def test_get_errors_mse():
mse = MSEMetric()
errors = mse.get_errors(np.random.rand(5, 10))
assert_equal(np.shape(errors), (5, ))
assert (np.all(np.array(errors) > 0))
errors = mse.get_errors(np.zeros((2, 10)))
assert_equal(np.shape(errors), (2, ))
assert_equal(errors, [0., 0.])
def test_calc_mse_t_start():
mse = MSEMetric(t_start=1 * ms)
out = np.random.rand(2, 20)
inp = np.random.rand(5, 2, 20)
errors = mse.calc(inp, out, 0.1 * ms)
assert_equal(np.shape(errors), (5, ))
assert (np.all(errors > 0))
# Everything before 1ms should be ignored, so having the same values for
# the rest should give an error of 0
inp[:, :, 10:] = out[None, :, 10:]
assert_equal(mse.calc(inp, out, 0.1 * ms), np.zeros(5))
def test_get_features_mse():
mse = MSEMetric()
out_mse = np.random.rand(2, 20)
inp_mse = np.random.rand(5, 2, 20)
features = mse.get_features(inp_mse, out_mse, 0.1 * ms)
assert_equal(np.shape(features), (5, 2))
assert (np.all(np.array(features) > 0))
features = mse.get_features(np.tile(out_mse, (5, 1, 1)), out_mse, 0.1 * ms)
assert_equal(np.shape(features), (5, 2))
assert_equal(features, np.zeros((5, 2)))
def test_spikefitter_fit_errors(setup):
dt, sf = setup
with pytest.raises(TypeError):
results, errors = sf.fit(n_rounds=2,
optimizer=n_opt,
metric=MSEMetric(),
gL=[20*nS, 40*nS],
C=[0.5*nF, 1.5*nF])
with pytest.raises(TypeError):
results, errors = sf.fit(n_rounds=2,
optimizer=None,
metric=MSEMetric(),
gL=[20*nS, 40*nS],
C=[0.5*nF, 1.5*nF])
def test_calc_mse_t_weights():
with pytest.raises(ValueError):
# t_start and t_weights
MSEMetric(t_start=1 * ms, t_weights=np.ones(20))
with pytest.raises(ValueError):
# all values zero
MSEMetric(t_weights=np.zeros(20))
with pytest.raises(ValueError):
# negative values
weights = np.ones(20)
weights[17] = -1
MSEMetric(t_weights=weights)
weights = np.ones(20)
weights[:10] = 0
mse = MSEMetric(t_weights=weights)
out = np.random.rand(2, 20)
inp = np.random.rand(5, 2, 20)
errors = mse.calc(inp, out, 0.1 * ms)
assert_equal(np.shape(errors), (5, ))
assert (np.all(errors > 0))
# Everything before 1ms should be ignored, so having the same values for
# the rest should give an error of 0
inp[:, :, 10:] = out[None, :, 10:]
assert_equal(mse.calc(inp, out, 0.1 * ms), np.zeros(5))
def test_fitter_fit_tstart(setup_constant):
dt, tf = setup_constant
# Ignore the first 50 steps at 10mV
params, result = tf.fit(n_rounds=10, optimizer=n_opt,
metric=MSEMetric(t_start=50*dt),
c=[0 * mV, 30 * mV])
# Fit should be close to 20mV
assert np.abs(params['c'] - 20*mV) < 1*mV
def test_fitter_fit_tsteps(setup_constant):
dt, tf = setup_constant
with pytest.raises(ValueError):
# Incorrect weight size
tf.fit(n_rounds=10,
optimizer=n_opt,
metric=MSEMetric(t_weights=np.ones(101)),
c=[0 * mV, 30 * mV])
# Ignore the first 50 steps at 10mV
weights = np.ones(100)
weights[:50] = 0
params, result = tf.fit(n_rounds=10,
optimizer=n_opt,
metric=MSEMetric(t_weights=weights),
c=[0 * mV, 30 * mV])
# Fit should be close to 20mV
assert np.abs(params['c'] - 20 * mV) < 1 * mV
def test_fitter_refine_reuse_tstart(setup_constant):
dt, tf = setup_constant
# Ignore the first 50 steps at 10mV but do not actually fit (0 rounds)
params, result = tf.fit(n_rounds=0, optimizer=n_opt,
metric=MSEMetric(t_start=50*dt),
c=[0 * mV, 30 * mV])
# t_start should be reused
params, result = tf.refine({'c': 5 * mV}, c=[0 * mV, 30 * mV])
# Fit should be close to 20mV
assert np.abs(params['c'] - 20 * mV) < 1 * mV
def test_fitter_refine_tsteps_normalization(setup_constant):
dt, tf = setup_constant
model_traces = tf.generate(params={'c': 5 * mV})
mse_error = MSEMetric(t_start=50 * dt).calc(model_traces[None, :, :],
tf.output, dt)
all_errors = []
def callback(parameters, errors, best_parameters, best_error, index):
all_errors.append(float(errors[0]))
return True # stop simulation
# Ignore the first 50 steps at 10mV
tf.refine({'c': 5 * mV},
c=[0 * mV, 30 * mV],
t_start=50 * dt,
callback=callback)
weights = np.ones(100)
weights[:50] = 0
tf.refine({'c': 5 * mV},
c=[0 * mV, 30 * mV],
t_weights=weights,
callback=callback)
tf.refine({'c': 5 * mV},
c=[0 * mV, 30 * mV],
t_weights=weights * 2,
callback=callback)
tf.fit(n_rounds=0,
optimizer=n_opt,
metric=MSEMetric(t_weights=weights * 3),
c=[0 * mV, 30 * mV])
tf.refine({'c': 5 * mV}, c=[0 * mV, 30 * mV], callback=callback)
assert_almost_equal(float(mse_error[0]), all_errors[0])
assert_almost_equal(all_errors[0], all_errors[1])
assert_almost_equal(all_errors[1], all_errors[2])
assert_almost_equal(all_errors[2], all_errors[3])
def test_calc_mse():
mse = MSEMetric()
out = np.random.rand(2, 20)
inp = np.random.rand(5, 2, 20)
errors = mse.calc(inp, out, 0.01 * ms)
assert_equal(np.shape(errors), (5, ))
assert_equal(mse.calc(np.tile(out, (5, 1, 1)), out, 0.1 * ms), np.zeros(5))
assert (np.all(mse.calc(inp, out, 0.1 * ms) > 0))
inp = np.vstack([np.ones((1, 3, 10)), np.zeros((1, 3, 10))])
out = np.ones((3, 10))
errors = mse.calc(inp, out, 0.01 * ms)
assert_equal(errors, [0, 1])
mse = MSEMetric(normalization=1 / 2)
errors = mse.calc(inp, out, 0.01 * ms)
# The normalization factor scales the traces, so the squared error scales
# with the square of the normalization factor
assert_equal(errors, [0, 4])
def test_calc_mse_t_weights_normalization():
# check that normalization works correctly
dt = 0.1 * ms
metric1 = MSEMetric(t_start=50 * dt)
weights = np.ones(100)
weights[:50] = 0
metric2 = MSEMetric(t_weights=weights)
weights2 = weights * 2 # should not make any difference
metric3 = MSEMetric(t_weights=weights2)
data_traces = np.random.rand(3, 100)
model_traces = np.random.rand(2, 3, 100)
error_1 = metric1.calc(model_traces=model_traces,
data_traces=data_traces,
dt=dt)
error_2 = metric2.calc(model_traces=model_traces,
data_traces=data_traces,
dt=dt)
error_3 = metric3.calc(model_traces=model_traces,
data_traces=data_traces,
dt=dt)
assert_almost_equal(error_1, error_2)
assert_almost_equal(error_1, error_3)
def test_init():
MSEMetric()
GammaFactor(10 * ms, time=10 * ms)
I = g*(v-E) : amp
g : siemens (constant)
'''
constant_model = Equations('''
v = c + x: volt
c : volt (constant)''')
all_constant_model = Equations('''
v = 10*mV + x: volt
c : volt (constant)
penalty_fixed = 10*mV**2: volt**2
penalty_wrong_unit = 10*mV : volt''')
n_opt = NevergradOptimizer()
metric = MSEMetric()
@pytest.fixture
def setup(request):
dt = 0.01 * ms
tf = TraceFitter(dt=dt,
model=model,
input_var='v',
output_var='I',
input=input_traces,
output=output_traces,
n_samples=30)
def fin():
reinit_devices()
