def main(session_id):
ctx = drawing.Context(2000, 1000)
for sensor in [2, 3, 4]:
result = db.branches.find({'session': session_id, 'sensor': sensor}).sort([('t', ASCENDING)])
if not result.count():
continue
result = list(result)
ts = [r['t'] for r in result]
xs = [r['sample'][0] for r in result]
ys = [r['sample'][1] for r in result]
zs = [r['sample'][2] for r in result]
rms = [r['sample'][3] for r in result]
duration = ts[-1] - ts[0]
SAMPLING_RATE = 50 # hz
x_signal = (sp.resample(ts, xs, duration * SAMPLING_RATE) - RANGE[0]) / (RANGE[1] - RANGE[0])
y_signal = (sp.resample(ts, ys, duration * SAMPLING_RATE) - RANGE[0]) / (RANGE[1] - RANGE[0])
z_signal = (sp.resample(ts, zs, duration * SAMPLING_RATE) - RANGE[0]) / (RANGE[1] - RANGE[0])
rms_signal = (sp.resample(ts, rms, duration * SAMPLING_RATE) - RANGE[0]) / (RANGE[1] - RANGE[0])
# ctx.plot(x_signal, stroke=(1.0, 0.0, 0.0, 1.0), thickness=2.0)
# ctx.plot(y_signal, stroke=(0.0, 1.0, 0.0, 1.0), thickness=2.0)
# ctx.plot(z_signal, stroke=(0.0, 0.0, 1.0, 1.0), thickness=2.0)
rms_signal = sp.normalize(rms_signal)
ctx.plot(rms_signal, stroke=colors[sensor], thickness=2.0)
ctx.output("graphs")
def run(self):
spectrum_sum = None
num_spectrums = 0
while True:
# filename = "robin_chat_sample_11k_16_mono.wav"
# sound = Sound().load(filename)
sound = Sound().record(10, keep_file=False)
block_size = 512
block_overlap = block_size / 2 # power of two, default is 128
spectrum, freqs, ts, image = plt.specgram(sound.signal, NFFT=block_size, Fs=sound.rate, noverlap=block_overlap)
log.info("--> freq bins %s" % len(freqs))
log.info("--> time columns %s" % len(ts))
spectrum = sp.normalize(np.sqrt(spectrum), 0.0, 100.0) # sqrt compresses, good for power. 200 is a clipping threshold.
spectrum = sp.rescale(spectrum, 0, 255)
s = np.sum(spectrum)
log.info("--> SUM %s" % s)
if s < THRESHOLD:
continue
num_spectrums += 1
if spectrum_sum is None:
spectrum_sum = spectrum
else:
sum_multiplier = (num_spectrums - 1) / num_spectrums
spectrum_sum *= sum_multiplier
add_multiplier = 1 / num_spectrums
spectrum *= add_multiplier
spectrum_sum += spectrum
ds = np.copy(spectrum_sum).astype(np.uint32) # convert to uint32
ds = ((ds & 0xFF) <<8) + ((ds & 0xFF) <<16) + ((ds & 0xFF)) # make white
# print(ds[80])
display_spectrums.put(ds)
def process_readings(kind, color, thickness=5):
global t, db, ctx
data = model.fetch_readings(kind, t - DURATION, t)
if not len(data):
return False
data.sort(key=lambda d: d['t'])
ts = [d['t'] for d in data]
vs = [d['v'] for d in data]
signal = sp.resample(ts, vs, DURATION)
signal = sp.normalize(signal, RANGE[kind][0], RANGE[kind][1])
db[kind] = list(signal)
if config['draw']:
ctx.line([(float(i) / DURATION, sample) for (i, sample) in enumerate(signal)], thickness=thickness, stroke=color)
def run(self):
while True:
# filename = "robin_chat_sample_11k_16_mono.wav"
# sound = Sound().load(filename)
sound = Sound().record(10)
block_size = 512
block_overlap = block_size / 2 # power of two, default is 128
spectrum, freqs, ts, image = plt.specgram(sound.signal, NFFT=block_size, Fs=sound.rate, noverlap=block_overlap)
log.info("--> freq bins %s" % len(freqs))
log.info("--> time columns %s" % len(ts))
spectrum = sp.normalize(np.sqrt(spectrum), 0.0, 100.0) # sqrt compresses, good for power. 200 is a clipping threshold.
spectrum = sp.rescale(spectrum, 0, 255).astype(np.uint32) # convert to uint32
spectrum = ((spectrum & 0xFF) <<8) + ((spectrum & 0xFF) <<16) + ((spectrum & 0xFF))
print(spectrum[80])
spectrums.put(spectrum)
def get_data(hidden=False):
log.debug("HIDDEN %s" % hidden)
data = {}
try:
walks = model.fetch_walks(hidden=hidden)
notes = []
v = 0
ids = []
for walk in walks:
sequence = model.fetch_sequence(walk['id'])
if len(sequence) < config['min_steps']:
continue
for step in sequence: #[:config['max_steps']]:
notes.append((step[0], v, 0 if step[1] == 'left' else 1))
v += 1
ids.append(walk['id'])
# sort and normalize onsets
notes.sort(key=lambda x: x[0])
onsets = [note[0] for note in notes]
onsets = sp.normalize(onsets)
notes = [(onsets[i], note[1], note[2])
for (i, note) in enumerate(notes)]
log.info("NOTES %s" % len(notes))
data['notes'] = notes
data['walk_ids'] = ids
except Exception as e:
log.error(log.exc(e))
return {}
return json.dumps(data)
def get_data(hidden=False):
log.debug("HIDDEN %s" % hidden)
data = {}
try:
walks = model.fetch_walks(hidden=hidden)
notes = []
v = 0
ids = []
for walk in walks:
sequence = model.fetch_sequence(walk['id'])
if len(sequence) < config['min_steps']:
continue
for step in sequence:#[:config['max_steps']]:
notes.append((step[0], v, 0 if step[1] == 'left' else 1))
v += 1
ids.append(walk['id'])
# sort and normalize onsets
notes.sort(key=lambda x: x[0])
onsets = [note[0] for note in notes]
onsets = sp.normalize(onsets)
notes = [(onsets[i], note[1], note[2]) for (i, note) in enumerate(notes)]
log.info("NOTES %s" % len(notes))
data['notes'] = notes
data['walk_ids'] = ids
except Exception as e:
log.error(log.exc(e))
return {}
return json.dumps(data)
def main(session_id):
ctx = drawing.Context(2000, 1000)
for sensor in [2, 3, 4]:
result = db.branches.find({
'session': session_id,
'sensor': sensor
}).sort([('t', ASCENDING)])
if not result.count():
continue
result = list(result)
ts = [r['t'] for r in result]
xs = [r['sample'][0] for r in result]
ys = [r['sample'][1] for r in result]
zs = [r['sample'][2] for r in result]
rms = [r['sample'][3] for r in result]
duration = ts[-1] - ts[0]
SAMPLING_RATE = 50 # hz
x_signal = (sp.resample(ts, xs, duration * SAMPLING_RATE) -
RANGE[0]) / (RANGE[1] - RANGE[0])
y_signal = (sp.resample(ts, ys, duration * SAMPLING_RATE) -
RANGE[0]) / (RANGE[1] - RANGE[0])
z_signal = (sp.resample(ts, zs, duration * SAMPLING_RATE) -
RANGE[0]) / (RANGE[1] - RANGE[0])
rms_signal = (sp.resample(ts, rms, duration * SAMPLING_RATE) -
RANGE[0]) / (RANGE[1] - RANGE[0])
# ctx.plot(x_signal, stroke=(1.0, 0.0, 0.0, 1.0), thickness=2.0)
# ctx.plot(y_signal, stroke=(0.0, 1.0, 0.0, 1.0), thickness=2.0)
# ctx.plot(z_signal, stroke=(0.0, 0.0, 1.0, 1.0), thickness=2.0)
rms_signal = sp.normalize(rms_signal)
ctx.plot(rms_signal, stroke=colors[sensor], thickness=2.0)
ctx.output("graphs")
def generate():
# load data into t and count arrays per species
species = OrderedDict()
start_t = util.timestamp(util.parse_date(str(config['start'])))
end_t = util.timestamp(util.parse_date(str(config['end'])))
max_count = 0
with open("data.csv") as f:
data = csv.reader(f)
for r, row in enumerate(data):
if r == 0:
continue
plot = row[1]
name = row[2]
if len(config['species_list']
) and name not in config['species_list']:
continue
dt = datetime.datetime(int(row[3]), 1,
1) + datetime.timedelta(int(row[4]) - 1)
t = util.timestamp(dt)
if t < start_t or t > end_t:
continue
count = 0 if row[5] == "NA" else int(row[5])
if count > max_count:
max_count = count
if name not in species:
species[name] = {'ts': [start_t, t - 1], 'counts': [0, 0]}
species[name]['ts'].append(t)
species[name]['counts'].append(count)
species = OrderedDict(sorted(species.items()))
print("--> loaded")
# add a zero count at the start and end of every year
yts = [
util.timestamp(datetime.datetime(y, 1, 1)) for y in range(1974, 2017)
]
for name in species:
ts = species[name]['ts']
for yt in yts:
i = 0
while i < len(ts) and ts[i] < yt:
i += 1
if i > 0:
end_season_t = ts[i - 1]
if i < len(ts):
start_season_t = ts[i]
ts.insert(i, start_season_t - config['tail'])
species[name]['counts'].insert(i, 0)
ts.insert(i, end_season_t + config['tail'])
species[name]['counts'].insert(i, 0)
species[name]['ts'].append(end_t)
species[name]['counts'].append(0)
print("--> onsets added")
# create and draw signals
signals = []
names = []
i = 0
for name, data in species.items():
print("Processing %s..." % name)
# create signal from bloom counts
signal = sp.resample(data['ts'], data['counts'])
if config['normalize']:
signal = sp.normalize(signal)
else:
signal = sp.normalize(signal, 0, max_count)
signal = sp.smooth(signal, size=8)
signal = sp.limit(
signal,
max(signal)) # get rid of noise below 0 for onset detection
# add spikes for peaks
if config['peak_spikes']:
peaks, valleys = sp.detect_peaks(signal, lookahead=50)
peak_signal = np.zeros(len(signal))
for peak in peaks:
peak_signal[peak[0]] = 1.0
signal += peak_signal
# add spikes for onsets
if config['onset_spikes']:
onsets = sp.detect_onsets(signal)
onset_signal = np.zeros(len(signal))
for onset in onsets:
onset_signal[onset] = 0.5
onset_signal[onset + 1] = 0.4
onset_signal[onset + 2] = 0.25
signal += onset_signal
# limit
signal = sp.limit(signal, 1.0)
signal *= 0.9 # hack, just controlling gain
signals.append(signal)
names.append(name)
i += 1
return signals, names
def process_walk(walk_id, force=False):
if not model.process_check(walk_id):
log.error("Walk %s already processed" % walk_id)
if force:
log.info("--> forcing...")
model.remove_sequence(walk_id)
else:
return
log.info("Processing walk %s" % walk_id)
# fetch data
data = model.fetch_accels(walk_id)
data = [(reading['t'], reading['x'], reading['y'], reading['z'])
for reading in data]
# let's sample every millisecond, so the time of the last reading is how many samples we need
data = np.array(data)
ts = data[:, 0]
total_samples = int(ts[-1])
# need at least 10s of data
# add 2000 for trimming at nd
if total_samples < 10000 + 2000:
log.info("No footsteps detected (too short)")
model.hide(walk_id)
return
# resample the values
xs = sp.resample(ts, data[:, 1], total_samples)
ys = sp.resample(ts, data[:, 2], total_samples)
zs = sp.resample(ts, data[:, 3], total_samples)
# skip for accelerometer startup and for phone out of pocket at end
skipin, skipout = 0, 2000
xs = xs[skipin:-skipout]
ys = ys[skipin:-skipout]
zs = zs[skipin:-skipout]
total_samples -= (skipin + skipout)
log.info("TOTAL SAMPLES %s (%fs)" % (total_samples,
(total_samples / 1000.0)))
# get 3d magnitude (not RMS) -- orientation shouldnt matter
ds = np.sqrt(np.power(xs, 2) + np.power(ys, 2) + np.power(zs, 2))
# prep the raw values for display
# normalize the values to a given range (this is Gs)
MIN = -10.0
MAX = 10.0
xs = (xs - MIN) / (MAX - MIN)
ys = (ys - MIN) / (MAX - MIN)
zs = (zs - MIN) / (MAX - MIN)
# smooth them
xs = sp.smooth(xs, 300)
ys = sp.smooth(ys, 300)
zs = sp.smooth(zs, 300)
# process the magnitude signal
ds = sp.smooth(ds, 500)
ds = np.clip(ds, -10.0, 10.0) # limit the signal to +-10 Gs
ds = sp.normalize(ds)
ds = 1 - ds
ds = sp.compress(ds, 3.0)
ds = sp.normalize(ds)
# detect peaks
peaks, valleys = sp.detect_peaks(ds, lookahead=50, delta=0.10)
peaks = np.array(peaks)
valleys = np.array(valleys)
log.info("PEAKS %s" % len(peaks))
if not len(peaks):
log.info("No footsteps detected")
model.hide(walk_id)
return
# get foot separator line
fxs = [int(peak[0]) for peak in peaks]
fys = [peak[1] for peak in peaks]
avs = np.average([peak[1] for peak in peaks])
fys[0] = avs # it's going to start with a peak, so we need to bring it up or down accordingly
fxs.append(total_samples - 1)
fys.append(avs)
fs = sp.resample(fxs, fys, total_samples)
fs = sp.smooth(fs, 3000)
# print out
log.info("Saving sequence (%s)..." % walk_id)
sequence = []
for p, peak in enumerate(peaks):
foot = 'right' if peak[1] > fs[int(peak[0])] else 'left'
t = peak[0]
t += 250 # turns out the peak hits just before the step
sequence.append((t, foot))
# fix triples
for i in range(len(sequence) - 2):
if sequence[i][1] == sequence[i + 1][1] == sequence[i + 2][1]:
sequence[i + 1] = (sequence[i + 1][0],
'right') if sequence[i + 1][1] == 'left' else (
sequence[i + 1][0], 'left')
model.insert_sequence(walk_id, sequence)
plot(walk_id, xs, ys, zs, ds, peaks, total_samples, fs)
def generate():
# load data into t and count arrays per species
species = OrderedDict()
start_t = util.timestamp(util.parse_date(str(config['start'])))
end_t = util.timestamp(util.parse_date(str(config['end'])))
max_count = 0
with open("data.csv") as f:
data = csv.reader(f)
for r, row in enumerate(data):
if r == 0:
continue
plot = row[1]
name = row[2]
if len(config['species_list']) and name not in config['species_list']:
continue
dt = datetime.datetime(int(row[3]), 1, 1) + datetime.timedelta(int(row[4]) - 1)
t = util.timestamp(dt)
if t < start_t or t > end_t:
continue
count = 0 if row[5] == "NA" else int(row[5])
if count > max_count:
max_count = count
if name not in species:
species[name] = {'ts': [start_t, t - 1], 'counts': [0, 0]}
species[name]['ts'].append(t)
species[name]['counts'].append(count)
species = OrderedDict(sorted(species.items()))
print("--> loaded")
# add a zero count at the start and end of every year
yts = [util.timestamp(datetime.datetime(y, 1, 1)) for y in range(1974, 2017)]
for name in species:
ts = species[name]['ts']
for yt in yts:
i = 0
while i < len(ts) and ts[i] < yt:
i += 1
if i > 0:
end_season_t = ts[i-1]
if i < len(ts):
start_season_t = ts[i]
ts.insert(i, start_season_t - config['tail'])
species[name]['counts'].insert(i, 0)
ts.insert(i, end_season_t + config['tail'])
species[name]['counts'].insert(i, 0)
species[name]['ts'].append(end_t)
species[name]['counts'].append(0)
print("--> onsets added")
# create and draw signals
signals = []
names = []
i = 0
for name, data in species.items():
print("Processing %s..." % name)
# create signal from bloom counts
signal = sp.resample(data['ts'], data['counts'])
if config['normalize']:
signal = sp.normalize(signal)
else:
signal = sp.normalize(signal, 0, max_count)
signal = sp.smooth(signal, size=8)
signal = sp.limit(signal, max(signal)) # get rid of noise below 0 for onset detection
# add spikes for peaks
if config['peak_spikes']:
peaks, valleys = sp.detect_peaks(signal, lookahead=50)
peak_signal = np.zeros(len(signal))
for peak in peaks:
peak_signal[peak[0]] = 1.0
signal += peak_signal
# add spikes for onsets
if config['onset_spikes']:
onsets = sp.detect_onsets(signal)
onset_signal = np.zeros(len(signal))
for onset in onsets:
onset_signal[onset] = 0.5
onset_signal[onset+1] = 0.4
onset_signal[onset+2] = 0.25
signal += onset_signal
# limit
signal = sp.limit(signal, 1.0)
signal *= 0.9 # hack, just controlling gain
signals.append(signal)
names.append(name)
i += 1
return signals, names
def process(t):
log.info("////////// process %s //////////" % t)
filename = "audio_tmp/%s.wav" % t
sample_rate, signal = wavfile.read(filename)
log.info("AUDIO SAMPLES %s" % len(signal))
log.info("SAMPLE RATE %s" % sample_rate)
duration = float(len(signal)) / sample_rate
log.info("AUDIO DURATION %ss" % util.format_time(duration))
signal = (np.array(signal).astype('float') / (2**16 * 0.5)) # assuming 16-bit PCM, -1 - 1
log.info("--> preprocessing")
magnitude = abs(signal)
thresholded_magnitude = (magnitude > THRESHOLD) * magnitude
# level = sp.smooth(thresholded_magnitude, size=10000) # shit -- smooth is too expensive for raspi
level = thresholded_magnitude
log.info("--> scanning")
TOLERANCE = sample_rate / 10 # within a tenth of a second, same sound (poor man's smoothing?)
indexes = []
maxes = []
durations = []
zeros = 0
on_chunk = False
for index, sample in enumerate(level):
if sample > 0.0:
if not on_chunk:
indexes.append(index)
durations.append(0)
maxes.append(0)
on_chunk = True
durations[-1] += 1
if sample > maxes[-1]:
maxes[-1] = sample
zeros = 0
if sample == 0.0:
if on_chunk:
zeros += 1
if zeros == TOLERANCE:
on_chunk = False
events = []
for i in xrange(len(indexes)):
value, t_, duration = maxes[i], t + int(float(indexes[i]) / sample_rate), float(durations[i]) / sample_rate
events.append((value, t_, duration))
for event in events:
log.debug(event)
if 'draw' in config and config['draw']:
from housepy import drawing
log.info("--> drawing")
ctx = drawing.Context(width=2000, height=500, background=(0., 0., 1.), hsv=True, flip=True, relative=True)
ctx.line([(float(i) / len(magnitude), sample) for (i, sample) in enumerate(magnitude)], thickness=1, stroke=(0., 0., 0.5))
ctx.line([(float(i) / len(thresholded_magnitude), sample) for (i, sample) in enumerate(thresholded_magnitude)], thickness=1, stroke=(0., 0., 0.))
ctx.line([(float(i) / len(level), sample) for (i, sample) in enumerate(level)], thickness=1, stroke=(0., 1., 1.))
level = sp.normalize(level)
ctx.line([(float(i) / len(level), sample) for (i, sample) in enumerate(level)], thickness=1, stroke=(0.15, 1., 1.))
ctx.line(0.0, THRESHOLD, 1.0, THRESHOLD, thickness=1, stroke=(0.55, 1., 1.))
ctx.show()
try:
data = []
for event in events:
value, t_, duration = event
data.append({'device': config['device'], 'kind': "sound", 'value': value, 't': t_, 'duration': duration})
response = net.read("http://%s:%s" % (config['server']['host'], config['server']['port']), json.dumps(data))
log.info(response)
except Exception as e:
log.error(log.exc(e))
if config['device'] != "Granu":
os.remove(filename)
def main(session_id):
result = db.branches.find({'session': session_id}).sort([('t', ASCENDING)])
if not result.count():
print("NO DATA!")
exit()
log.info("Start processing...")
result = list(result)
ts = [r['t'] for r in result]
rms = [r['sample'][3] for r in result]
duration = ts[-1] - ts[0]
SAMPLING_RATE = 60 # hz
log.info("DURATION %fs" % duration)
signal = sp.resample(ts, rms, duration * SAMPLING_RATE)
signal = sp.remove_shots(signal)
signal = sp.normalize(signal)
signal = sp.smooth(signal, 15)
# this number should match some lower frequency bound. ie, put this in hz.
# the smaller the number, the more it will affect small motion
# so this should be higher than the slowest motion we care about
# ie, dont care about motion over 0.5hz, which is 120 samples
trend = sp.smooth(signal, 120)
signal -= trend
signal += 0.5
atrend = sp.smooth(signal, 500)
## autocorrelation
auto = sp.autocorrelate(signal)
# this should be small -- if 60hz, fastest gesture would reasonably be half of that, so 30
peaks, valleys = sp.detect_peaks(auto, 10)
peaks = [peak for peak in peaks[1:] if peak[1] > 0.5]
partials = []
for peak in peaks:
frequency = SAMPLING_RATE / peak[0]
partial = frequency * 1000
partials.append([partial, float(peak[1])])
log.info("%d samps\t%fhz\t%f magnitude\t%f map" % (peak[0], frequency, peak[1], partial))
log.info(partials)
ctx = drawing.Context(2000, 750)
ctx.plot(auto, stroke=(0.0, 0.0, 0.0, 1.0), thickness=2.0)
for peak in peaks:
x = peak[0] / len(auto)
ctx.line(x, 0.0, x, peak[1], stroke=(1.0, 0.0, 0.0, 1.0))
ctx.output("graphs")
## audio
audio_signal = sp.make_audio(signal)
spectrum(audio_signal, SAMPLING_RATE)
AUDIO_RATE = 11025
filename = "%s.wav" % util.timestamp()
sound.write_audio(audio_signal, filename, AUDIO_RATE)
subprocess.call(["open", filename])
log.info("AUDIO DURATION %fs" % (duration / (AUDIO_RATE / SAMPLING_RATE)))
ctx = drawing.Context(2000, 750)
ctx.plot(signal, stroke=(0.0, 0.0, 0.0, 1.0), thickness=2.0)
ctx.plot(trend, stroke=(1.0, 0.0, 0.0, 1.0), thickness=2.0)
ctx.plot(atrend, stroke=(0.0, 0.0, 1.0, 1.0), thickness=2.0)
ctx.output("graphs")
log.info("--> done") # around 300ms
points, rates = util.load("data/last_snap.pkl")
log.info("INPUT: %s POINTS, %s DIMENSIONS" % points.shape)
# points = manifold.Isomap().fit_transform(points)
# points = manifold.LocallyLinearEmbedding(method="modified").fit_transform(points)
# points = manifold.SpectralEmbedding().fit_transform(points)
# points = manifold.MDS().fit_transform(points)
# points = manifold.TSNE(n_iter=2000).fit_transform(points)
# points = decomposition.PCA(n_components=2).fit_transform(points)
# points = manifold.TSNE().fit_transform(points)
points = decomposition.PCA(n_components=2).fit_transform(points)
log.info("OUTPUT: %s POINTS, %s DIMENSIONS" % points.shape)
# labels = cluster.DBSCAN(eps=0.1, min_samples=5).fit_predict(points)
clusterer = cluster.KMeans(n_clusters=8)
labels = clusterer.fit_predict(points)
centroids = clusterer.cluster_centers_
labels += abs(min(labels))
max_label = max(labels)
log.info("CENTROIDS\n%s" % centroids)
centroids = np.column_stack((sp.normalize(centroids[:,0], np.min(points[:,0]), np.max(points[:,0])), sp.normalize(centroids[:,1], np.min(points[:,1]), np.max(points[:,1]))))
points = np.column_stack((sp.normalize(points[:,0], np.min(points[:,0]), np.max(points[:,0])), sp.normalize(points[:,1], np.min(points[:,1]), np.max(points[:,1]))))
chart.plot(points, sample_axis=True, scatter=False, c=(0., 0., 1., 1.), linewidth=2)
chart.plot(centroids, sample_axis=True, scatter=True, c=(1., 0., 0., 1.), linewidth=0, s=100)
chart.show("charts/")
def spectrum(signal, rate):
log.info("Computing spectrogram...")
block_size = 512
block_overlap = block_size / 2 # power of two, default is 128
# freqs, ts, spectrum = spectrogram(sound.signal, fs=sound.rate, noverlap=block_overlap, nfft=block_size, detrend='constant', return_onesided=True, scaling='density', axis=-1, mode='psd', window=('tukey', 0.25), nperseg=block_overlap*2)
spectrum, freqs, ts, image = plt.specgram(signal,
NFFT=block_size,
Fs=rate,
noverlap=block_overlap)
# (plt is 3k smaller)
# print("spectrum", spectrum) # freq rows of time columns.
# print()
# print(freqs)
# print()
# print(ts)
log.info("--> done")
log.info("--> freq bins %s" % len(freqs))
log.info("--> time columns %s" % len(ts))
log.info("Drawing...")
# with gzip.open("spectrum.pklz", 'wb') as f:
# f.write(pickle.dumps(spectrum))
ctx = drawing.Context(
len(ts) * 1, len(freqs) * 1,
relative=True) # if it's not an even multiple, artifacts happen
pixel_width = ctx.width / len(spectrum[0])
pixel_height = ctx.height / len(spectrum)
# for y, row in enumerate(spectrum):
# for x, value in enumerate(row):
# v = min(value / (allmax / 500), 1.0)
# v = 1 - v
# # print((x * pixel_width) / ctx.width, (y * pixel_height) / ctx.height, pixel_width / ctx.width, pixel_height / ctx.height)
# ctx.rect((x * pixel_width) / ctx.width, (y * pixel_height) / ctx.height, pixel_width / ctx.width, pixel_height / ctx.height, fill=(v, v, v, 1.), stroke=(1., 0., 0., 0.), thickness=0.0)
# mx = spectrum.flatten().max()
# print("maximum", spectrum.flatten().max())
# print("minimum", spectrum.flatten().min())
spectrum = sp.normalize(
np.sqrt(spectrum), 0.0,
200.0) # sqrt compresses, good for power. 200 is a clipping threshold.
for y, row in enumerate(spectrum):
for x, v in enumerate(row):
ctx.line((x * pixel_width) / ctx.width,
(y * pixel_height) / ctx.height,
((x * pixel_width) + pixel_width) / ctx.width,
(y * pixel_height) / ctx.height,
stroke=(v, v, v, 1.),
thickness=pixel_height)
log.info("--> done")
ctx.output("charts/")
def sample(draw=False):
log.info("START SAMPLE")
# get the time
# dt = timeutil.get_dt(tz=config['tz'])
# dt -= datetime.timedelta(days=300) # time adjustment if necessary for testing
# t_utc = timeutil.t_utc(dt)
t_utc = timeutil.t_utc()
dt = timeutil.get_dt(t_utc, tz=config['tz'])
log.info("CURRENT TIME %s" % timeutil.get_string(t_utc, tz=config['tz']))
# pull the last 24 hours worth -- we're going to normalize over that to set our dynamic levels
log.info(config['sites'][config['sample']])
# # this is the real-time last 24 hours
# query = {'site': config['sample'], 't_utc': {'$gt': t_utc - 86400, '$lt': t_utc}}
# log.info(query)
# results = db.entries.find(query)
# this is the last 24 hours we have
# assume updating every 15 minutes, last 24 hours is the last 96 results
results = db.entries.find({
'site': config['sample']
}).sort([('t_utc', DESCENDING)]).limit(96)
results = list(results)
results.reverse()
log.info("%s results" % len(results)) # should be 96
log.info(json.dumps(results[-1], indent=4,
default=lambda d: str(d))) # show the last one
# resample signals for each
ts = [d['t_utc'] for d in results]
duration = ts[-1] - ts[0]
log.info("DURATION %s %s" % (duration, timeutil.format_seconds(duration)))
signals = []
rates = []
labels = list(config['labels'].values())
labels.sort()
for i, label in enumerate(labels):
# log.debug(label)
try:
values = [d[label] if label in d else None for d in results]
values = sp.remove_shots(values,
nones=True) # repair missing values
signal = sp.resample(ts, values)
num_samples = len(signal)
sample_rate = num_samples / duration
rates.append(sample_rate)
signal = sp.normalize(signal)
signal = sp.smooth(signal, 15)
signals.append(signal)
except KeyError as e:
log.error(log.exc(e))
log.error(values)
# draw if desired
if draw:
from housepy import drawing
ctx = drawing.Context(1200, 500, margin=20, hsv=True)
for i, label in enumerate(labels):
color = i / len(labels), .8, .8, 1.
signal = signals[i]
ctx.plot(signal, stroke=color, thickness=2)
ctx.output("charts/")
# collapse into n-dimensional points
points = []
for i in range(len(signals[0])):
point = [signal[i] for signal in signals]
points.append(point)
# PCA to 4D -- this takes whatever data we've got and maximizes variation for our four panels
points = np.array(points)
# log.debug("INPUT: %s POINTS, %s DIMENSIONS" % points.shape)
points = decomposition.PCA(n_components=4).fit_transform(points)
# log.debug("OUTPUT: %s POINTS, %s DIMENSIONS" % points.shape)
# normalize each dimension independently, again amplifying dynamics
points = np.column_stack((sp.normalize(points[:, 0], np.min(points[:, 0]),
np.max(points[:, 0])),
sp.normalize(points[:, 1], np.min(points[:, 1]),
np.max(points[:, 1])),
sp.normalize(points[:, 1], np.min(points[:, 1]),
np.max(points[:, 2])),
sp.normalize(points[:, 2], np.min(points[:, 3]),
np.max(points[:, 3]))))
# now, for each time this is queried we want to return an interpolation between the last two points
# this essentially implements a delay that closes in on the most recent query
# ...hopefully to be refreshed with a new USGS reading when it gets there
# if that reading doesnt come, it's ok, it just hovers there until we proceed
# aaandd actually we want a couple of hours delay, because these come in at bulk every 1-4 hours
# we know we have 96 points. four hours back is 16 points
# interpolating between points -17 and -16 should give the most recent guaranteed smooth transitions
# transduction takes time, pues
point_a = points[-17]
point_b = points[-16]
# log.debug(point_a)
# log.debug(point_b)
# linear interpolation over 15 minutes
position = (((dt.minute % 15) * 60) + dt.second) / (15 * 60)
# log.debug(position)
point = [(point_a[i] * (1.0 - position)) + (point_b[i] * position)
for i in range(len(point_a))]
log.info("RESULT: %s" % point)
return point
def process_walk(walk_id, force=False):
if not model.process_check(walk_id):
log.error("Walk %s already processed" % walk_id)
if force:
log.info("--> forcing...")
model.remove_sequence(walk_id)
else:
return
log.info("Processing walk %s" % walk_id)
# fetch data
data = model.fetch_accels(walk_id)
data = [(reading['t'], reading['x'], reading['y'], reading['z']) for reading in data]
# let's sample every millisecond, so the time of the last reading is how many samples we need
data = np.array(data)
ts = data[:,0]
total_samples = int(ts[-1])
# need at least 10s of data
# add 2000 for trimming at nd
if total_samples < 10000 + 2000:
log.info("No footsteps detected (too short)")
model.hide(walk_id)
return
# resample the values
xs = sp.resample(ts, data[:,1], total_samples)
ys = sp.resample(ts, data[:,2], total_samples)
zs = sp.resample(ts, data[:,3], total_samples)
# skip for accelerometer startup and for phone out of pocket at end
skipin, skipout = 0, 2000
xs = xs[skipin:-skipout]
ys = ys[skipin:-skipout]
zs = zs[skipin:-skipout]
total_samples -= (skipin + skipout)
log.info("TOTAL SAMPLES %s (%fs)" % (total_samples, (total_samples / 1000.0)))
# get 3d magnitude (not RMS) -- orientation shouldnt matter
ds = np.sqrt(np.power(xs, 2) + np.power(ys, 2) + np.power(zs, 2))
# prep the raw values for display
# normalize the values to a given range (this is Gs)
MIN = -10.0
MAX = 10.0
xs = (xs - MIN) / (MAX - MIN)
ys = (ys - MIN) / (MAX - MIN)
zs = (zs - MIN) / (MAX - MIN)
# smooth them
xs = sp.smooth(xs, 300)
ys = sp.smooth(ys, 300)
zs = sp.smooth(zs, 300)
# process the magnitude signal
ds = sp.smooth(ds, 500)
ds = np.clip(ds, -10.0, 10.0) # limit the signal to +-10 Gs
ds = sp.normalize(ds)
ds = 1 - ds
ds = sp.compress(ds, 3.0)
ds = sp.normalize(ds)
# detect peaks
peaks, valleys = sp.detect_peaks(ds, lookahead=50, delta=0.10)
peaks = np.array(peaks)
valleys = np.array(valleys)
log.info("PEAKS %s" % len(peaks))
if not len(peaks):
log.info("No footsteps detected")
model.hide(walk_id)
return
# get foot separator line
fxs = [int(peak[0]) for peak in peaks]
fys = [peak[1] for peak in peaks]
avs = np.average([peak[1] for peak in peaks])
fys[0] = avs # it's going to start with a peak, so we need to bring it up or down accordingly
fxs.append(total_samples-1)
fys.append(avs)
fs = sp.resample(fxs, fys, total_samples)
fs = sp.smooth(fs, 3000)
# print out
log.info("Saving sequence (%s)..." % walk_id)
sequence = []
for p, peak in enumerate(peaks):
foot = 'right' if peak[1] > fs[int(peak[0])] else 'left'
t = peak[0]
t += 250 # turns out the peak hits just before the step
sequence.append((t, foot))
# fix triples
for i in range(len(sequence) - 2):
if sequence[i][1] == sequence[i+1][1] == sequence[i+2][1]:
sequence[i+1] = (sequence[i+1][0], 'right') if sequence[i+1][1] == 'left' else (sequence[i+1][0], 'left')
model.insert_sequence(walk_id, sequence)
plot(walk_id, xs, ys, zs, ds, peaks, total_samples, fs)
duration = ts[-1] - ts[0]
print("DURATION %s %s" % (duration, strings.format_time(duration)))
signals = []
rates = []
labels = list(config['labels'].values())
labels.sort()
for i, label in enumerate(labels):
log.info(label)
try:
values = [d[label] if label in d else None for d in results]
values = sp.remove_shots(values, nones=True) # repair missing values
signal = sp.resample(ts, values)
num_samples = len(signal)
sample_rate = num_samples / duration
rates.append(sample_rate)
signal = sp.normalize(signal)
signal = sp.smooth(signal, 15)
signals.append(signal)
# color = colors[i]
color = i / len(labels), .8, .8, 1.
ctx.plot(signal, stroke=color, thickness=2)
ctx.line(10 / ctx.width, 1 - ((10 + (i * 10)) / ctx.height), 30 / ctx.width, 1 - ((10 + (i * 10)) / ctx.height), stroke=color, thickness=2)
ctx.label(35 / ctx.width, 1 - ((13 + (i * 10)) / ctx.height), label.upper(), size=8)
except KeyError as e:
log.error(log.exc(e))
log.error(values)
ctx.output("charts/")
points = []
for i in range(len(signals[0])):
db.stream.find(
{
"t_utc": {
"$gt": util.timestamp(util.parse_date(START, tz="America/New_York")),
"$lt": util.timestamp(util.parse_date(END, tz="America/New_York")),
}
}
).sort([("t_utc", pymongo.ASCENDING)])
)
log.info("--> done")
##
ts = [r["t_utc"] for r in results]
xs = [r["x"] for r in results]
duration = ts[-1] - ts[0]
SAMPLING_RATE = 100
log.info("Resampling...")
signal = sp.resample(ts, xs, duration * SAMPLING_RATE)
signal += 1 # change -1,1 to 0,2
signal = sp.normalize(signal, 0, 2)
log.info("--> done")
log.info("Drawing...")
ctx = drawing.Context(1000, 400)
ctx.plot(signal, stroke=(0.0, 0.0, 0.0, 1.0), thickness=1.0)
ctx.output("graphs")
log.info("--> done")
log.info("T_MAX %s" % t_max)
signals = []
labels = list(streams.keys())
log.info("LABELS %s" % labels)
for label in labels:
log.info(label)
ts = tses[label]
ts = [t_min] + ts + [t_max]
values = [d[label] if label in d else None for d in streams[label]]
values = [values[0]] + values + [values[-1]]
values = sp.remove_shots(values, nones=True) # repair missing values
signal = sp.resample(ts, values)
num_samples = len(signal)
sample_rate = num_samples / duration
signal = sp.normalize(signal)
signal = sp.smooth(signal, 15)
signals.append(signal)
log.info("Drawing...")
ctx = drawing.Context(1200, 500, margin=20, hsv=True)
for b in range(12):
ctx.line(b / 12, 0, b / 12, 1, stroke=(0.5, 0.5, 0.5, 0.5), thickness=0.5)
ctx.line(1, 0, 1, 1, stroke=(0.5, 0.5, 0.5, 0.5), thickness=0.5)
for i, signal in enumerate(signals):
color = i / (len(signals) + 4) + .1, 1., .8, 1.
ctx.plot(signal, stroke=color, thickness=1.5)
ctx.line(10 / ctx.width, 1 - ((10 + (i * 10)) / ctx.height), 30 / ctx.width, 1 - ((10 + (i * 10)) / ctx.height), stroke=color, thickness=2)
ctx.label(35 / ctx.width, 1 - ((13 + (i * 10)) / ctx.height), labels[i].upper(), size=8)
ctx.output("charts/")
voices = []
for walk in walks:
sequence = model.fetch_sequence(walk['id'])
if len(sequence) < MIN_STEPS:
continue
voices.append(Swerve(v + 1))
for step in sequence:
notes.append((step[0], v, 0 if step[1] == 'left' else 1))
v += 1
if v == len(steps):
break
# sort and normalize onsets
notes.sort(key=lambda x: x[0])
onsets = [note[0] for note in notes]
onsets = sp.normalize(onsets)
notes = [(onsets[i], note[1], note[2]) for (i, note) in enumerate(notes)]
for voice in voices:
voice.synth = 'cycle'
voice.attack = 350
# voice.sustain = 350
voice.sustain = 2000
voice.decay = 600
voice.reverb = 0.5, 0.5, 0.45, 0.5, 0.0
# voice.chord = C3
DURATION = 20 * 60.0
DURATION = 10 * 60.0
