
tensor only on the first K examples and visualizing detections overlaid
on this original_image.
Args:
tensor_dicts: 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. Necessary only
for matching function signiture in third_party eval_util.py.
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. Necessary only for matching function signiture in
third_party eval_util.py.
"""
if batch_index % 10 == 0:
logging.info('Running eval ops batch %d', batch_index)
if not losses_dict:
losses_dict = {}
try:
result_dicts, result_losses_dict = sess.run([tensor_dicts, losses_dict])
counters['success'] += 1
except tf.errors.InvalidArgumentError:
logging.info('Skipping image')
counters['skipped'] += 1
return {}
num_images = len(tensor_dicts)
for i in range(num_images):
result_dict = result_dicts[i]
global_step = tf.train.global_step(sess, tf.train.get_global_step())
tag = 'image-%d' % (batch_index * num_images + i)
if batch_index < eval_config.num_visualizations / num_images:
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)
if num_images > 1:
return result_dicts, result_losses_dict
else:
return result_dicts[0], 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()
for key in variables_to_restore.keys():
if 'moving_mean' in key:
variables_to_restore[key.replace(
'moving_mean', 'moving_mean/ExponentialMovingAverage')] = (
variables_to_restore[key])
del variables_to_restore[key]
if 'moving_variance' in key:
variables_to_restore[key.replace(
'moving_variance', 'moving_variance/ExponentialMovingAverage')] = (
variables_to_restore[key])
del variables_to_restore[key]
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_dicts,
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, 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, 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,
categories,
checkpoint_dir,
eval_dir,
graph_hook_fn=None,
evaluator_list=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.
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()
#hotspot = 0
#not_hotspot = 0
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 = {}
'''
testing
'''
# # tensor_dict = [n.name for n in tf.get_default_graph().as_graph_def().node]
# test_layer = tf.get_default_graph().get_tensor_by_name('SecondStageFeatureExtractor/InceptionV2/Mixed_5b/Branch_1/Conv2d_0a_1x1/Relu:0')
# test_layer_dict = {"test_layer":test_layer}
# feature_dict, result_dict, result_losses_dict = sess.run([test_layer, tensor_dict, losses_dict])
# # feature_dict = sess.run([test_layer, tensor_dict])
# import numpy as np
# import scipy.misc
# feature_tensor = feature_dict[0]
# origin_image = result_dict['original_image']
# feature_map = np.sum(feature_tensor,axis=(2))
# # feature_map /= feature_map.max()
# # feature_map = 1 - feature_map
# # rgb_feature_map = np.zeros((256, 256, 3))
# # rgb_feature_map[..., 0] = feature_map * 30
# # rgb_feature_map[..., 1] = feature_map * 144
# # rgb_feature_map[..., 2] = feature_map * 255
# # rgb_feature_map[..., 0] = np.where(feature_map==0, 30, 255)
# # rgb_feature_map[..., 1] = np.where(feature_map==0, 144, 255)
# # rgb_feature_map[..., 2] = np.where(feature_map==0, 255, 255)
# origin_image = origin_image[0]
# # origin_image[..., 0] = np.where(origin_image[..., 0]==0, 30, 255)
# # origin_image[..., 1] = np.where(origin_image[..., 1]==0, 144, 255)
# # origin_image[..., 2] = np.where(origin_image[..., 2]==0, 255, 255)
# scipy.misc.imsave('feature_map_{}.jpg'.format(counters['success']), feature_map)
# # scipy.misc.imsave('origin_image_{}.jpg'.format(counters['success']), origin_image)
# # from IPython import embed;embed()
''''''
result_dict, result_losses_dict = sess.run(
[tensor_dict, losses_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(
eval_config.num_visualizations,
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,
num_examples=eval_config.num_examples,
keep_image_id_for_visualization_export=eval_config.
keep_image_id_for_visualization_export)
return result_dict, result_losses_dict
if graph_hook_fn: graph_hook_fn()
variables_to_restore = tf.global_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)
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)
print("metrics:", metrics)
return metrics