[FIX] dss_line() with nremove>1

meegkit/dss.py

@@ -2,9 +2,9 @@
 import numpy as np
 from scipy import linalg
 
-from .utils import (demean, mean_over_trials, smooth, pca, theshapeof,
-                    tscov, gaussfilt, wpwr)
 from .tspca import tsr
+from .utils import (demean, gaussfilt, mean_over_trials, pca, smooth,
+                    theshapeof, tscov, wpwr)
 
 
 def dss1(data, weights=None, keep1=None, keep2=1e-12):
@@ -175,7 +175,7 @@ def dss_line(x, fline, sfreq, nremove=1, nfft=1024, nkeep=None, show=False):
     if x.shape[0] < nfft:
         print('reducing nfft to {}'.format(x.shape[0]))
         nfft = x.shape[0]
-
+    n_samples, n_chans, n_trials = theshapeof(x)
     x = demean(x)
 
     # cancels line_frequency and harmonics, light lowpass
@@ -208,7 +208,12 @@ def dss_line(x, fline, sfreq, nremove=1, nfft=1024, nkeep=None, show=False):
         plt.show()
 
     idx_remove = np.arange(nremove)
-    xxxx = xxxx @ todss[:, idx_remove]  # line-dominated components
+    if x.ndim == 3:
+        for t in range(n_trials):  # line-dominated components
+            xxxx[..., t] = xxxx[..., t] @ todss[:, idx_remove]
+    elif x.ndim == 2:
+        xxxx = xxxx @ todss[:, idx_remove]
+
     xxx, _, _, _ = tsr(xxx, xxxx)  # project them out
 
     # reconstruct clean signal

meegkit/tspca.py

@@ -120,7 +120,6 @@ def tsr(X, R, shifts=None, wX=None, wR=None, keep=None, thresh=1e-12):
     offset1 = np.max((0, -np.min(shifts)))
     idx = np.arange(offset1, X.shape[0])
     # X = X[idx, ...]
-
     # if len(wX) > 0:
     #     wX = wX[idx, ...]
     # if len(wR) > 0:
@@ -130,10 +129,9 @@ def tsr(X, R, shifts=None, wX=None, wR=None, keep=None, thresh=1e-12):
 
     # adjust size of X
     offset2 = np.max((0, np.max(shifts)))
-    idx = np.arange(X.shape[0]) - offset2
-    idx = idx[idx >= 0]
+    # idx = np.arange(X.shape[0]) - offset2
+    # idx = idx[idx >= 0]
     # X = X[idx, ...]
-
     # if len(wX) > 0:
     #     wX = wX[idx, ...]
 
@@ -157,21 +155,21 @@ def tsr(X, R, shifts=None, wX=None, wR=None, keep=None, thresh=1e-12):
             weights[..., t] = wr
 
     wX = weights
-    wR = np.zeros((n_samples_R, 1, n_trials_R))
     wR = weights
 
     # remove weighted means
     X, mean1 = demean(X, wX, return_mean=True)
     R = demean(R, wR)
 
     # equalize power of R channels, the equalize power of the R PCs
-    if R.shape[1] > 1:
-        R = normcol(R, wR)
-        C, _ = tscov(R, [])
-        V, _ = pca(C, thresh=1e-6)
-        R = R * V
-
+    # if R.shape[1] > 1:
     R = normcol(R, wR)
+    C, _ = tscov(R)
+    V, _ = pca(C, thresh=1e-6)
+    z = np.zeros((n_samples_X, V.shape[1], n_trials_R))
+    for t in range(n_trials_R):
+        z[..., t] = R[..., t] @ V
+    R = normcol(z, wR)
 
     # covariances and cross-covariance with time-shifted refs
     Cr, twcr = tscov(R, shifts, wR)
@@ -182,10 +180,10 @@ def tsr(X, R, shifts=None, wX=None, wR=None, keep=None, thresh=1e-12):
 
     # TSPCA: clean x by removing regression on time-shifted refs
     y = np.zeros((n_samples_X, n_chans_X, n_trials_X))
-    for trial in np.arange(n_trials_X):
-        r = multishift(R[..., trial], shifts, reshape=True)
+    for t in np.arange(n_trials_X):
+        r = multishift(R[..., t], shifts, reshape=True)
         z = r @ regression
-        y[..., trial] = X[:z.shape[0], :, trial] - z
+        y[..., t] = X[:z.shape[0], :, t] - z
 
     y, mean2 = demean(y, wX, return_mean=True)
 

meegkit/utils/matrix.py

@@ -553,8 +553,8 @@ def normcol(X, weights=None, return_norm=False):
     """
     if X.ndim == 3:
         n_samples, n_chans, n_trials = theshapeof(X)
-        X = unfold(X)
         weights = _check_weights(weights, X)
+        X = unfold(X)
         if not weights.any():  # no weights
             X_norm, N = normcol(X, return_norm=True)
             N = N ** 2
@@ -568,9 +568,6 @@ def normcol(X, weights=None, return_norm=False):
             if weights.ndim == 2 and weights.shape[1] == 1:
                 weights = np.tile(weights, (1, n_samples, n_trials))
 
-            if weights.shape != weights.shape:
-                raise ValueError("Weight array should have be same shape as X")
-
             weights = unfold(weights)
             X_norm, N = normcol(X, weights, return_norm=True)
             N = N ** 2

meegkit/utils/sig.py

@@ -399,7 +399,11 @@ def gaussfilt(data, srate, f, fwhm, n_harm=1, shift=0, return_empvals=False,
 
     # filter
     tmp = np.fft.fft(data, axis=0)
-    tmp *= fx[:, None]
+    if data.ndim == 2:
+        tmp *= fx[:, None]
+    elif data.ndim == 3:
+        tmp *= fx[:, None, None]
+
     filtdat = 2 * np.real(np.fft.ifft(tmp, axis=0))
 
     if return_empvals or show:

tests/test_dss.py

@@ -73,22 +73,38 @@ def test_dss_line():
     artifact = artifact ** 3
     s = x + 10 * artifact
 
+    def _plot(x):
+        f, ax = plt.subplots(1, 2, sharey=True)
+        f, Pxx = signal.welch(x, sr, nperseg=1024, axis=0,
+                              return_onesided=True)
+        ax[1].semilogy(f, Pxx)
+        f, Pxx = signal.welch(s, sr, nperseg=1024, axis=0,
+                              return_onesided=True)
+        ax[0].semilogy(f, Pxx)
+        ax[0].set_xlabel('frequency [Hz]')
+        ax[1].set_xlabel('frequency [Hz]')
+        ax[0].set_ylabel('PSD [V**2/Hz]')
+        ax[0].set_title('before')
+        ax[1].set_title('after')
+        plt.show()
+
+    # 2D case, n_outputs == 1
     out, _ = dss.dss_line(s, 20, sr)
-    f, ax = plt.subplots(1, 2, sharey=True)
-    print(out.shape)
-    f, Pxx = signal.welch(out, sr, nperseg=1024, axis=0,
-                          return_onesided=True)
-    ax[1].semilogy(f, Pxx)
-    f, Pxx = signal.welch(s, sr, nperseg=1024, axis=0,
-                          return_onesided=True)
-    ax[0].semilogy(f, Pxx)
-    ax[0].set_xlabel('frequency [Hz]')
-    ax[1].set_xlabel('frequency [Hz]')
-    ax[0].set_ylabel('PSD [V**2/Hz]')
-    ax[0].set_title('before')
-    ax[1].set_title('after')
-    plt.show()
+    _plot(out)
+
+    # Test n_outputs > 1
+    out, _ = dss.dss_line(s, 20, sr, nremove=2)
+    # _plot(out)
+
+    # Test n_trials > 1
+    x = np.random.randn(nsamples, nchans, 4)
+    artifact = np.sin(np.arange(nsamples) / sr * 2 * np.pi * 20)[:, None, None]
+    artifact[artifact < 0] = 0
+    artifact = artifact ** 3
+    s = x + 10 * artifact
+    out, _ = dss.dss_line(s, 20, sr, nremove=1)
+
 
 if __name__ == '__main__':
-    pytest.main([__file__])
-    # test_dss_line()
+    # pytest.main([__file__])
+    test_dss_line()

tests/test_tspca.py

@@ -64,39 +64,46 @@ def test_tsr(show=True):
     # artifact + harmonics
     artifact = np.sin(np.arange(nsamples) / sr * 2 * np.pi * 10)[:, None]
     artifact[artifact < 0] = 0
-    artifact = artifact ** 3
-    signal = x + 10 * artifact
+    artifact = 5 * artifact ** 3
+    signal = x + artifact
+
+    signal -= np.mean(signal, keepdims=True)
+    artifact -= np.mean(artifact, keepdims=True)
+
     # Without shifts
     y, idx, mean_total, weights = tspca.tsr(
         signal,
         artifact,
         shifts=[0])
 
     if show:
-        f, ax = plt.subplots(2, 1)
-        ax[0].plot(y[:500, 0], 'grey', label='cleaned')
-        ax[0].plot(x[:500, 0], ':', label='signal')
-        ax[1].plot((y - x)[:500], label='cleaned - signal')
+        f, ax = plt.subplots(2, 1, num='without shifts')
+        ax[0].plot(y[:500, 0], 'grey', label='recovered signal')
+        ax[0].plot(x[:500, 0], ':', label='real signal')
+        ax[1].plot((y - x)[:500], label='residual')
         ax[0].legend()
         ax[1].legend()
         # plt.show()
 
     # Test residual almost 0.0
     np.testing.assert_almost_equal(y - x, np.zeros_like(y), decimal=1)
 
-    # With shifts
+    # With shifts. We slide the input array by one sample, and check that the
+    # artifact is successfully regressed.
     y, idx, mean_total, weights = tspca.tsr(
-        signal + artifact,
+        signal,
         np.roll(artifact, 1, axis=0),
         shifts=[-1, 0, 1])
 
     if show:
-        f, ax = plt.subplots(2, 1)
-        ax[0].plot(y[:500, 0], 'grey', label='cleaned')
-        ax[0].plot(x[:500, 0], ':', label='signal')
-        ax[1].plot((y - x)[:500, 0], label='cleaned - signal')
+        f, ax = plt.subplots(3, 1, num='with shifts')
+        ax[0].plot(signal[:500], label='signal + noise')
+        ax[1].plot(x[:500, 0], 'grey', label='real signal')
+        ax[1].plot(y[:500, 0], ':', label='recovered signal')
+        ax[2].plot((signal - y)[:500, 0], label='before - after')
         ax[0].legend()
         ax[1].legend()
+        ax[2].legend()
         plt.show()
 
 if __name__ == '__main__':