incorrect number of iterations in `snippets._get_all_mpdist_profiles`

[Nima Sarajpoor]

Issue

In the function _get_all_mpdist_profiless, the time series T is extended so that to have a length that is a multiples of m.

stumpy/stumpy/snippets.py

Lines 103 to 108 in 1ddb950

 right_pad = 0 
 T_subseq_isconstant = core.process_isconstant(T, s, mpdist_T_subseq_isconstant) 
 if T.shape[0] % m != 0: 
     right_pad = int(m * np.ceil(T.shape[0] / m) - T.shape[0]) 
     pad_width = (0, right_pad) 
     T = np.pad(T, pad_width, mode="constant", constant_values=np.nan) 

Note that the length of the original T does not change IF len(T) % m == 0. In such case, it seems that the following for-loop misses the last non-overlapping window of length m

stumpy/stumpy/snippets.py

Lines 113 to 123 in 1ddb950

 n_padded = T.shape[0] 
 D = np.empty(((n_padded // m) - 1, n_padded - m + 1), dtype=np.float64) 
  
 M_T, Σ_T = core.compute_mean_std(T, s) 
  
 # Iterate over non-overlapping subsequences, see Definition 3 
 for i in range((n_padded // m) - 1): 
     start = i * m 
     stop = (i + 1) * m 
     S_i = T[start:stop] 
     D[i, :] = _mpdist_vect( 

Solution

Before the if check if T.shape[0] % m != 0:, we may add: n_windows = T.shape[0] // m and then we do for-loop using range(n_windows)

[Sean M. Law]

@NimaSarajpoor In both cases (with and without padding), wouldn't it be sufficient to do:

for i in range(n_padded // m): 

So we omit the - 1

[Nima Sarajpoor]

First, I should mention that my aim was to fix the shape of distance matrix D as well. D.shape[0] should be n_windows.

wouldn't it be sufficient to do:

I don't think it would since I am inclined to believe that the snippet needs to have full-size m. Therefore, if len(T) % m != 0, then the T[-r: ] (where r = len(T) % m) should not be considered as snippet.

[Sean M. Law]

Before the if check if T.shape[0] % m != 0:, we may add: n_windows = T.shape[0] // m and then we do for-loop using range(n_windows)

I accept your proposal. However, can we call it n_contiguous_windows ("contiguous" == "non-overlapping")?

[Nima Sarajpoor]

Before the if check if T.shape[0] % m != 0:, we may add: n_windows = T.shape[0] // m and then we do for-loop using range(n_windows)

I accept your proposal. However, can we call it n_contiguous_windows ("contiguous" == "non-overlapping")?

Sure. n_contiguous_windows is more informative. Will make changes and submit a PR.