Merge pull request #2227 from h-mayorquin/add_sorting_generator

Add spike-train based lazy SortingGenerator

Author: alejoe91

src/spikeinterface/core/generate.py

@@ -742,10 +742,10 @@ def synthesize_poisson_spike_vector(
 
     # Calculate the number of frames in the refractory period
     refractory_period_seconds = refractory_period_ms / 1000.0
-    refactory_period_frames = int(refractory_period_seconds * sampling_frequency)
+    refractory_period_frames = int(refractory_period_seconds * sampling_frequency)
 
-    is_refactory_period_too_long = np.any(refractory_period_seconds >= 1.0 / firing_rates)
-    if is_refactory_period_too_long:
+    is_refractory_period_too_long = np.any(refractory_period_seconds >= 1.0 / firing_rates)
+    if is_refractory_period_too_long:
         raise ValueError(
             f"The given refractory period {refractory_period_ms} is too long for the firing rates {firing_rates}"
         )
@@ -764,9 +764,9 @@ def synthesize_poisson_spike_vector(
     binomial_p_modified = modified_firing_rate / sampling_frequency
     binomial_p_modified = np.minimum(binomial_p_modified, 1.0)
 
-    # Generate inter spike frames, add the refactory samples and accumulate for sorted spike frames
+    # Generate inter spike frames, add the refractory samples and accumulate for sorted spike frames
     inter_spike_frames = rng.geometric(p=binomial_p_modified[:, np.newaxis], size=(num_units, num_spikes_max))
-    inter_spike_frames[:, 1:] += refactory_period_frames
+    inter_spike_frames[:, 1:] += refractory_period_frames
     spike_frames = np.cumsum(inter_spike_frames, axis=1, out=inter_spike_frames)
     spike_frames = spike_frames.ravel()
 
@@ -1054,6 +1054,176 @@ def synthetize_spike_train_bad_isi(duration, baseline_rate, num_violations, viol
     return spike_train
 
 
+from spikeinterface.core.basesorting import BaseSortingSegment, BaseSorting
+
+
+class SortingGenerator(BaseSorting):
+    def __init__(
+        self,
+        num_units: int = 20,
+        sampling_frequency: float = 30_000.0,  # in Hz
+        durations: List[float] = [10.325, 3.5],  #  in s for 2 segments
+        firing_rates: float | np.ndarray = 3.0,
+        refractory_period_ms: float | np.ndarray = 4.0,  # in ms
+        seed: int = 0,
+    ):
+        """
+        A class for lazily generate synthetic sorting objects with Poisson spike trains.
+
+        We have two ways of representing spike trains in SpikeInterface:
+
+        - Spike vector (sample_index, unit_index)
+        - Dictionary of unit_id to spike times
+
+        This class simulates a sorting object that uses a representation based on unit IDs to lists of spike times,
+        rather than pre-computed spike vectors. It is intended for testing performance differences and functionalities
+        in data handling and analysis frameworks. For the normal use case of sorting objects with spike_vectors use the
+        `generate_sorting` function.
+
+        Parameters
+        ----------
+        num_units : int, optional
+            The number of distinct units (neurons) to simulate. Default is 20.
+        sampling_frequency : float, optional
+            The sampling frequency of the spike data in Hz. Default is 30_000.0.
+        durations : list of float, optional
+            A list containing the duration in seconds for each segment of the sorting data. Default is [10.325, 3.5],
+            corresponding to 2 segments.
+        firing_rates : float or np.ndarray, optional
+            The firing rate(s) in Hz, which can be specified as a single value applicable to all units or as an array
+            with individual firing rates for each unit. Default is 3.0.
+        refractory_period_ms : float or np.ndarray, optional
+            The refractory period in milliseconds. Can be specified either as a single value for all units or as an
+            array with different values for each unit. Default is 4.0.
+        seed : int, default: 0
+            The seed for the random number generator to ensure reproducibility.
+
+        Raises
+        ------
+        ValueError
+            If the refractory period is too long for the given firing rates, which could result in unrealistic
+            physiological conditions.
+
+        Notes
+        -----
+        This generator simulates the spike trains using a Poisson process. It takes into account the refractory periods
+        by adjusting the firing rates accordingly. See the notes on `synthesize_poisson_spike_vector` for more details.
+
+        """
+
+        unit_ids = np.arange(num_units)
+        super().__init__(sampling_frequency, unit_ids)
+
+        self.num_units = num_units
+        self.num_segments = len(durations)
+        self.firing_rates = firing_rates
+        self.durations = durations
+        self.refractory_period_seconds = refractory_period_ms / 1000.0
+
+        is_refractory_period_too_long = np.any(self.refractory_period_seconds >= 1.0 / firing_rates)
+        if is_refractory_period_too_long:
+            raise ValueError(
+                f"The given refractory period {refractory_period_ms} is too long for the firing rates {firing_rates}"
+            )
+
+        seed = _ensure_seed(seed)
+        self.seed = seed
+
+        for segment_index in range(self.num_segments):
+            segment_seed = self.seed + segment_index
+            segment = SortingGeneratorSegment(
+                num_units=num_units,
+                sampling_frequency=sampling_frequency,
+                duration=durations[segment_index],
+                firing_rates=firing_rates,
+                refractory_period_seconds=self.refractory_period_seconds,
+                seed=segment_seed,
+                t_start=None,
+            )
+            self.add_sorting_segment(segment)
+
+        self._kwargs = {
+            "num_units": num_units,
+            "sampling_frequency": sampling_frequency,
+            "durations": durations,
+            "firing_rates": firing_rates,
+            "refractory_period_ms": refractory_period_ms,
+            "seed": seed,
+        }
+
+
+class SortingGeneratorSegment(BaseSortingSegment):
+    def __init__(
+        self,
+        num_units: int,
+        sampling_frequency: float,
+        duration: float,
+        firing_rates: float | np.ndarray,
+        refractory_period_seconds: float | np.ndarray,
+        seed: int,
+        t_start: Optional[float] = None,
+    ):
+        self.num_units = num_units
+        self.duration = duration
+        self.sampling_frequency = sampling_frequency
+        self.refractory_period_seconds = refractory_period_seconds
+
+        if np.isscalar(firing_rates):
+            firing_rates = np.full(num_units, firing_rates, dtype="float64")
+
+        self.firing_rates = firing_rates
+
+        if np.isscalar(self.refractory_period_seconds):
+            self.refractory_period_seconds = np.full(num_units, self.refractory_period_seconds, dtype="float64")
+
+        self.segment_seed = seed
+        self.units_seed = {unit_id: self.segment_seed + hash(unit_id) for unit_id in range(num_units)}
+        self.num_samples = math.ceil(sampling_frequency * duration)
+        super().__init__(t_start)
+
+    def get_unit_spike_train(self, unit_id, start_frame: int | None = None, end_frame: int | None = None) -> np.ndarray:
+        unit_seed = self.units_seed[unit_id]
+        unit_index = self.parent_extractor.id_to_index(unit_id)
+
+        rng = np.random.default_rng(seed=unit_seed)
+
+        firing_rate = self.firing_rates[unit_index]
+        refractory_period = self.refractory_period_seconds[unit_index]
+
+        # p is the probably of an spike per tick of the sampling frequency
+        binomial_p = firing_rate / self.sampling_frequency
+        # We estimate how many spikes we will have in the duration
+        max_frames = int(self.duration * self.sampling_frequency) - 1
+        max_binomial_p = float(np.max(binomial_p))
+        num_spikes_expected = ceil(max_frames * max_binomial_p)
+        num_spikes_std = int(np.sqrt(num_spikes_expected * (1 - max_binomial_p)))
+        num_spikes_max = num_spikes_expected + 4 * num_spikes_std
+
+        # Increase the firing rate to take into account the refractory period
+        modified_firing_rate = firing_rate / (1 - firing_rate * refractory_period)
+        binomial_p_modified = modified_firing_rate / self.sampling_frequency
+        binomial_p_modified = np.minimum(binomial_p_modified, 1.0)
+
+        inter_spike_frames = rng.geometric(p=binomial_p_modified, size=num_spikes_max)
+        spike_frames = np.cumsum(inter_spike_frames)
+
+        refractory_period_frames = int(refractory_period * self.sampling_frequency)
+        spike_frames[1:] += refractory_period_frames
+
+        if start_frame is not None:
+            start_index = np.searchsorted(spike_frames, start_frame, side="left")
+        else:
+            start_index = 0
+
+        if end_frame is not None:
+            end_index = np.searchsorted(spike_frames[start_index:], end_frame, side="left")
+        else:
+            end_index = int(self.duration * self.sampling_frequency)
+
+        spike_frames = spike_frames[start_index:end_index]
+        return spike_frames
+
+
 ## Noise generator zone ##
 class NoiseGeneratorRecording(BaseRecording):
     """

src/spikeinterface/core/tests/test_generate.py

@@ -10,6 +10,7 @@
     generate_recording,
     generate_sorting,
     NoiseGeneratorRecording,
+    SortingGenerator,
     TransformSorting,
     generate_recording_by_size,
     InjectTemplatesRecording,
@@ -94,6 +95,73 @@ def measure_memory_allocation(measure_in_process: bool = True) -> float:
     return memory
 
 
+def test_memory_sorting_generator():
+    # Test that get_traces does not consume more memory than allocated.
+
+    bytes_to_MiB_factor = 1024**2
+    relative_tolerance = 0.05  # relative tolerance of 5 per cent
+
+    sampling_frequency = 30000  # Hz
+    durations = [60.0]
+    num_units = 1000
+    seed = 0
+
+    before_instanciation_MiB = measure_memory_allocation() / bytes_to_MiB_factor
+    sorting = SortingGenerator(
+        num_units=num_units,
+        sampling_frequency=sampling_frequency,
+        durations=durations,
+        seed=seed,
+    )
+    after_instanciation_MiB = measure_memory_allocation() / bytes_to_MiB_factor
+    memory_usage_MiB = after_instanciation_MiB - before_instanciation_MiB
+    ratio = memory_usage_MiB / before_instanciation_MiB
+    expected_allocation_MiB = 0
+    assert (
+        ratio <= 1.0 + relative_tolerance
+    ), f"SortingGenerator wrong memory {memory_usage_MiB} instead of {expected_allocation_MiB}"
+
+
+def test_sorting_generator_consisency_across_calls():
+    sampling_frequency = 30000  # Hz
+    durations = [1.0]
+    num_units = 3
+    seed = 0
+
+    sorting = SortingGenerator(
+        num_units=num_units,
+        sampling_frequency=sampling_frequency,
+        durations=durations,
+        seed=seed,
+    )
+
+    for unit_id in sorting.get_unit_ids():
+        spike_train = sorting.get_unit_spike_train(unit_id=unit_id)
+        spike_train_again = sorting.get_unit_spike_train(unit_id=unit_id)
+
+        assert np.allclose(spike_train, spike_train_again)
+
+
+def test_sorting_generator_consisency_within_trains():
+    sampling_frequency = 30000  # Hz
+    durations = [1.0]
+    num_units = 3
+    seed = 0
+
+    sorting = SortingGenerator(
+        num_units=num_units,
+        sampling_frequency=sampling_frequency,
+        durations=durations,
+        seed=seed,
+    )
+
+    for unit_id in sorting.get_unit_ids():
+        spike_train = sorting.get_unit_spike_train(unit_id=unit_id, start_frame=0, end_frame=1000)
+        spike_train_again = sorting.get_unit_spike_train(unit_id=unit_id, start_frame=0, end_frame=1000)
+
+        assert np.allclose(spike_train, spike_train_again)
+
+
 def test_noise_generator_memory():
     # Test that get_traces does not consume more memory than allocated.