def parse_csv(filename, hdim):
"""Parses ECG data from csv file."""
df = pd.read_csv(filename)
labels = df[df['input channel '] == 0].label.values
seqs = []
for i, row in df.iterrows():
if i % 2:
assert row[1] == 1
continue
assert row[1] == 0
ch1_vals = row[3:].values.astype(np.float32).reshape(-1, 1)
ch2_vals = df.iloc[i + 1][3:].values.astype(np.float32).reshape(-1, 1)
if FLAGS.channel == 'both':
values = np.concatenate([ch1_vals, ch2_vals], axis=1)
elif FLAGS.channel == '1':
values = ch1_vals
elif FLAGS.channel == '2':
values = ch2_vals
else:
raise ValueError('Channel flag expected to be both, 1, or 2.')
seq_len = values.shape[0]
if np.isnan(values).any():
seq_len = np.min(np.where(np.isnan(values).any(axis=1)))
assert np.isnan(values[seq_len:]).all()
values = values[:seq_len]
seqs.append(
lds.LinearDynamicalSystemSequence(np.zeros((seq_len, 1)),
np.zeros((seq_len, hdim)),
values))
return seqs, labels
def parse_csv(filename, hdim):
"""Reads ECG data from csv file."""
labels = []
seqs = []
unprocessed_key = None
unprocessed_label = None
unprocessed_ch0 = None
not_full_length = 0
with open(filename, 'rb') as csvfile:
reader = csv.reader(csvfile)
for row in reader:
key = row[0]
channel = row[1]
label = row[2]
channel_signal = np.array(row[3:]).reshape(-1, 1)
try:
channel_signal = channel_signal.astype(np.float32)
except ValueError:
channel_signal = np.array([
float(x) if x else np.nan for x in row[3:]
]).reshape(-1, 1)
# logging.info('Partial signal of len %d with key %s',
# sum(~np.isnan(channel_signal)), key)
not_full_length += 1
if channel == '0':
assert unprocessed_ch0 is None
unprocessed_ch0 = channel_signal
unprocessed_key = key
unprocessed_label = label
if channel == '1':
assert unprocessed_ch0 is not None
seq_len = len(channel_signal)
assert len(unprocessed_ch0) == seq_len
if FLAGS.channel == 'both':
vals = np.concatenate([unprocessed_ch0, channel_signal],
axis=1)
elif FLAGS.channel == '0':
vals = unprocessed_ch0
elif FLAGS.channel == '1':
vals = channel_signal
else:
raise ValueError('Unexpected FLAGS.channel value: %s' %
FLAGS.channel)
seqs.append(
lds.LinearDynamicalSystemSequence(
np.zeros((seq_len, 1)), np.zeros((seq_len, hdim)),
vals))
assert label == unprocessed_label
assert key.split(':')[:2] == unprocessed_key.split(':')[:2]
labels.append(label)
unprocessed_label = None
unprocessed_key = None
unprocessed_ch0 = None
logging.info('Total seqs: %d, partial length seqs: %d.', len(seqs),
not_full_length)
if FLAGS.filter_type:
seqs, labels = filter_type(seqs, labels)
seqs, labels = drop_infreq_labels(seqs, labels)
return seqs, labels
def subsample(sequences, step_size=5):
subsampled = []
for s in sequences:
subsampled.append(
lds.LinearDynamicalSystemSequence(
_subsample_rows(s.inputs, step_size),
_subsample_rows(s.hidden_states, step_size),
_subsample_rows(s.outputs, step_size)))
return subsampled