Comparison failure in text_axes:test_phase_spectrum_freqs

[mdboom]

In #5307, where I am lowering our testing tolerance down to zero, I am discovering a few recent changes that slipped through because the tolerance on our tests was too high.

This test has been broken since at least v1.4.0 (before which I can't build with current Numpys).

BASELINE RESULT:

ACTUAL RESULT:

Test code:

@image_comparison(baseline_images=['phase_spectrum_freqs'],
                  remove_text=True,
                  extensions=['png'])
def test_phase_spectrum_freqs():
    '''test axes.phase_spectrum with sinusoidal stimuli'''
    n = 10000
    Fs = 100.

    fstims1 = [Fs/4, Fs/5, Fs/11]

    NFFT = int(1000 * Fs / min(fstims1))
    pad_to = int(2 ** np.ceil(np.log2(NFFT)))

    x = np.arange(0, n, 1/Fs)

    y = np.zeros(x.size)
    for i, fstim1 in enumerate(fstims1):
        y += np.sin(fstim1 * x * np.pi * 2) * 10**i
    y = y

    fig = plt.figure()
    ax1 = fig.add_subplot(3, 1, 1)
    ax2 = fig.add_subplot(3, 1, 2)
    ax3 = fig.add_subplot(3, 1, 3)

    spec1, freqs1, line1 = ax1.phase_spectrum(y, Fs=Fs, pad_to=pad_to,
                                              sides='default')
    spec2, freqs2, line2 = ax2.phase_spectrum(y, Fs=Fs, pad_to=pad_to,
                                              sides='onesided')
    spec3, freqs3, line3 = ax3.phase_spectrum(y, Fs=Fs, pad_to=pad_to,
                                              sides='twosided')

    ax1.set_xlabel('')
    ax2.set_xlabel('')
    ax3.set_xlabel('')
    ax1.set_ylabel('')
    ax2.set_ylabel('')
    ax3.set_ylabel('')

@toddrjen: You appear to have written the test. Any thoughts?

[dopplershift]

Looks to me like the original result had the phase wrap around pi, and that's not in the current result. Hard to tell without ticks though,

[mdboom]

Just an update: I thought this might be related to a change in numpy, since that's where most of the actual calculation happens. However, going back to 1.7 (the earliest that matplotlib head compiles with), the result is all the same.

I'm leaning on just updating the baseline image to reflect the new result if I don't hear otherwise.

[dopplershift]

No argument here. Minus the wrap (which is going to depend on implementation details of the calculation), the images do appear to match still.

[tacaswell]

👍 to updating the images. This is probably an API change in mlab, but no one has complained about that yet so we should probably let it go.

[QuLogic]

I tried out c89ce08 where the test image was added (against NumPy 1.8.2), but it is also incorrect there. In fact, I went back a little further to 7fc90de where the phase_spectrum plotting method was added and it also does not produce the baseline image.

[QuLogic]

It also fails against NumPy 1.6.2, since you thought it might be related.

[mdboom]

Ok. All signs point to just updating the baseline image on this one then. Thanks for the extra investigation. I'll probably just update the image as part of #5307.

[mdboom]

Closed by #5307.