Note

This documentation is for a development version. Click here for the latest stable release (v1.0.0).

DVS from file

This example demonstrates how to load pre-recorded Dynamic Vision Sensor (DVS) event data from a file.

import numpy as np
import matplotlib.pyplot as plt
from IPython.display import HTML
from matplotlib.animation import ArtistAnimation

%matplotlib inline

import nengo
import nengo_loihi

# All NengoLoihi models should call this before model construction
nengo_loihi.set_defaults()

rng = np.random.RandomState(0)

Generate synthetic data

Rather than using real DVS data, we will generate some synthetic data and save it in a .events file. In most applications, this will not be necessary, since you will already have a .events or .aedat file from a real DVS camera.

def jitter_time(n, t, jitter, rng, dtype="<u4"):
    assert jitter >= 0
    assert t - jitter >= 0
    tt = (t - jitter) * np.ones(n, dtype=dtype)
    if jitter > 0:
        tt += rng.randint(0, 2 * jitter + 1, size=tt.shape, dtype=dtype)
    return tt


# the height and width of the DVS sensor
dvs_height = 180
dvs_width = 240

# our timestep in microseconds (μs)
dt_us = 1000

# the maximum amount by which to jitter spikes around the timestep (in microseconds)
t_jitter_us = 100

assert t_jitter_us < dt_us // 2

# the length of time to generate data for, in seconds and in microseconds
t_length = 1.0
t_length_us = int(1e6 * t_length)

# the maximum rate of input spikes (per pixel)
max_rate = 10
max_prob = max_rate * 1e-6 * dt_us

# the period of the sine wave, in pixels
period = 120

# these functions control the angle (theta) and phase of the sine wave over time
theta_fn = lambda t: 1
phase_fn = lambda t: 10 * t

X, Y = np.meshgrid(np.linspace(-1, 1, dvs_width), np.linspace(-1, 1, dvs_height))

events = []
for t_us in range(dt_us, t_length_us + 1, dt_us):
    t = t_us * 1e-6
    theta = theta_fn(t)
    phase = phase_fn(t)

    X1 = np.cos(theta) * X + np.sin(theta) * Y

    x = np.linspace(-1.5, 1.5, 50)
    prob = np.sin((np.pi * dvs_height / period) * x + phase) * max_prob
    prob = np.interp(X1, x, prob)

    u = rng.rand(*prob.shape)
    s_on = u < prob
    s_off = u < -prob

    y, x = s_off.nonzero()
    tt = jitter_time(len(x), t_us, t_jitter_us, rng, dtype="<u4")
    events.append((tt, 0, x, y))

    y, x = s_on.nonzero()
    tt = jitter_time(len(x), t_us, t_jitter_us, rng, dtype="<u4")
    events.append((tt, 1, x, y))

dvs_events = nengo_loihi.dvs.DVSEvents()
dvs_events.init_events(n_events=sum(len(xx) for _, _, xx, _ in events))

i = 0
for tt, p, xx, yy in events:
    ee = dvs_events.events[i : i + len(xx)]
    ee["t"] = tt
    ee["p"] = p
    ee["x"] = xx
    ee["y"] = yy
    i += len(xx)

events_file_name = "dvs-from-file-events.events"
dvs_events.write_file(events_file_name)
print("Wrote %r" % events_file_name)
del dvs_events

Wrote 'dvs-from-file-events.events'

We can view the data by using the DVSEvents class to load the events, group the events into frames, and then make the frames into a video with the help of Matplotlib’s animation support.

dvs_events = nengo_loihi.dvs.DVSEvents.from_file(events_file_name)

dt_frame_us = 20e3
t_frames = dt_frame_us * np.arange(int(round(t_length_us / dt_frame_us)))

fig = plt.figure()
imgs = []
for t_frame in t_frames:
    t0_us = t_frame
    t1_us = t_frame + dt_frame_us
    t = dvs_events.events[:]["t"]
    m = (t >= t0_us) & (t < t1_us)
    events_m = dvs_events.events[m]

    # show "off" (0) events as -1 and "on" (1) events as +1
    events_sign = 2.0 * events_m["p"] - 1

    frame_img = np.zeros((dvs_height, dvs_width))
    frame_img[events_m["y"], events_m["x"]] = events_sign

    img = plt.imshow(frame_img, vmin=-1, vmax=1, animated=True)
    imgs.append([img])

del dvs_events

ani = ArtistAnimation(fig, imgs, interval=50, blit=True)
HTML(ani.to_html5_video())