the create_track_objects method

createTracks.py

def create_track_objects(self):
	df = self.data_df
	df = self.split_train_val_as_pd(data=df, ratio=self.config.get('valratio', 6))
	df.sort_values(by=['track_id', 'segment_id'], inplace=True)
	df.replace({'animal': 0, 'human': 1}, inplace=True)
	df['target_type'] = df['target_type'].astype(int)
	# validating that each track consists of segments with same values in following columns
	columns_to_check = ['geolocation_type', 'geolocation_id', 'sensor_id', 'snr_type', 'date_index', 'target_type']
	# creating boolean matrix for np.select
	conditions = [(df.groupby('track_id')[col].shift(0) == df.groupby('track_id')[col].shift(1).bfill())
				  for col in columns_to_check]
	df['usable'] = np.select(conditions, conditions, default=False)
	df.loc[df['is_validation'], 'usable'] = False
	df.loc[df['is_validation'].shift(1).fillna(False), 'usable'] = False
	df['usable'] = ~df['usable']
	# save validation segments to object and drop from current DF
	val_df = df.loc[df.is_validation].copy().set_index(['track_id', 'segment_id'])
	df = df.loc[~df.is_validation].copy()
	# Creating a subtrack id for grouping into contiguous segments
	df['subtrack_id'] = df.groupby('track_id').usable.cumsum()
	# create track-level dataframes
	df_tracks = df.groupby(['track_id', 'subtrack_id']).agg(
		target_type=pd.NamedAgg(column="target_type", aggfunc='unique'),
		output_array=pd.NamedAgg(column="iq_sweep_burst", aggfunc=list),
		doppler_burst=pd.NamedAgg(column="doppler_burst", aggfunc=list),
		segment_count=pd.NamedAgg(column="segment_id", aggfunc='count'),
		)
	val_tracks = val_df.groupby(['track_id', 'segment_id']).agg(
		target_type=pd.NamedAgg(column="target_type", aggfunc='unique'),
		output_array=pd.NamedAgg(column="iq_sweep_burst", aggfunc=list),
		doppler_burst=pd.NamedAgg(column="doppler_burst", aggfunc=list),
		segment_count=pd.NamedAgg(column="target_type", aggfunc='count'),
		)
	# turning array into scalar
	df_tracks['target_type'] = df_tracks['target_type'].apply(lambda x: x[0])
	val_tracks['target_type'] = val_tracks['target_type'].apply(lambda x: x[0])
	# concatenating doppler bursts
	df_tracks['doppler_burst'] = df_tracks['doppler_burst'].apply(lambda x: np.concatenate(x, axis=-1))
	val_tracks['doppler_burst'] = val_tracks['doppler_burst'].apply(lambda x: np.concatenate(x, axis=-1))
	# concatenating IQ matrices along slow-time axis
	df_tracks['output_array'] = df_tracks['output_array'].apply(lambda x: np.concatenate(x, axis=1))
	val_tracks['output_array'] = val_tracks['output_array'].apply(lambda x: np.concatenate(x, axis=1))
	# transforming IQ to scalogram or spectrogram
	if self.output_data_type == 'scalogram':
		print('Converting IQ matricies to Scalogram')
		df_tracks['output_array'] = df_tracks['output_array'].progress_apply(iq_to_scalogram,
										     transformation=self.config['mother_wavelet'],
										     scale=self.config['scale'])
		val_tracks['output_array'] = val_tracks['output_array'].progress_apply(iq_to_scalogram,
										       transformation=self.config['mother_wavelet'],
										       scale=self.config['scale'])
	else:
		print('Converting IQ matricies to Spectrogram')
		df_tracks['output_array'] = df_tracks['output_array'].progress_apply(iq_to_spectogram)
		val_tracks['output_array'] = val_tracks['output_array'].progress_apply(iq_to_spectogram)

		if self.config.get('include_doppler'):
			df_tracks['output_array'] = df_tracks.progress_apply(lambda row: max_value_on_doppler(row['output_array'], 
													      row['doppler_burst']), axis=1)
			val_tracks['output_array'] = val_tracks.progress_apply(lambda row: max_value_on_doppler(row['output_array'], 
														row['doppler_burst']), axis=1)
		df_tracks['output_array'] = df_tracks['output_array'].progress_apply(normalize)
		val_tracks['output_array'] = val_tracks['output_array'].progress_apply(normalize)
	# generating list of _Segment objects
	train_segments = [_Segment(segment_id=f'{k[0]}_{k[1]}', **v) for k, v in
					  df_tracks.to_dict(orient='index').items()]
	val_segments = [_Segment(segment_id=f'{k[0]}_{k[1]}', **v) for k, v in
					val_tracks.to_dict(orient='index').items()]
	return train_segments, val_segments