def transform_csv(self, to_transform_path):
"""
Main transformation method
:param to_transform_path:
:return:
"""
self.reset()
trace_video_pair_name = to_transform_path.split('/')[-2]
client_logger_file = to_transform_path.replace(
'raw_dataframe.csv', 'local_client_state_logger.csv')
client_logger_file = pd.read_csv(client_logger_file, index_col=0)
client_logger_file['buffer_level'] = client_logger_file[
'buffered_until'] - client_logger_file['played_until']
client_logger_file['time_elapsed'] = pd.to_timedelta(
client_logger_file.timestamp_s -
client_logger_file.timestamp_s.iloc[0],
unit='s')
client_logger_file['fps'] = client_logger_file.decodedFrames.diff()
decoded_frame_unit = client_logger_file[['time_elapsed', 'fps']]
client_logger_file = client_logger_file.drop('fps', axis=1)
decoded_frame_unit = decoded_frame_unit.set_index('time_elapsed')
decoded_frame_unit = decoded_frame_unit.resample(
'%ds' % self.fps_smoothing_s).sum() / float(self.fps_smoothing_s)
decoded_frame_unit['fps'] = decoded_frame_unit['fps'].map(
self.map_experimental_fps_avail_fps)
decoded_frame_unit = decoded_frame_unit.reset_index()
client_logger_file = pd.merge_asof(client_logger_file,
decoded_frame_unit,
on='time_elapsed',
direction='nearest')
client_logger_file['current_level'] = (
client_logger_file['videoWidth'] *
client_logger_file['videoHeight'] * client_logger_file['fps']).map(
self.map_resolution_n_pixels)
client_logger_file = client_logger_file.sort_values('timestamp_s')
### We have a quality level progression which we need to simulate
quality_level_progression_list = []
insertion_points = (self.video_information_csv.time_s -
self.video_information_csv.seg_len_s).values
for n_segment, points in enumerate(insertion_points):
client_logger_index = np.searchsorted(
client_logger_file['played_until'], points)
if points >= client_logger_file['played_until'].iloc[-1]:
break
quality_level_progression_list.append(
client_logger_file['current_level'].iloc[client_logger_index])
trajectory_object = Trajectory()
trajectory_object.new_trace_video_pair_name(trace_video_pair_name)
previous_quality = 0
current_quality = 0
previous_observation = self.generate_observation_dictionary()
previous_likelihood = np.zeros((1, len(self.quality_change_arr)))
previous_likelihood[0, len(previous_likelihood) // 2 + 1] = 1.
del previous_observation[
'streaming_environment'] # We can't make use of this in the trajectory
video_finished = False
quality_iteration_idx = 0
logging_list = []
while not video_finished:
observation, reward, video_finished, info = self.get_video_chunk(
current_quality)
switch = current_quality - previous_quality
action_idx = self.map_switch_idx(switch)
previous_quality = current_quality
if video_finished or quality_iteration_idx >= len(
quality_level_progression_list):
# Add the last observation
del observation[
'streaming_environment'] # We can't make use of this in the trajectory
trajectory_object.add_trajectory_triple(
previous_observation, observation, action_idx)
trajectory_object.add_likelihood(previous_likelihood, False)
break
# We actually need to map this
# which is the closest quality to the played we can reach
current_quality = quality_level_progression_list[
quality_iteration_idx]
# Should we actually map this. I probaly should map the quality progression to the closes mapping of the quality switches
switch = int(
np.clip(current_quality - previous_quality,
a_min=-self.max_switch_allowed,
a_max=self.max_switch_allowed))
current_quality = previous_quality + switch
del observation[
'streaming_environment'] # We can't make use of this in the trajectory
trajectory_object.add_trajectory_triple(previous_observation,
observation, action_idx)
trajectory_object.add_likelihood(previous_likelihood, False)
previous_likelihood = np.zeros((1, len(self.quality_change_arr)))
previous_likelihood[0, action_idx] = 1.
previous_observation = observation
quality_iteration_idx += 1
streaming_session_evaluation = pd.DataFrame(
self.return_log_state(), columns=self.get_logging_columns())
logging_list.append(
StreamingSessionEvaluation(
streaming_session_evaluation=streaming_session_evaluation,
name=trace_video_pair_name,
max_buffer_length_s=self.buffer_threshold_ms / 1000.,
max_switch_allowed=self.max_switch_allowed))
return logging_list, trajectory_object
def transform_csv(self, to_transform_path):
self.reset()
trace_video_pair_name = to_transform_path.split('/')[-1]
evaluation_dataframe = pd.read_csv(to_transform_path, index_col=0)
quality_level_progression_list = evaluation_dataframe[
'quality_level_chosen'].values
trajectory_object = Trajectory()
trajectory_object.new_trace_video_pair_name(trace_video_pair_name)
previous_quality = 0
current_quality = 0
previous_observation = self.generate_observation_dictionary()
previous_likelihood = np.zeros((1, len(self.quality_change_arr)))
previous_likelihood[0, len(previous_likelihood) // 2 + 1] = 1.
del previous_observation[
'streaming_environment'] # We can't make use of this in the trajectory
video_finished = False
quality_iteration_idx = 0
logging_list = []
while not video_finished:
observation, reward, video_finished, info = self.get_video_chunk(
current_quality)
switch = current_quality - previous_quality
action_idx = self.map_switch_idx(switch)
previous_quality = current_quality
if video_finished or quality_iteration_idx >= len(
quality_level_progression_list):
# Add the last observation
del observation[
'streaming_environment'] # We can't make use of this in the trajectory
trajectory_object.add_trajectory_triple(
previous_observation, observation, action_idx)
trajectory_object.add_likelihood(previous_likelihood, False)
break
current_quality = quality_level_progression_list[
quality_iteration_idx]
# Should we actually map this. I probaly should map the quality progression to the closes mapping of the quality switches
switch = int(
np.clip(current_quality - previous_quality,
a_min=-self.max_switch_allowed,
a_max=self.max_switch_allowed))
current_quality = previous_quality + switch
del observation[
'streaming_environment'] # We can't make use of this in the trajectory
trajectory_object.add_trajectory_triple(previous_observation,
observation, action_idx)
trajectory_object.add_likelihood(previous_likelihood, False)
previous_likelihood = np.zeros((1, len(self.quality_change_arr)))
previous_likelihood[0, action_idx] = 1.
previous_observation = observation
quality_iteration_idx += 1
streaming_session_evaluation = pd.DataFrame(
self.return_log_state(), columns=self.get_logging_columns())
logging_list.append(
StreamingSessionEvaluation(
streaming_session_evaluation=streaming_session_evaluation,
name=trace_video_pair_name,
max_buffer_length_s=self.buffer_threshold_ms / 1000.,
max_switch_allowed=self.max_switch_allowed))
return logging_list, trajectory_object