Merge pull request #1958 from zm711/isi-refactor

Refactor ISI calculation numpy and numba

src/spikeinterface/postprocessing/__init__.py

@@ -38,7 +38,6 @@
 
 from .isi import (
     ISIHistogramsCalculator,
-    compute_isi_histograms_from_spiketrain,
     compute_isi_histograms,
     compute_isi_histograms_numpy,
     compute_isi_histograms_numba,

src/spikeinterface/postprocessing/isi.py

@@ -65,61 +65,6 @@ def get_extension_function():
 WaveformExtractor.register_extension(ISIHistogramsCalculator)
 
 
-def compute_isi_histograms_from_spiketrain(spike_train: np.ndarray, max_time: int, bin_size: int, sampling_f: float):
-    """
-    Computes the Inter-Spike Intervals histogram from a given spike train.
-
-    This implementation only works if you have numba installed, to accelerate the
-    computation time.
-
-    Parameters
-    ----------
-    spike_train: np.ndarray
-        The ordered spike train to compute the ISI.
-    max_time: int
-        Compute the ISI from 0 to +max_time (in sampling time).
-    bin_size: int
-        Size of a bin (in sampling time).
-    sampling_f: float
-        Sampling rate/frequency (in Hz).
-
-    Returns
-    -------
-    tuple (ISI, bins)
-    ISI: np.ndarray[int64]
-        The computed ISI histogram.
-    bins: np.ndarray[float64]
-        The bins for the ISI histogram.
-    """
-    if not HAVE_NUMBA:
-        print("Error: numba is not installed.")
-        print("compute_ISI_from_spiketrain cannot run without numba.")
-        return 0
-
-    return _compute_isi_histograms_from_spiketrain(spike_train.astype(np.int64), max_time, bin_size, sampling_f)
-
-
-if HAVE_NUMBA:
-
-    @numba.jit((numba.int64[::1], numba.int32, numba.int32, numba.float32), nopython=True, nogil=True, cache=True)
-    def _compute_isi_histograms_from_spiketrain(spike_train, max_time, bin_size, sampling_f):
-        n_bins = int(max_time / bin_size)
-
-        bins = np.arange(0, max_time + bin_size, bin_size) * 1e3 / sampling_f
-        ISI = np.zeros(n_bins, dtype=np.int64)
-
-        for i in range(1, len(spike_train)):
-            diff = spike_train[i] - spike_train[i - 1]
-
-            if diff >= max_time:
-                continue
-
-            bin = int(diff / bin_size)
-            ISI[bin] += 1
-
-        return ISI, bins
-
-
 def compute_isi_histograms(
     waveform_or_sorting_extractor,
     load_if_exists=False,
@@ -140,7 +85,7 @@ def compute_isi_histograms(
     bin_ms : float, optional
         The bin size in ms, by default 1.0.
     method : str, optional
-        "auto" | "numpy" | "numba". If _auto" and numba is installed, numba is used, by default "auto"
+        "auto" | "numpy" | "numba". If "auto" and numba is installed, numba is used, by default "auto"
 
     Returns
     -------
@@ -191,21 +136,19 @@ def compute_isi_histograms_numpy(sorting, window_ms: float = 50.0, bin_ms: float
     """
     fs = sorting.get_sampling_frequency()
     num_units = len(sorting.unit_ids)
-
+    assert bin_ms * 1e-3 >= 1 / fs, f"bin size must be larger than the sampling period {1e3 / fs}"
+    assert bin_ms <= window_ms
     window_size = int(round(fs * window_ms * 1e-3))
     bin_size = int(round(fs * bin_ms * 1e-3))
     window_size -= window_size % bin_size
-    num_bins = int(window_size / bin_size)
-    assert num_bins >= 1
-
-    ISIs = np.zeros((num_units, num_bins), dtype=np.int64)
     bins = np.arange(0, window_size + bin_size, bin_size) * 1e3 / fs
+    ISIs = np.zeros((num_units, len(bins) - 1), dtype=np.int64)
 
     # TODO: There might be a better way than a double for loop?
     for i, unit_id in enumerate(sorting.unit_ids):
         for seg_index in range(sorting.get_num_segments()):
             spike_train = sorting.get_unit_spike_train(unit_id, segment_index=seg_index)
-            ISI = np.histogram(np.diff(spike_train), bins=num_bins, range=(0, window_size - 1))[0]
+            ISI = np.histogram(np.diff(spike_train), bins=bins)[0]
             ISIs[i] += ISI
 
     return ISIs, bins
@@ -224,18 +167,18 @@ def compute_isi_histograms_numba(sorting, window_ms: float = 50.0, bin_ms: float
 
     assert HAVE_NUMBA
     fs = sorting.get_sampling_frequency()
+    assert bin_ms * 1e-3 >= 1 / fs, f"the bin_ms must be larger than the sampling period: {1e3 / fs}"
+    assert bin_ms <= window_ms
     num_units = len(sorting.unit_ids)
 
     window_size = int(round(fs * window_ms * 1e-3))
     bin_size = int(round(fs * bin_ms * 1e-3))
     window_size -= window_size % bin_size
-    num_bins = int(window_size / bin_size)
-    assert num_bins >= 1
 
     bins = np.arange(0, window_size + bin_size, bin_size) * 1e3 / fs
     spikes = sorting.to_spike_vector(concatenated=False)
 
-    ISIs = np.zeros((num_units, num_bins), dtype=np.int64)
+    ISIs = np.zeros((num_units, len(bins) - 1), dtype=np.int64)
 
     for seg_index in range(sorting.get_num_segments()):
         spike_times = spikes[seg_index]["sample_index"].astype(np.int64)
@@ -245,9 +188,7 @@ def compute_isi_histograms_numba(sorting, window_ms: float = 50.0, bin_ms: float
             ISIs,
             spike_times,
             spike_labels,
-            window_size,
-            bin_size,
-            fs,
+            bins,
         )
 
     return ISIs, bins
@@ -256,16 +197,15 @@ def compute_isi_histograms_numba(sorting, window_ms: float = 50.0, bin_ms: float
 if HAVE_NUMBA:
 
     @numba.jit(
-        (numba.int64[:, ::1], numba.int64[::1], numba.int32[::1], numba.int32, numba.int32, numba.float32),
+        (numba.int64[:, ::1], numba.int64[::1], numba.int32[::1], numba.float64[::1]),
         nopython=True,
         nogil=True,
         cache=True,
-        parallel=True,
     )
-    def _compute_isi_histograms_numba(ISIs, spike_trains, spike_clusters, max_time, bin_size, sampling_f):
+    def _compute_isi_histograms_numba(ISIs, spike_trains, spike_clusters, bins):
         n_units = ISIs.shape[0]
 
-        for i in numba.prange(n_units):
+        units_loop = numba.prange(n_units) if n_units > 300 else range(n_units)
+        for i in units_loop:
             spike_train = spike_trains[spike_clusters == i]
-
-            ISIs[i] += _compute_isi_histograms_from_spiketrain(spike_train, max_time, bin_size, sampling_f)[0]
+            ISIs[i] += np.histogram(np.diff(spike_train), bins=bins)[0]