def summarize_video_metrics(hook_args):
"""Computes video metrics summaries using the decoder output."""
problem_name = hook_args.problem.name
current_problem = hook_args.problem
hparams = hook_args.hparams
output_dirs = hook_args.output_dirs
predictions = hook_args.predictions
frame_shape = [
current_problem.frame_height, current_problem.frame_width,
current_problem.num_channels
]
metrics_graph = tf.Graph()
with metrics_graph.as_default():
if predictions:
metrics_results = video_metrics.compute_video_metrics_from_predictions(
predictions)
else:
metrics_results, _ = video_metrics.compute_video_metrics_from_png_files(
output_dirs, problem_name, hparams.video_num_target_frames,
frame_shape)
summary_values = []
for name, array in six.iteritems(metrics_results):
for ind, val in enumerate(array):
tag = "metric_{}/{}".format(name, ind)
summary_values.append(tf.Summary.Value(tag=tag, simple_value=val))
return summary_values
def summarize_video_metrics(hook_args):
"""Computes video metrics summaries using the decoder output."""
problem_name = hook_args.problem.name
current_problem = hook_args.problem
hparams = hook_args.hparams
output_dirs = hook_args.output_dirs
predictions = hook_args.predictions
frame_shape = [
current_problem.frame_height, current_problem.frame_width,
current_problem.num_channels
]
metrics_graph = tf.Graph()
with metrics_graph.as_default():
if predictions:
metrics_results = video_metrics.compute_video_metrics_from_predictions(
predictions)
else:
metrics_results, _ = video_metrics.compute_video_metrics(
output_dirs, problem_name,
hparams.video_num_target_frames, frame_shape)
summary_values = []
for name, array in six.iteritems(metrics_results):
for ind, val in enumerate(array):
tag = name + "_" + str(ind)
summary_values.append(tf.Summary.Value(tag=tag, simple_value=val))
return summary_values
def display_video_hooks(hook_args):
"""Hooks to display videos at decode time."""
predictions = hook_args.predictions
max_outputs = hook_args.decode_hparams.max_display_outputs
max_decodes = hook_args.decode_hparams.max_display_decodes
with tf.Graph().as_default():
_, best_decodes = video_metrics.compute_video_metrics_from_predictions(
predictions, decode_hparams=hook_args.decode_hparams)
all_summaries = []
# Displays decodes corresponding to the best/worst metric,
for metric, metric_decode_inds in best_decodes.items():
curr_metric_inds = metric_decode_inds[:max_outputs]
best_inputs, best_outputs, best_targets = [], [], []
for sample_ind, decode_ind in enumerate(curr_metric_inds):
curr_decode = predictions[decode_ind][sample_ind]
best_inputs.append(curr_decode["inputs"])
best_outputs.append(curr_decode["outputs"])
best_targets.append(curr_decode["targets"])
best_inputs = np.array(best_inputs, dtype=np.uint8)
best_outputs = np.array(best_outputs, dtype=np.uint8)
best_targets = np.array(best_targets, dtype=np.uint8)
summaries = convert_videos_to_summaries(
best_inputs,
best_outputs,
best_targets,
tag=metric,
decode_hparams=hook_args.decode_hparams)
all_summaries.extend(summaries)
# Display random decodes for ten conditioning frames.
for decode_ind, decode in enumerate(predictions[:max_decodes]):
target_videos = video_metrics.stack_data_given_key(decode, "targets")
output_videos = video_metrics.stack_data_given_key(decode, "outputs")
input_videos = video_metrics.stack_data_given_key(decode, "inputs")
target_videos = np.asarray(target_videos, dtype=np.uint8)
output_videos = np.asarray(output_videos, dtype=np.uint8)
input_videos = np.asarray(input_videos, dtype=np.uint8)
summaries = convert_videos_to_summaries(
input_videos,
output_videos,
target_videos,
tag="decode_%d" % decode_ind,
decode_hparams=hook_args.decode_hparams,
display_ground_truth=decode_ind == 0)
all_summaries.extend(summaries)
return all_summaries
