def _generate_timeseries(self, inp_data, tgt_data, timestep_start):
self.num_timesteps = inp_data.shape[0]
self.times = (np.arange(self.num_timesteps) + timestep_start) * self.dt
t_stop = (timestep_start + self.num_timesteps) * self.dt
self.input = TSContinuous(self.times, inp_data, t_stop=t_stop)
self.target = TSContinuous(self.times, tgt_data)
self.duration = self.num_timesteps * self.dt
def test_continuous_call():
from rockpool import TSContinuous
# - Generate series
times = np.arange(1, 5) * 0.1
samples = np.arange(4).reshape(-1, 1) + np.arange(2) * 2
ts = TSContinuous(times, samples)
ts_empty = TSContinuous()
ts_single = TSContinuous(2, [3, 2])
# - Call ts
assert np.allclose(ts(0.1), np.array([[0, 2]]))
assert np.allclose(ts(0.25), np.array([[1.5, 3.5]]))
assert (np.isnan(ts(0))).all() and ts(0).shape == (1, 2)
samples_ts = ts([0, 0.1, 0.25])
assert np.allclose(samples_ts[1:], np.array([[0, 2], [1.5, 3.5]]))
assert (np.isnan(samples_ts[0])).all() and samples_ts.shape == (3, 2)
# - Call ts_empty
assert ts_empty(1).shape == (1, 0)
assert ts_empty([0, 1, 4]).shape == (3, 0)
# - Call ts_single
assert (ts_single(2) == np.array([[3, 2]])).all()
assert np.isnan(ts_single(0)).all() and ts_single(0).shape == (1, 2)
samples_single = ts_single([0, 2, 1, 3, 2, 4])
assert (np.isnan(samples_single) == np.array([
2 * [True], 2 * [False], 2 * [True], 2 * [True], 2 * [False],
2 * [True]
])).all()
assert (samples_single[1] == np.array([3, 2])).all()
assert (samples_single[4] == np.array([3, 2])).all()
def test_ForceRateEulerJax():
""" Test ForceRateEulerJax """
from rockpool import TSContinuous
from rockpool.layers import ForceRateEulerJax
# - Generic parameters
w_in = 2 * np.random.rand(1, 2) - 1
w_out = 2 * np.random.rand(2, 1) - 1
bias = 2 * np.random.rand(2) - 1
tau = 20e-3 * np.ones(2)
# - Layer generation
fl0 = ForceRateEulerJax(w_in=w_in,
w_out=w_out,
bias=bias,
noise_std=0.1,
tau=tau,
dt=0.01)
# - Input signal
tsInCont = TSContinuous(times=np.arange(15) * 0.01,
samples=np.ones((15, 1)))
tsForceCont = TSContinuous(times=np.arange(15) * 0.01,
samples=np.ones((15, 2)))
# - Compare states and time before and after
vStateBefore = np.copy(fl0.state)
ts_output = fl0.evolve(tsInCont, tsForceCont, duration=0.1)
assert fl0.t == 0.1
assert (vStateBefore != fl0.state).any()
fl0.reset_all()
assert fl0.t == 0
assert (vStateBefore == fl0.state).all()
# - Test that some errors are caught
with pytest.raises(AssertionError):
fl1 = ForceRateEulerJax(
w_in=np.zeros((1, 2)),
w_out=np.zeros((3, 1)),
tau=np.zeros(3),
bias=np.zeros(3),
)
with pytest.raises(AssertionError):
fl1 = ForceRateEulerJax(
w_in=np.zeros((1, 2)),
w_out=np.zeros((2, 1)),
tau=np.zeros(3),
bias=np.zeros(3),
)
with pytest.raises(AssertionError):
fl1 = ForceRateEulerJax(
w_in=np.zeros((1, 2)),
w_out=np.zeros((2, 1)),
tau=np.zeros(2),
bias=np.zeros(3),
)
def test_save_load():
"""
Test saving and loading function for timeseries
"""
from rockpool import TSEvent, TSContinuous, load_ts_from_file
from os import remove
# - Generate time series objects
times = [1, 3, 6]
samples = np.random.randn(3)
channels = [0, 1, 1]
tsc = TSContinuous(times,
samples,
t_start=-1,
t_stop=8,
periodic=True,
name="continuous")
tse = TSEvent(
times,
channels,
num_channels=3,
t_start=-1,
t_stop=8,
periodic=True,
name="events",
)
# - Store objects
tsc.save("test_tsc")
tse.save("test_tse")
# - Load objects
tscl = load_ts_from_file("test_tsc.npz")
tsel = load_ts_from_file("test_tse.npz")
# - Verify that attributes are still correct
assert (tscl.times == times).all(), "TSContinuous: times changed."
assert (tscl.samples == samples.reshape(
-1, 1)).all(), "TSContinuous: samples changed."
assert tscl.name == "continuous", "TSContinuous: name changed."
assert tscl.t_start == -1, "TSContinuous: t_start changed."
assert tscl.t_stop == 8, "TSContinuous: t_stop changed."
assert tscl.periodic, "TSContinuous: periodic changed."
assert (tsel.times == times).all(), "TSEvent: times changed."
assert (tsel.channels == channels).all(), "TSEvent: channels changed."
assert tsel.name == "events", "TSEvent: name changed."
assert tsel.t_start == -1, "TSEvent: t_start changed."
assert tsel.t_stop == 8, "TSEvent: t_stop changed."
assert tsel.periodic, "TSEvent: periodic changed."
assert tsel.num_channels == 3, "TSEvent: num_channels changed."
# - Remove saved files
remove("test_tsc.npz")
remove("test_tse.npz")
def test_updown():
""" Test FFUpDown """
from rockpool import TSContinuous
from rockpool.layers import FFUpDown
# - Generic parameters
weights = np.random.rand(2, 4)
# - Layer generation
fl0 = FFUpDown(weights=weights, dt=0.01, thr_down=0.02, thr_up=0.01, tau_decay=0.1)
# - Check layer properties
assert fl0.size == 4, "Problem with size"
assert fl0.size_in == 2, "Problem with size_in"
assert (fl0.thr_down == np.array([0.02, 0.02])).all(), "Problem with thr_down"
assert (fl0.thr_up == np.array([0.01, 0.01])).all(), "Problem with thr_up"
# - Input signal
tsInCont = TSContinuous(
times=np.arange(15) * 0.01,
samples=np.vstack(
(np.sin(np.linspace(0, 1, 15)), np.cos(np.linspace(0, 1, 15)))
).T,
)
# - Compare states and time before and after evolution
vStateBefore = np.copy(fl0.state)
fl0.evolve(tsInCont, duration=0.1)
assert fl0.t == 0.1
assert (vStateBefore != fl0.state).any()
fl0.reset_all()
assert fl0.t == 0
assert (vStateBefore == fl0.state).all()
def test_updown_in_net():
""" Test RecRateEuler """
from rockpool import TSContinuous
from rockpool.networks import Network
from rockpool.layers import FFUpDown
from rockpool.layers import RecDIAF
# - Generic parameters
weights = np.random.rand(2, 4)
# - Layer generation
fl0 = FFUpDown(weights=weights)
fl1 = RecDIAF(np.zeros((4, 2)), np.zeros((2, 2)), dt=0.002)
# - Generate network
net = Network(fl0, fl1)
# - Input signal
tsInCont = TSContinuous(
times=np.arange(15) * 0.01,
samples=np.vstack(
(np.sin(np.linspace(0, 1, 15)), np.cos(np.linspace(0, 1, 15)))
).T,
)
# - Compare states and time before and after evolution
vStateBefore = np.copy(fl1.state)
net.evolve(tsInCont, duration=0.1)
assert net.t == 0.1
assert (vStateBefore != fl1.state).any()
def test_continuous_merge():
"""
Test merge method of TSContinuous
"""
from rockpool import TSContinuous
# - Generate a few TSEvent objects
samples = np.random.randint(10, size=(2, 6))
empty_series = TSContinuous(t_start=1)
series_list = []
series_list.append(
TSContinuous([1, 2], samples[:2, :2], t_start=-1, t_stop=3))
series_list.append(
TSContinuous([1.5], samples[0, 2:4], t_start=0, t_stop=4))
series_list.append(
TSContinuous([2, 2.5], samples[:2, 4:6], t_start=-2, t_stop=3))
# Merging two series
merged_fromtwo = series_list[0].merge(series_list[1])
assert merged_fromtwo.t_start == -1, "Wrong t_start for merged series."
assert merged_fromtwo.t_stop == 4, "Wrong t_stop for merged series."
assert (merged_fromtwo.times == np.array(
[1, 1.5, 2])).all(), "Wrong time trace for merged series."
correct_samples = np.vstack((samples[0, :2], samples[0,
2:4], samples[1, :2]))
assert (merged_fromtwo.samples == correct_samples
).all(), "Wrong samples for merged series."
# Merging with empty series
merged_empty_first = empty_series.merge(series_list[0])
assert (merged_empty_first.t_start == series_list[0].t_start
), "Wrong t_start when merging with empty."
assert (merged_empty_first.t_stop == series_list[0].t_stop
), "Wrong t_stop when merging with empty."
assert (merged_empty_first.samples == series_list[0].samples
).all(), "Wrong samples when merging with empty"
assert (merged_empty_first.times == series_list[0].times
).all(), "Wrong time trace when merging with empty"
merged_empty_last = series_list[0].merge(empty_series)
assert (merged_empty_last.t_start == series_list[0].t_start
), "Wrong t_start when merging with empty."
assert (merged_empty_last.t_stop == series_list[0].t_stop
), "Wrong t_stop when merging with empty."
assert (merged_empty_last.samples == series_list[0].samples
).all(), "Wrong samples when merging with empty"
assert (merged_empty_last.times == series_list[0].times
).all(), "Wrong time trace when merging with empty"
# Merging with list of series
merged_with_list = empty_series.merge(series_list, remove_duplicates=True)
assert (merged_with_list.num_channels == 2
), "Wrong channel count when merging with list."
assert merged_with_list.t_start == -2, "Wrong t_start when merging with list."
assert merged_with_list.t_stop == 4, "Wrong t_stop when merging with list."
assert (merged_with_list.times == np.array(
[1, 1.5, 2, 2.5])).all(), "Wrong time trace when merging with list."
assert (merged_with_list.samples == np.vstack(
(samples[0, :2], samples[0, 2:4], samples[1, :2],
samples[1, 4:6]))).all(), "Wrong samples when merging with list."
def test_continuous_append_c():
"""
Test append_c method of TSContinuous
"""
from rockpool import TSContinuous
# - Generate a few TSContinuous objects
samples = np.random.randint(10, size=(2, 4))
empty_series = TSContinuous()
series_list = []
series_list.append(
TSContinuous([1, 2], samples[:2, :2], t_start=-1, t_stop=2))
series_list.append(TSContinuous([1], samples[0, -2:], t_start=0, t_stop=2))
# Appending two series
appended_fromtwo = series_list[0].append_c(series_list[1])
assert appended_fromtwo.t_start == -1, "Wrong t_start for appended series."
assert appended_fromtwo.t_stop == 2, "Wrong t_stop for appended series."
assert (appended_fromtwo.times == np.array(
[1, 2])).all(), "Wrong time trace for appended series."
assert ((appended_fromtwo.samples[:, :2] == samples[:2, :2]).all()
and (appended_fromtwo.samples[0, -2:] == samples[0, -2:]).all()
and (np.isnan(appended_fromtwo.samples[1, -2:])).all()
), "Wrong samples for appended series."
# Appending with empty series
appended_empty_first = empty_series.append_c(series_list[0])
assert (appended_empty_first.t_start == empty_series.t_start
), "Wrong t_start when appending with empty."
assert (appended_empty_first.t_stop == empty_series.t_stop
), "Wrong t_stop when appending with empty."
assert (appended_empty_first.samples.flatten() == empty_series.samples.
flatten()).all(), "Wrong samples when appending with empty"
assert (appended_empty_first.times == empty_series.times
).all(), "Wrong time trace when appending with empty"
appended_empty_last = series_list[0].append_c(empty_series)
assert (appended_empty_last.t_start == series_list[0].t_start
), "Wrong t_start when appending with empty."
assert (appended_empty_last.t_stop == series_list[0].t_stop
), "Wrong t_stop when appending with empty."
assert (appended_empty_last.samples == series_list[0].samples
).all(), "Wrong samples when appending with empty"
assert (appended_empty_last.times == series_list[0].times
).all(), "Wrong time trace when appending with empty"
def test_continuous_operators():
"""
Test creation and manipulation of a continuous time series
"""
from rockpool import TSContinuous
# - Creation
ts = TSContinuous([0], [0])
ts = TSContinuous([0, 1, 2, 3], [1, 2, 3, 4])
ts2 = TSContinuous([1, 2, 3, 4], [5, 6, 7, 8])
# - Samples don't match time
with pytest.raises(ValueError):
TSContinuous([0, 1, 2], [0])
# - Addition
ts = ts + 1
ts += 5
ts = ts + ts2
ts += ts2
# - Subtraction
ts = ts - 3
ts -= 2
ts = ts - ts2
ts -= ts2
# - Multiplication
ts = ts * 0.9
ts *= 0.2
ts = ts * ts2
ts *= ts2
# - Division
ts = ts / 2.0
ts /= 1.0
ts = ts / ts2
ts /= ts2
# - Floor division
ts = ts // 1.0
ts //= 1.0
ts = ts // ts2
ts //= ts2
def test_largescale():
from rockpool import TSEvent, TSContinuous
from rockpool.layers import RecLIFCurrentInJax, RecLIFJax, RecLIFJax_IO
# Numpy
import numpy as np
# - Define network
N = 200
Nin = 500
Nout = 1
tau_mem = 50e-3
tau_syn = 100e-3
bias = 0.0
def rand_params(N, Nin, Nout, tau_mem, tau_syn, bias):
return {
"w_in": np.random.rand(Nin, N) - 0.5,
"w_recurrent": 0.1 * np.random.randn(N, N) / np.sqrt(N),
"w_out": 2 * np.random.rand(N, Nout) - 1,
"tau_mem": tau_mem,
"tau_syn": tau_syn,
"bias": (np.ones(N) * bias).reshape(N),
}
# - Build a random network
params0 = rand_params(N, Nin, Nout, tau_mem, tau_syn, bias)
lyrIO = RecLIFJax_IO(**params0)
# - Define input and target
numRepeats = 1
dur_input = 1000e-3
dt = 1e-3
T = int(np.round(dur_input / dt))
timebase = np.linspace(0, T * dt, T)
trigger = np.atleast_2d(timebase < 50e-3).T
chirp = np.atleast_2d(np.sin(timebase * 2 * np.pi * (timebase * 10))).T
target_ts = TSContinuous(timebase, chirp, periodic=True, name="Target")
spiking_prob = 0.01
sp_in_ts = np.random.rand(T * numRepeats, Nin) < spiking_prob * trigger
spikes = np.argwhere(sp_in_ts)
input_sp_ts = TSEvent(
timebase[spikes[:, 0]],
spikes[:, 1],
name="Input",
periodic=True,
t_start=0.0,
t_stop=dur_input,
)
lyrIO.evolve(input_sp_ts)
def test_continuous_indexing():
from rockpool import TSContinuous
# - Generate series
times = np.arange(6) * 0.1
samples = np.arange(6).reshape(-1, 1) + np.arange(4)
ts = TSContinuous(times, samples)
# - Indexing time
ts0 = ts[:]
assert (ts0.times == times).all()
assert (ts0.samples == samples).all()
ts1 = ts[0.2:0.5]
assert (ts1.times == times[[2, 3, 4]]).all()
assert (ts1.samples == samples[[2, 3, 4]]).all()
ts2 = ts[0.1:0.5:-0.2]
assert (ts2.times == times[[3, 1]]).all()
assert (ts2.samples == samples[[3, 1]]).all()
ts3 = ts[[0.5, 0, 0.1]]
assert (ts3.times == times[[5, 0, 1]]).all()
assert (ts3.samples == samples[[5, 0, 1]]).all()
ts4 = ts[0.2]
assert (ts4.times == times[[2]]).all()
assert (ts4.samples == samples[2]).all()
# - Indexing channels
ts0 = ts[:, :]
assert (ts0.times == times).all()
assert (ts0.samples == samples).all()
ts1 = ts[None, 1:3]
assert (ts1.times == times).all()
assert (ts1.samples == samples[:, 1:3]).all()
ts2 = ts[:, 1::-2]
assert (ts2.times == times).all()
assert (ts2.samples == samples[:, [3, 1]]).all()
ts3 = ts[None, [2, 0, 3]]
assert (ts3.times == times).all()
assert (ts3.samples == samples[:, [2, 0, 3]]).all()
ts4 = ts[:, 2]
assert (ts4.times == times).all()
assert (ts4.samples == samples[:, [2]]).all()
# - Indexing channels and time
ts0 = ts[:0.4, [3, 1]]
assert (ts0.times == times[:4]).all()
assert (ts0.samples == samples[:4, [3, 1]]).all()
def test_evolve():
from rockpool import TSContinuous
# - Get a network
netRes = test_build()
# - Generate an input
time_trace = np.linspace(0, 10, 100)
ts_input = TSContinuous(time_trace, np.random.rand(100))
# - Evolve the network
resp = netRes.evolve(ts_input)
def test_adam():
from rockpool.layers import RecRateEulerJax
from rockpool.layers.training import add_train_output
from rockpool import TSContinuous
# - Generic parameters
w_in = 2 * np.random.rand(1, 2) - 1
w_recurrent = 2 * np.random.rand(2, 2) - 1
w_out = 2 * np.random.rand(2, 1) - 1
# - Layer generation
fl0 = RecRateEulerJax(
w_in=w_in,
w_recurrent=w_recurrent,
w_out=w_out,
bias=0,
noise_std=0.1,
tau=20,
dt=1,
)
# - Add training shim
fl0 = add_train_output(fl0)
# - Define simple input and target
ts_input = TSContinuous([0, 1, 2, 3], [0, 1, 0, 0])
ts_target = TSContinuous([0, 1, 2, 3], [0.1, 0.2, 0.3, 0.4])
# - Initialise training
loss_fcn, grad_fcn = fl0.train_adam(ts_input, ts_target, is_first=True)
# - Test loss and gradient functions
loss_fcn()
grad_fcn()
# - Perform intermediate training step
fl0.train_adam(ts_input, ts_target)
# - Perform final training step
fl0.train_adam(ts_input, ts_target, is_last=True)
def test_continuous_clip():
from rockpool import TSContinuous
# - Generate series
times = np.arange(1, 6) * 0.1
samples = np.arange(5).reshape(-1, 1) + np.arange(2) * 2
ts = TSContinuous(times, samples)
ts_empty = TSContinuous()
# - Clip ts in time
assert (ts.clip(0.2, 0.4, include_stop=True).times == times[1:4]).all()
assert (ts.clip(0.2, 0.4, include_stop=True).samples == samples[1:4]).all()
assert (ts.clip(0.2, 0.4, include_stop=False).times == times[1:3]).all()
assert (ts.clip(0.2, 0.4,
include_stop=False).samples == samples[1:3]).all()
ts_limits = ts.clip(0.2, 0.35, include_stop=True, sample_limits=True)
assert np.allclose(ts_limits.times, np.array([0.2, 0.3, 0.35]))
expected_samples = np.vstack((samples[1:3], [2.5, 4.5]))
assert np.allclose(ts_limits.samples, expected_samples)
ts_beyond = ts.clip(0.4, 0.6, sample_limits=False)
assert (ts_beyond.times == times[-2:]).all()
assert (ts_beyond.samples == samples[-2:]).all()
# - Clip ts channels
assert (ts.clip(channels=1).times == times).all()
assert (ts.clip(channels=1).samples == samples[:, [1]]).all()
assert (ts.clip(channels=[1, 0]).times == times).all()
assert (ts.clip(channels=[1, 0]).samples == samples[:, [1, 0]]).all()
# - Clip ts channels and time
ts_ch_t = ts.clip(0.2, 0.4, channels=1, include_stop=True)
assert (ts_ch_t.times == times[1:4]).all()
assert (ts_ch_t.samples == samples[1:4, [1]]).all()
# - Clip empty
ts_et = ts_empty.clip(0.2, 0.4, sample_limits=False)
assert ts_et.isempty()
assert ts_et.t_start == 0.2 and ts_et.t_stop == 0.4
with pytest.raises(IndexError):
ts_empty.clip(channels=0)
ts_empty.clip(2, 4, channels=0)
def test_FFRateEuler():
""" Test FFRateEuler """
from rockpool import TSContinuous
from rockpool.layers import FFRateEuler
# - Generic parameters
weights = 2 * np.random.rand(2, 3) - 1
bias = 2 * np.random.rand(3) - 1
# - Layer generation
fl0 = FFRateEuler(weights=weights, bias=bias, noise_std=0.1, dt=0.01)
# - Input signal
tsInCont = TSContinuous(times=np.arange(15) * 0.01,
samples=np.ones((15, 2)))
# - Compare states and time before and after
vStateBefore = np.copy(fl0.state)
fl0.evolve(tsInCont, duration=0.1)
assert fl0.t == 0.1
assert (vStateBefore != fl0.state).any()
fl0.reset_all()
assert fl0.t == 0
assert (vStateBefore == fl0.state).all()
# - Test that some errors are caught
with pytest.raises(TypeError):
fl1 = FFRateEuler(weights=None)
with pytest.raises(AssertionError):
fl1 = FFRateEuler(weights=1, bias=[1, 1])
with pytest.raises(AssertionError):
fl1 = FFRateEuler(weights=1, tau=[1, 1])
with pytest.raises(AssertionError):
fl1 = FFRateEuler(weights=1, gain=[1, 1])
tau_decay=np.array((None, None)),
name="updown",
multiplex_spikes=True,
)
# - Load data
classprobs = {0: 0.8, 1: 0.05, 2: 0, 3: 0.05, 4: 0.05, 5: 0, 18: 0.05}
annotations, recordings = rec.load_from_file(rec.save_path)
data_in, data_tgt, anno_curr = rec.generate_data(
num_beats=1000,
annotations=annotations,
rec_data=recordings,
use_recordings=None,
probs=classprobs,
include_bad_signal=False,
min_len_segment=2,
use_cont_segments=True,
)
rhythm_starts = anno_curr.idx_new_start
# - Convert to spikes
times = np.arange(data_in.shape[0]) * dt_in
spike_input = aslayer.evolve(TSContinuous(times, data_in))
# - Load to chip
reservoir.load_events(
tsAS=spike_input,
vtRhythmStart=np.array(rhythm_starts) * dt_in,
tTotalDuration=(data_in.shape[0] + 1) * dt_in,
)
def test_training_FFwd():
from rockpool import TSEvent, TSContinuous
from rockpool.layers import RecLIFCurrentInJax, RecLIFJax, RecLIFJax_IO, FFLIFJax_IO
import numpy as np
N = 100
Nin = 100
Nout = 1
tau_mem = 50e-3
tau_syn = 100e-3
bias = 0.0
dt = 1e-3
def rand_params(N, Nin, Nout, tau_mem, tau_syn, bias):
return {
"w_in": (np.random.rand(Nin, N) - 0.5) / Nin,
"w_out": 2 * np.random.rand(N, Nout) - 1,
"tau_mem": tau_mem,
"tau_syn": tau_syn,
"bias": (np.ones(N) * bias).reshape(N),
}
# - Generate a network
params0 = rand_params(N, Nin, Nout, tau_mem, tau_syn, bias)
lyrIO = FFLIFJax_IO(**params0, dt=dt)
# - Define input and target
numRepeats = 1
dur_input = 1000e-3
dt = 1e-3
T = int(np.round(dur_input / dt))
timebase = np.linspace(0, T * dt, T)
trigger = np.atleast_2d(timebase < dur_input).T
chirp = np.atleast_2d(np.sin(timebase * 2 * np.pi * (timebase * 10))).T
target_ts = TSContinuous(timebase, chirp, periodic=True, name="Target")
spiking_prob = 0.01
sp_in_ts = np.random.rand(T * numRepeats, Nin) < spiking_prob * trigger
spikes = np.argwhere(sp_in_ts)
input_sp_ts = TSEvent(
timebase[spikes[:, 0]],
spikes[:, 1],
name="Input",
periodic=True,
t_start=0,
t_stop=dur_input,
)
# - Simulate initial network state
lyrIO.randomize_state()
lyrIO.evolve(input_sp_ts)
# - Add training shim
from rockpool.layers.training import add_shim_lif_jax_sgd
lyrIO = add_shim_lif_jax_sgd(lyrIO)
# - Train
steps = 100
for t in range(steps):
lyrIO.randomize_state()
l_fcn, g_fcn = lyrIO.train_output_target(input_sp_ts,
target_ts,
is_first=(t == 0))
l_fcn()
g_fcn()
def test_RecLIFCurrentInJax():
""" Test RecLIFCurrentInJax """
from rockpool import TSContinuous
from rockpool.layers import RecLIFCurrentInJax
# - Generic parameters
net_size = 2
dt = 1e-3
w_recurrent = 2 * np.random.rand(net_size, net_size) - 1
bias = 2 * np.random.rand(net_size) - 1
tau_m = 20e-3 * np.ones(net_size)
tau_s = 20e-3 * np.ones(net_size)
# - Layer generation
fl0 = RecLIFCurrentInJax(
w_recurrent=w_recurrent,
bias=bias,
noise_std=0.1,
tau_mem=tau_m,
tau_syn=tau_s,
dt=dt,
)
# - Input signal
tsInCont = TSContinuous(times=np.arange(100),
samples=np.ones((100, net_size)))
# - Compare states and time before and after
vStateBefore = np.copy(fl0.state["Vmem"])
ts_output = fl0.evolve(tsInCont, duration=0.1)
assert fl0.t == 0.1
assert (vStateBefore != fl0.state["Vmem"]).any()
# - Test TS only evolution
fl0.reset_all()
ts_output = fl0.evolve(tsInCont)
assert fl0.t == 99
fl0.reset_all()
assert fl0.t == 0
assert (vStateBefore == fl0.state["Vmem"]).all()
# - Test that some errors are caught
with pytest.raises(AssertionError):
fl1 = RecLIFCurrentInJax(
w_recurrent=np.zeros((3, 2)),
tau_mem=np.zeros(3),
tau_syn=np.zeros(3),
bias=np.zeros(3),
)
with pytest.raises(AssertionError):
fl1 = RecLIFCurrentInJax(
w_recurrent=np.zeros((2, 2)),
tau_mem=np.zeros(3),
tau_syn=np.zeros(3),
bias=np.zeros(3),
)
with pytest.raises(AssertionError):
fl1 = RecLIFCurrentInJax(
w_recurrent=np.zeros((2, 2)),
tau_mem=np.zeros(2),
tau_syn=np.zeros(3),
bias=np.zeros(3),
)
def test_setWeightsIn():
""" Test weight setting"""
from rockpool.layers import FFIAFNest, RecIAFSpkInNest, RecAEIFSpkInNest
from rockpool import TSEvent, TSContinuous
import numpy as np
# - Generic parameters
weights_in = np.array([[-0.5, 0.02, 0.4], [0.2, -0.3, -0.15]])
weights_rec = np.random.rand(3, 3) * 0.01
bias = 0.01
tau_mem = [0.02, 0.05, 0.1]
tau_syn_exc = [0.2, 0.01, 0.01]
tau_syn_inh = tau_syn_exc
# - Different input weights for initialization of fl1
weights_in1 = np.array([[-0.1, 0.02, 0.4], [0.2, -0.3, -0.15]])
# - Input time series
tsInpCont = TSContinuous(np.arange(15) * 0.01, np.ones(15) * 0.1)
tsInp = TSEvent([0.1], [0])
## -- FFIAFNEst
# - Layer generation
fl0 = FFIAFNest(
weights=weights_in, dt=0.001, bias=bias, tau_mem=tau_mem, refractory=0.001
)
fl1 = FFIAFNest(
weights=weights_in1, dt=0.001, bias=bias, tau_mem=tau_mem, refractory=0.001
)
# - Set input weights to same as fl0
fl1.weights = weights_in
assert (fl1.weights_ == weights_in).all()
# - Compare states before and after
fl0.evolve(tsInpCont, duration=0.12)
fl1.evolve(tsInpCont, duration=0.12)
assert (fl0.state == fl1.state).all()
fl0.terminate()
fl1.terminate()
## -- RecIAFSpkInNest
# - Layer generation
fl0 = RecIAFSpkInNest(
weights_in=weights_in,
weights_rec=weights_rec,
dt=0.001,
bias=bias,
tau_mem=tau_mem,
tau_syn_exc=tau_syn_exc,
tau_syn_inh=tau_syn_inh,
refractory=0.001,
record=True,
)
fl1 = RecIAFSpkInNest(
weights_in=weights_in1,
weights_rec=weights_rec,
dt=0.001,
bias=bias,
tau_mem=tau_mem,
tau_syn_exc=tau_syn_exc,
tau_syn_inh=tau_syn_inh,
refractory=0.001,
record=True,
)
# - Set input weights to same as fl0
fl1.weights_in = weights_in
assert (fl1.weights_in_ == weights_in).all()
# - Compare states before and after
fl0.evolve(tsInp, duration=0.12)
fl1.evolve(tsInp, duration=0.12)
assert (fl0.state == fl1.state).all()
fl0.terminate()
fl1.terminate()
## -- RecAEIFSpkInNest
# - Layer generation
fl0 = RecAEIFSpkInNest(
weights_in=weights_in,
weights_rec=weights_rec,
dt=0.001,
bias=bias,
tau_mem=tau_mem,
tau_syn_exc=tau_syn_exc,
tau_syn_inh=tau_syn_inh,
refractory=0.001,
record=True,
)
fl1 = RecAEIFSpkInNest(
weights_in=weights_in1,
weights_rec=weights_rec,
dt=0.001,
bias=bias,
tau_mem=tau_mem,
tau_syn_exc=tau_syn_exc,
tau_syn_inh=tau_syn_inh,
refractory=0.001,
record=True,
)
# - Set input weights to same as fl0
fl1.weights_in[0, 0] = weights_in[0, 0]
assert (fl1.weights_in_ == weights_in).all()
# - Compare states before and after
fl0.evolve(tsInp, duration=0.12)
fl1.evolve(tsInp, duration=0.12)
assert (fl0.state == fl1.state).all()
fl0.terminate()
fl1.terminate()
def test_ff_rate_euler_train():
"""Test ridge regression for FFRateEuler"""
from rockpool import TSContinuous, TSEvent
from rockpool.layers import FFRateEuler, FFExpSyn, PassThrough
# - Layers
size_in = 6
size = 3
dt = 0.001
tau_syn = 0.15
# - FFExpSyn layer to filter spike trains
fl_exp_prepare = FFExpSyn(np.eye(size_in), dt=dt, tau_syn=tau_syn)
# - Spiking input signal
tDur = 0.01
nSpikes = 5
vnC = np.tile(np.arange(size_in),
int(np.ceil(1.0 / nSpikes * size)))[:nSpikes]
vtT = np.linspace(0, tDur, nSpikes, endpoint=False)
tsIn = TSEvent(vtT, vnC, num_channels=size_in)
# - Filter signal
ts_filtered = fl_exp_prepare.evolve(tsIn)
# - Another FFExpSyn layer to compare training
fl_exp_train = FFExpSyn(np.zeros((size_in, size)), dt=dt, tau_syn=tau_syn)
# - Rate layers to be trained
fl_rate = FFRateEuler(np.zeros((size_in, size)), dt=dt, noise_std=0)
fl_pt = PassThrough(np.zeros((size_in, size)), dt=dt, noise_std=0)
# - Target and training
tgt_samples = np.array([
np.sin(np.linspace(0, 10 * tDur, int(tDur / dt)) + fPhase)
for fPhase in np.linspace(0, 3, size)
]).T
ts_tgt = TSContinuous(np.arange(int(tDur / dt)) * dt, tgt_samples)
fl_exp_train.train_rr(ts_tgt,
tsIn,
regularize=0.1,
is_first=True,
is_last=True)
fl_rate.train_rr(ts_tgt,
ts_filtered,
regularize=0.1,
is_first=True,
is_last=True)
fl_pt.train_rr(ts_tgt,
ts_filtered,
regularize=0.1,
is_first=True,
is_last=True)
assert (
np.isclose(fl_exp_train.weights, fl_rate.weights, rtol=1e-4,
atol=1e-2).all()
and np.isclose(fl_exp_train.bias, fl_rate.bias, rtol=1e-4,
atol=1e-2).all()
and np.isclose(
fl_exp_train.weights, fl_pt.weights, rtol=1e-4, atol=1e-2).all()
and np.isclose(fl_exp_train.bias, fl_pt.bias, rtol=1e-4,
atol=1e-2).all()), "Training led to different results"
def test_continuous_methods():
from rockpool import TSContinuous
ts1 = TSContinuous([0, 1, 2], [0, 1, 2])
# - Interpolation
assert ts1(0) == 0
assert ts1(2) == 2
assert ts1(1.5) == 1.5
assert ts1._interpolate(0) == 0
assert ts1._interpolate(2) == 2
assert ts1._interpolate(1.5) == 1.5
# - Delay
ts2 = ts1.delay(1)
assert ts1.t_start == 0
assert ts2.t_start == 1
ts20 = ts1.start_at_zero()
assert ts20.t_start == 0
# - Contains
assert ts1.contains(0)
assert ~ts1.contains(-1)
assert ts1.contains([0, 1, 2])
assert ~ts1.contains([0, 1, 2, 3])
# - Resample
ts2 = ts1.resample([0.1, 1.1, 1.9])
# - Merge
ts1 = TSContinuous([0, 1, 2], [0, 1, 2])
ts2 = TSContinuous([0, 1, 2], [1, 2, 3])
ts3 = ts1.merge(ts2, remove_duplicates=True)
assert np.size(ts3.samples) == 3
assert np.size(ts1.samples) == 3
assert np.size(ts2.samples) == 3
ts3 = ts1.merge(ts2, remove_duplicates=False)
assert np.size(ts3.samples) == 6
# - Append
ts1 = TSContinuous([0, 1, 2], [0, 1, 2])
ts2 = TSContinuous([0, 1, 2], [1, 2, 3])
ts3 = ts1.append_t(ts2)
assert np.size(ts3.times) == 6
ts3 = ts1.append_c(ts2)
assert ts3.num_channels == 2
# - isempty
assert ~ts1.isempty()
assert TSContinuous().isempty()
# - clip
ts2 = ts1.clip(0.5, 1.5)
# - Min / Max
ts1 = TSContinuous([0, 1, 2], [0, 1, 2])
assert ts1.min == 0
assert ts1.max == 2
def test_continuous_inplace_mutation():
from rockpool import TSContinuous
ts1 = TSContinuous([0, 1, 2], [0, 1, 2])
# - Delay
ts1.delay(1, inplace=True)
assert ts1.t_start == 1
# - Resample
ts1.resample([0.125, 1.1, 1.9], inplace=True)
assert ts1.t_start == 0.125
# - Merge
ts1 = TSContinuous([0, 1, 2], [0, 1, 2])
ts2 = TSContinuous([0, 1, 2], [1, 2, 3])
ts1.merge(ts2, remove_duplicates=True, inplace=True)
assert np.size(ts1.samples) == 3
ts3 = ts1.merge(ts2, remove_duplicates=False, inplace=True)
assert np.size(ts3.samples) == 6
# - Append
ts1 = TSContinuous([0, 1, 2], [0, 1, 2])
ts2 = TSContinuous([0, 1, 2], [1, 2, 3])
ts1.append_t(ts2, inplace=True)
assert np.size(ts1.times) == 6
ts1 = TSContinuous([0, 1, 2], [0, 1, 2])
ts2 = TSContinuous([0, 1, 2], [1, 2, 3])
ts1.append_c(ts2, inplace=True)
assert ts1.num_channels == 2
# - clip
ts1.clip(0.5, 1.5, inplace=True)
assert ts1.t_start == 0.5
# - Start at 0
ts1.start_at_zero(inplace=True)
assert ts1.t_start == 0
def test_continuous_append_t():
"""
Test append method of TSEvent
"""
from rockpool import TSContinuous
# - Generate a few TSEvent objects
samples = np.random.randint(10, size=(2, 6))
empty_series = TSContinuous(t_start=1)
series_list = []
series_list.append(
TSContinuous([1, 2], samples[:2, :2], t_start=-1, t_stop=3))
series_list.append(TSContinuous([1], samples[0, 2:4], t_start=0, t_stop=2))
series_list.append(
TSContinuous([1, 3], samples[:2, 4:], t_start=0, t_stop=3))
# Appending two series
appended_fromtwo = series_list[0].append_t(series_list[1])
assert appended_fromtwo.t_start == -1, "Wrong t_start for appended series."
assert appended_fromtwo.t_stop == 6, "Wrong t_stop for appended series."
assert (appended_fromtwo.times == np.array(
[1, 2, 5])).all(), "Wrong time trace for appended series."
assert ((appended_fromtwo.samples[:2] == samples[:2, :2]).all() and
(appended_fromtwo.samples[2:]
== samples[[0], 2:4])).all(), "Wrong samples for appended series."
# Appending with empty series
appended_empty_first = empty_series.append_t(series_list[0])
assert (appended_empty_first.t_start == empty_series.t_start
), "Wrong t_start when appending with empty."
assert (appended_empty_first.t_stop == series_list[0].duration +
empty_series.t_start), "Wrong t_stop when appending with empty."
assert (appended_empty_first.samples == series_list[0].samples
).all(), "Wrong samples when appending with empty"
assert (appended_empty_first.times == series_list[0].times +
2).all(), "Wrong time trace when appending with empty"
appended_empty_last = series_list[0].append_t(empty_series)
assert (appended_empty_last.t_start == series_list[0].t_start
), "Wrong t_start when appending with empty."
assert (
# - 1 is offset, which is np.median(np.diff(series_list[0].times))=1
appended_empty_last.t_stop == series_list[0].t_stop +
empty_series.duration), "Wrong t_stop when appending with empty."
assert (appended_empty_last.samples == series_list[0].samples
).all(), "Wrong samples when appending with empty"
assert (appended_empty_last.times == series_list[0].times
).all(), "Wrong time trace when appending with empty"
# Appending multiple time series
appended_fromthree = series_list[0].append_t(series_list[1:], offset=None)
exptd_offset = np.median(np.diff(series_list[0].times))
exptd_ts2_delay = exptd_offset + series_list[0].t_stop - series_list[
1].t_start
# - No offset between ts2 and ts3 because ts2 has only one element
exptd_ts3_delay = exptd_ts2_delay + series_list[1].t_stop - series_list[
2].t_start
exptd_ts2_times = series_list[1].times + exptd_ts2_delay
exptd_ts3_times = series_list[2].times + exptd_ts3_delay
assert (appended_fromthree.times == np.r_[series_list[0].times,
exptd_ts2_times, exptd_ts3_times]
).all(), "Wrong time trace when appending from list"
assert (appended_fromthree.samples == np.vstack([
series.samples for series in series_list
])).all(), "Wrong samples when appending from list"
# - Generating from list of TSContinuous
appended_fromlist = TSContinuous.concatenate_t(series_list)
assert (appended_fromthree.times == appended_fromlist.times
).all(), "Wrong time trace when appending from list"
assert (appended_fromthree.samples == appended_fromlist.samples
).all(), "Wrong samples when appending from list"