I have a time series data array of shape (#timestamp,#features). I would like to extract for each row (timestamp) n_lags (previous rows) and reshape the array such that I have the shape (#samples, #lags+now,#features) for an input to a LSTM layer of Keras.
Consider this toy example:
import numpy as np
n_rows = 6
n_feat= 3
n_lag = 2
a = np.array(range(n_rows*n_feat)).reshape(n_rows, n_feat)
>>> a.shape = (6, 3)
>>> a = array([[ 0, 1, 2],
[ 3, 4, 5],
[ 6, 7, 8],
[ 9, 10, 11],
[12, 13, 14],
[15, 16, 17]])
With iterating over rows I achieve the expected output:
b = np.empty(shape=(0, (n_lag + 1), n_feat))
for idx, row in enumerate(a):
temp = np.expand_dims(a[max(0, idx-n_lag):idx+1, :], 0)
if temp.shape[1:] == b.shape[1:]:
b = np.append(b, temp, axis=0)
>>> b.shape = (4, 3, 3)
>>> b = array([[[ 0., 1., 2.],
[ 3., 4., 5.],
[ 6., 7., 8.]],
[[ 3., 4., 5.],
[ 6., 7., 8.],
[ 9., 10., 11.]],
[[ 6., 7., 8.],
[ 9., 10., 11.],
[12., 13., 14.]],
[[ 9., 10., 11.],
[12., 13., 14.],
[15., 16., 17.]]])
Note: the first n_lags-1 rows do not have enough data and will be discarded in the final output
Question: I would like to know if there is a more elegant / nice way than iterating over the rows.
You can use the new np.lib.stride_tricks.sliding_window_view for this
n_rows = 6
n_feat= 3
n_lag = 2
a = np.array(range(n_rows*n_feat)).reshape(n_rows, n_feat)
b = np.lib.stride_tricks.sliding_window_view(a, window_shape=(n_feat, n_feat))
b
output:
array([[[[ 0, 1, 2],
[ 3, 4, 5],
[ 6, 7, 8]]],
[[[ 3, 4, 5],
[ 6, 7, 8],
[ 9, 10, 11]]],
[[[ 6, 7, 8],
[ 9, 10, 11],
[12, 13, 14]]],
[[[ 9, 10, 11],
[12, 13, 14],
[15, 16, 17]]]])
b will just change the shape and strides of a, so it will contain the same memory location of a multiple times. In other words, no need to allocate a new array.