def detect_spikey(local_chunk,thr=None,plotose=0):
if (thr is None): thr=np.repeat(50,local_chunk.shape[0])
from scipy.interpolate import interp1d
nchan=local_chunk.shape[1]
all_peaktimes=[];
for ich in range(nchan):
all_peaktimes.extend(algo.detect_peaks(np.abs(local_chunk[:, ich]), thr[ich], valley=False, mpd=10,show=False))
def histc(X, bins):
map_to_bins = np.digitize(X,bins)
r = np.zeros(bins.shape)
for iii in map_to_bins:
r[iii-1] += 1
return [r, map_to_bins]
# buf=np.random.gamma(1,3,[1000,64]) #test data
all_peaktimes=np.sort(all_peaktimes)
if (len(all_peaktimes)<6)|(local_chunk.shape[0]<90001):
return [],local_chunk
R,_=histc(all_peaktimes,np.arange(min(all_peaktimes),max(all_peaktimes),30000))
is_spikey=R> (nchan*35) # 20 spikes per second on every channel is too much
if (len(is_spikey)<3): return [],local_chunk
if plotose:
_ = algo.detect_peaks(np.abs(local_chunk[:, -2]), thr, valley=False, mpd=10,show=True)
plt.show()
plt.plot(all_peaktimes,'.');plt.title('all chan over time');plt.show();plt.close()
plt.plot(np.diff(all_peaktimes),'.');plt.title('all chan diff');plt.show();plt.close()
#%
plt.plot(R,'.');plt.title('Raster all chan over time');plt.show();plt.close()
plt.plot(R/nchan,'-');plt.title('all chan spikes per chan per sec');plt.show();plt.close()
plt.plot(is_spikey,'.');plt.title('is_spiky');plt.show();plt.close()
f=interp1d(range(len(is_spikey)),is_spikey)
is_spikey1=np.round(f(np.linspace(0,len(is_spikey)-1,local_chunk.shape[0])))
local_chunk[is_spikey1==1]=0#np.nan
return is_spikey1,local_chunk
def main(params, nb_cpu, nb_gpu, use_gpu):
# Part 1: Whitening
numpy.random.seed(420)
logger = init_logging(params.logfile)
logger = logging.getLogger('circus.whitening')
#################################################################
data_file = params.data_file
data_file.open()
N_e = params.getint('data', 'N_e')
N_total = params.nb_channels
N_t = params.getint('detection', 'N_t')
dist_peaks = params.getint('detection', 'dist_peaks')
template_shift = params.getint('detection', 'template_shift')
file_out_suff = params.get('data', 'file_out_suff')
file_out = params.get('data', 'file_out')
spike_thresh = params.getfloat('detection', 'spike_thresh')
matched_filter = params.getboolean('detection', 'matched-filter')
matched_thresh = params.getfloat('detection', 'matched_thresh')
sign_peaks = params.get('detection', 'peaks')
do_temporal_whitening = params.getboolean('whitening', 'temporal')
do_spatial_whitening = params.getboolean('whitening', 'spatial')
chunk_size = params.getint('whitening', 'chunk_size')
plot_path = os.path.join(params.get('data', 'data_file_noext'), 'plots')
nodes, edges = get_nodes_and_edges(params)
safety_time = params.getint('whitening', 'safety_time')
nb_temp_white = min(max(20, comm.size), N_e)
max_silence_1 = int(20 * params.rate // comm.size)
max_silence_2 = 5000
inv_nodes = numpy.zeros(N_total, dtype=numpy.int32)
inv_nodes[nodes] = numpy.argsort(nodes)
#################################################################
if comm.rank == 0:
print_and_log(
["Analyzing data to get whitening matrices and thresholds..."],
'default', logger)
if use_gpu:
import cudamat as cmt
## Need to properly handle multi GPU per MPI nodes?
if nb_gpu > nb_cpu:
gpu_id = int(comm.rank // nb_cpu)
else:
gpu_id = 0
cmt.cuda_set_device(gpu_id)
cmt.init()
cmt.cuda_sync_threads()
nb_chunks, last_chunk_len = data_file.analyze(chunk_size)
if nb_chunks < comm.size:
res = io.data_stats(params, show=False)
chunk_size = int(res * params.rate // comm.size)
if comm.rank == 0:
print_and_log(
["Too much cores, automatically resizing the data chunks"],
'debug', logger)
nb_chunks, last_chunk_len = data_file.analyze(chunk_size)
# I guess this is more relevant, to take signals from all over the recordings
all_chunks = numpy.random.permutation(
numpy.arange(nb_chunks, dtype=numpy.int32))
all_electrodes = numpy.random.permutation(N_e)
for gidx in [all_chunks[comm.rank]]:
#print "Node", comm.rank, "is analyzing chunk", gidx, "/", nb_chunks, " ..."
local_chunk, t_offset = data_file.get_data(gidx,
chunk_size,
nodes=nodes)
local_shape = len(local_chunk)
#print "Node", comm.rank, "computes the median absolute deviations in a random chunk"
thresholds = numpy.zeros(N_e, dtype=numpy.float32)
for i in xrange(N_e):
u = numpy.median(local_chunk[:, i], 0)
thresholds[i] = numpy.median(numpy.abs(local_chunk[:, i] - u), 0)
gdata = gather_array(thresholds, comm)
if comm.rank == 0:
gdata = gdata.reshape((comm.size, N_e))
thresholds = numpy.mean(gdata, 0)
bfile = h5py.File(file_out_suff + '.basis.hdf5',
'w',
libver='latest')
io.write_datasets(bfile, ['thresholds'],
{'thresholds': thresholds})
bfile.close()
comm.Barrier()
thresholds = io.load_data(params, 'thresholds')
#print "Extracting the peaks..."
local_peaktimes = numpy.zeros(0, dtype=numpy.int32)
for i in xrange(N_e):
peaktimes = algo.detect_peaks(numpy.abs(local_chunk[:, i]),
thresholds[i],
valley=False,
mpd=dist_peaks)
local_peaktimes = numpy.concatenate((local_peaktimes, peaktimes))
local_peaktimes = numpy.unique(local_peaktimes)
#print "Removing the useless borders..."
local_borders = (template_shift, local_shape - template_shift)
idx = (local_peaktimes >= local_borders[0]) & (local_peaktimes <
local_borders[1])
local_peaktimes = numpy.compress(idx, local_peaktimes)
if len(local_peaktimes) > 0:
diff_times = local_peaktimes[-1] - local_peaktimes[0]
all_times = numpy.zeros((N_e, diff_times + 1), dtype=numpy.bool)
min_times = numpy.maximum(
local_peaktimes - local_peaktimes[0] - safety_time, 0)
max_times = numpy.minimum(
local_peaktimes - local_peaktimes[0] + safety_time + 1,
diff_times)
argmax_peak = numpy.random.permutation(
numpy.arange(len(local_peaktimes)))
all_idx = numpy.take(local_peaktimes, argmax_peak)
#print "Selection of the peaks with spatio-temporal masks..."
for idx, peak in zip(argmax_peak, all_idx):
elec = numpy.argmax(numpy.abs(local_chunk[peak]))
indices = numpy.take(inv_nodes, edges[nodes[elec]])
all_times[indices, min_times[idx]:max_times[idx]] = True
else:
all_times = numpy.zeros((N_e, len(local_chunk)), dtype=numpy.bool)
all_times_Ne = numpy.any(all_times, 0)
subset = numpy.where(all_times_Ne == False)[0]
all_silences = []
if do_spatial_whitening:
local_silences = numpy.take(local_chunk, subset,
axis=0)[:max_silence_1]
all_silences = gather_array(local_silences, comm, 0, 1)
local_res = []
if do_temporal_whitening:
for elec in all_electrodes[numpy.arange(comm.rank, nb_temp_white,
comm.size)]:
res = numpy.zeros((0, N_t), dtype=numpy.float32)
scount = 0
indices = numpy.take(inv_nodes, edges[nodes[elec]])
all_times_elec = numpy.any(numpy.take(all_times, indices, axis=0),
0)
esubset = numpy.where(all_times_elec == False)[0]
bound = len(esubset) - N_t
while (scount < bound) and (len(res) < max_silence_2):
myslice = esubset[scount:scount + N_t]
if numpy.all((myslice - esubset[scount]) == numpy.arange(N_t)):
scount += N_t
res = numpy.vstack((res, local_chunk[myslice, elec]))
else:
scount += 1
if len(res) > 5:
local_res += [numpy.cov(res.T)]
nb_elecs = numpy.array([len(local_res)], dtype=numpy.float32)
local_res = numpy.array(local_res, dtype=numpy.float32)
if len(local_res) == 0:
local_res = numpy.zeros(0, dtype=numpy.float32)
else:
local_res = numpy.sum(local_res, 0)
all_res = gather_array(local_res.ravel(), comm, 0, 1)
all_elecs = gather_array(nb_elecs, comm, 0, 1)
if comm.rank == 0:
to_write = {}
if do_temporal_whitening:
try:
nb_silences = numpy.sum(all_elecs > 0)
all_res = all_res.reshape((nb_silences, N_t**2))
except Exception:
print_and_log([
"No silent periods detected: something wrong with the parameters?"
], 'error', logger)
all_res = numpy.sum(all_res, 0)
all_res = all_res.reshape((N_t, N_t)) / numpy.sum(all_elecs)
temporal_whitening = get_whitening_matrix(
all_res.astype(numpy.double),
fudge=1e-3)[template_shift].astype(numpy.float32)
temporal_whitening /= temporal_whitening.sum()
to_write['temporal'] = temporal_whitening
have_nans = numpy.sum(numpy.isnan(temporal_whitening))
if have_nans > 0:
temporal_whitening = numpy.zeros(N_t, dtype=numpy.float32)
temporal_whitening[N_t // 2] = 1
to_write['temporal'] = temporal_whitening
print_and_log(
["Disabling temporal whitening because of NaNs found"],
'info', logger)
if do_spatial_whitening:
if len(all_silences) / params.rate == 0:
print_and_log([
"No silent periods detected: something wrong with the parameters?"
], 'error', logger)
spatial_whitening = get_whitening_matrix(
all_silences.astype(numpy.double)).astype(numpy.float32)
to_write['spatial'] = spatial_whitening
print_and_log([
"Found %gs without spikes for whitening matrices..." %
(len(all_silences) / params.rate)
], 'default', logger)
have_nans = numpy.sum(numpy.isnan(spatial_whitening))
if have_nans > 0:
spatial_whitening = numpy.eye(spatial_whitening.shape[0],
dtype=numpy.float32)
to_write['spatial'] = spatial_whitening
print_and_log(
["Disabling spatial whitening because of NaNs found"],
'info', logger)
bfile = h5py.File(file_out_suff + '.basis.hdf5', 'r+', libver='latest')
io.write_datasets(bfile, to_write.keys(), to_write)
bfile.close()
del all_silences
comm.Barrier()
if do_spatial_whitening or do_temporal_whitening:
if comm.rank == 0:
print_and_log(
["Because of whitening, need to recompute the thresholds..."],
'default', logger)
if do_spatial_whitening:
spatial_whitening = io.load_data(params, 'spatial_whitening')
if use_gpu:
spatial_whitening = cmt.CUDAMatrix(spatial_whitening,
copy_on_host=False)
if do_temporal_whitening:
temporal_whitening = io.load_data(params, 'temporal_whitening')
for gidx in [all_chunks[comm.rank]]:
local_chunk, t_offset = data_file.get_data(gidx,
chunk_size,
nodes=nodes)
local_shape = len(local_chunk)
if do_spatial_whitening:
if use_gpu:
local_chunk = cmt.CUDAMatrix(local_chunk,
copy_on_host=False)
local_chunk = local_chunk.dot(spatial_whitening).asarray()
else:
local_chunk = numpy.dot(local_chunk, spatial_whitening)
if do_temporal_whitening:
local_chunk = scipy.ndimage.filters.convolve1d(
local_chunk, temporal_whitening, axis=0, mode='constant')
thresholds = numpy.zeros(N_e, dtype=numpy.float32)
for i in xrange(N_e):
u = numpy.median(local_chunk[:, i], 0)
thresholds[i] = numpy.median(numpy.abs(local_chunk[:, i] - u),
0)
gdata = gather_array(thresholds, comm)
if comm.rank == 0:
gdata = gdata.reshape((comm.size, N_e))
thresholds = numpy.mean(gdata, 0)
bfile = h5py.File(file_out_suff + '.basis.hdf5',
'r+',
libver='latest')
bfile.pop('thresholds')
io.write_datasets(bfile, ['thresholds'],
{'thresholds': thresholds})
bfile.close()
comm.Barrier()
#if comm.rank == 0:
#if not os.path.exists(plot_path):
# os.makedirs(plot_path)
#N_elec = min(int(numpy.sqrt(data_file.N_e)), 5)
#plot.view_fit(filename, t_start=0, t_stop=1, fit_on=False, square=True,
# n_elec=N_elec, save=[plot_path, 'electrodes'])
# Part 2: Basis
numpy.random.seed(422)
#################################################################
file_out = params.get('data', 'file_out')
alignment = params.getboolean('detection', 'alignment')
spike_thresh = params.getfloat('detection', 'spike_thresh')
nodes, edges = get_nodes_and_edges(params)
do_temporal_whitening = params.getboolean('whitening', 'temporal')
do_spatial_whitening = params.getboolean('whitening', 'spatial')
chunk_size = params.getint('data', 'chunk_size')
safety_time = params.getint('whitening', 'safety_time')
max_elts_elec = params.getint('whitening', 'max_elts')
nb_elts = int(
params.getfloat('whitening', 'nb_elts') * N_e * max_elts_elec)
output_dim = params.getfloat('whitening', 'output_dim')
inv_nodes = numpy.zeros(N_total, dtype=numpy.int32)
inv_nodes[nodes] = numpy.argsort(nodes)
if sign_peaks == 'both':
max_elts_elec *= 2
#################################################################
if comm.rank == 0:
print_and_log(["Searching spikes to construct the PCA basis..."],
'default', logger)
nb_chunks, last_chunk_len = data_file.analyze(chunk_size)
if nb_chunks < comm.size:
res = io.data_stats(params, show=False)
chunk_size = int(res * params.rate // comm.size)
if comm.rank == 0:
print_and_log(
["Too much cores, automatically resizing the data chunks"],
'debug', logger)
nb_chunks, last_chunk_len = data_file.analyze(chunk_size)
groups = {}
for i in xrange(N_e):
groups[i] = 0
# I guess this is more relevant, to take signals from all over the recordings
all_chunks = numpy.random.permutation(
numpy.arange(nb_chunks, dtype=numpy.int32))
max_elts_elec //= comm.size
nb_elts //= comm.size
elt_count_pos = 0
elt_count_neg = 0
if sign_peaks in ['positive', 'both']:
elts_pos = numpy.zeros((N_t, nb_elts), dtype=numpy.float32)
if sign_peaks in ['negative', 'both']:
elts_neg = numpy.zeros((N_t, nb_elts), dtype=numpy.float32)
chunks_to_load = all_chunks[comm.rank::comm.size]
thresholds = io.load_data(params, 'thresholds')
if comm.rank == 0:
pbar = get_progressbar(nb_elts)
if alignment:
cdata = numpy.linspace(-template_shift, template_shift, 5 * N_t)
xdata = numpy.arange(-2 * template_shift, 2 * template_shift + 1)
for gcount, gidx in enumerate(chunks_to_load):
if ((elt_count_pos + elt_count_neg) < nb_elts):
#print "Node", comm.rank, "is analyzing chunk", gidx, "/", nb_chunks, " ..."
local_chunk, t_offset = data_file.get_data(gidx,
chunk_size,
nodes=nodes)
local_shape = len(local_chunk)
if do_spatial_whitening:
if use_gpu:
local_chunk = cmt.CUDAMatrix(local_chunk,
copy_on_host=False)
local_chunk = local_chunk.dot(spatial_whitening).asarray()
else:
local_chunk = numpy.dot(local_chunk, spatial_whitening)
if do_temporal_whitening:
local_chunk = scipy.ndimage.filters.convolve1d(
local_chunk, temporal_whitening, axis=0, mode='constant')
#print "Extracting the peaks..."
all_peaktimes = numpy.zeros(0, dtype=numpy.int32)
all_extremas = numpy.zeros(0, dtype=numpy.int32)
for i in xrange(N_e):
if sign_peaks == 'negative':
peaktimes = algo.detect_peaks(local_chunk[:, i],
thresholds[i],
valley=True,
mpd=dist_peaks)
elif sign_peaks == 'positive':
peaktimes = algo.detect_peaks(local_chunk[:, i],
thresholds[i],
valley=False,
mpd=dist_peaks)
elif sign_peaks == 'both':
peaktimes = algo.detect_peaks(numpy.abs(local_chunk[:, i]),
thresholds[i],
valley=False,
mpd=dist_peaks)
all_peaktimes = numpy.concatenate((all_peaktimes, peaktimes))
all_extremas = numpy.concatenate(
(all_extremas,
i * numpy.ones(len(peaktimes), dtype=numpy.int32)))
#print "Removing the useless borders..."
if alignment:
local_borders = (2 * template_shift,
local_shape - 2 * template_shift)
else:
local_borders = (template_shift, local_shape - template_shift)
idx = (all_peaktimes >= local_borders[0]) & (all_peaktimes <
local_borders[1])
all_peaktimes = numpy.compress(idx, all_peaktimes)
all_extremas = numpy.compress(idx, all_extremas)
local_peaktimes = numpy.unique(all_peaktimes)
if len(local_peaktimes) > 0:
diff_times = local_peaktimes[-1] - local_peaktimes[0]
all_times = numpy.zeros((N_e, diff_times + 1),
dtype=numpy.bool)
min_times = numpy.maximum(
local_peaktimes - local_peaktimes[0] - safety_time, 0)
max_times = numpy.minimum(
local_peaktimes - local_peaktimes[0] + safety_time + 1,
diff_times)
n_times = len(local_peaktimes)
argmax_peak = numpy.random.permutation(numpy.arange(n_times))
all_idx = numpy.take(local_peaktimes, argmax_peak)
#print "Selection of the peaks with spatio-temporal masks..."
for midx, peak in zip(argmax_peak, all_idx):
if (elt_count_neg + elt_count_pos) == nb_elts:
break
if sign_peaks == 'negative':
elec = numpy.argmin(local_chunk[peak])
negative_peak = True
elif sign_peaks == 'positive':
elec = numpy.argmax(local_chunk[peak])
negative_peak = False
elif sign_peaks == 'both':
if numpy.abs(numpy.max(local_chunk[peak])) > numpy.abs(
numpy.min(local_chunk[peak])):
elec = numpy.argmax(local_chunk[peak])
negative_peak = False
else:
elec = numpy.argmin(local_chunk[peak])
negative_peak = True
indices = numpy.take(inv_nodes, edges[nodes[elec]])
myslice = all_times[indices,
min_times[midx]:max_times[midx]]
is_local_extrema = elec in all_extremas[all_peaktimes ==
peak]
if is_local_extrema and not myslice.any():
upper_bounds = max_elts_elec
if groups[elec] < upper_bounds:
if negative_peak:
elts_neg[:, elt_count_neg] = local_chunk[
peak - template_shift:peak +
template_shift + 1, elec]
else:
elts_pos[:, elt_count_pos] = local_chunk[
peak - template_shift:peak +
template_shift + 1, elec]
if alignment:
ydata = local_chunk[peak -
2 * template_shift:peak +
2 * template_shift + 1,
elec]
f = scipy.interpolate.UnivariateSpline(xdata,
ydata,
s=0)
if negative_peak:
rmin = (numpy.argmin(f(cdata)) -
len(cdata) / 2.) / 5.
else:
rmin = (numpy.argmax(f(cdata)) -
len(cdata) / 2.) / 5.
ddata = numpy.linspace(rmin - template_shift,
rmin + template_shift,
N_t)
if negative_peak:
elts_neg[:,
elt_count_neg] = f(ddata).astype(
numpy.float32)
else:
elts_pos[:,
elt_count_pos] = f(ddata).astype(
numpy.float32)
if negative_peak:
elt_count_neg += 1
else:
elt_count_pos += 1
groups[elec] += 1
all_times[indices,
min_times[midx]:max_times[midx]] = True
if comm.rank == 0:
pbar.update(elt_count_pos + elt_count_neg)
if comm.rank == 0:
if (elt_count_pos + elt_count_neg <
(gcount + 1) * max_elts_elec // len(chunks_to_load)):
pbar.update(
(gcount + 1) * max_elts_elec // len(chunks_to_load))
if comm.rank == 0:
pbar.finish()
print_and_log([
"Node %d has collected %d waveforms" %
(comm.rank, elt_count_pos + elt_count_neg)
], 'debug', logger)
if sign_peaks in ['negative', 'both']:
gdata_neg = gather_array(elts_neg[:, :elt_count_neg].T, comm, 0, 1)
if sign_peaks in ['positive', 'both']:
gdata_pos = gather_array(elts_pos[:, :elt_count_pos].T, comm, 0, 1)
if comm.rank == 0:
#DO PCA on elts and store the basis obtained.
nb_waveforms = 0
if sign_peaks in ['negative', 'both']:
nb_waveforms += gdata_neg.shape[0]
if sign_peaks in ['positive', 'both']:
nb_waveforms += gdata_pos.shape[0]
print_and_log([
"Found %d waveforms over %d requested" %
(nb_waveforms, int(nb_elts * comm.size))
], 'default', logger)
pca = PCA(output_dim, copy=False)
res = {}
if sign_peaks in ['negative', 'both']:
if len(gdata_neg) > 0:
res_pca = pca.fit_transform(gdata_neg.astype(
numpy.double)).astype(numpy.float32)
res['proj'] = pca.components_.T.astype(numpy.float32)
else:
res['proj'] = numpy.identity(N_t, dtype=numpy.float32)
res['rec'] = res['proj'].T
res['waveform'] = numpy.median(gdata_neg, 0)
idx = numpy.random.permutation(numpy.arange(
gdata_neg.shape[0]))[:1000]
res['waveforms'] = gdata_neg[idx, :]
if sign_peaks in ['positive', 'both']:
if len(gdata_pos) > 0:
res_pca = pca.fit_transform(gdata_pos.astype(
numpy.double)).astype(numpy.float32)
res['proj_pos'] = pca.components_.T.astype(numpy.float32)
else:
res['proj_pos'] = numpy.identity(N_t, dtype=numpy.float32)
res['rec_pos'] = res['proj_pos'].T
res['waveform_pos'] = numpy.median(gdata_pos, 0)
idx = numpy.random.permutation(numpy.arange(
gdata_pos.shape[0]))[:1000]
res['waveforms_pos'] = gdata_pos[idx, :]
bfile = h5py.File(file_out_suff + '.basis.hdf5', 'r+', libver='latest')
io.write_datasets(bfile, res.keys(), res)
if sign_peaks == 'positive':
print_and_log([
"A basis with %s dimensions has been built" %
res['proj_pos'].shape[1]
], 'info', logger)
elif sign_peaks == 'negative':
print_and_log([
"A basis with %s dimensions has been built" %
res['proj'].shape[1]
], 'info', logger)
elif sign_peaks == 'both':
print_and_log([
"Two basis with %s dimensions has been built" %
res['proj'].shape[1]
], 'info', logger)
bfile.close()
comm.Barrier()
if matched_filter:
if comm.rank == 0:
print_and_log([
"Because of matched filters, need to recompute the thresholds..."
], 'default', logger)
if do_spatial_whitening:
spatial_whitening = io.load_data(params, 'spatial_whitening')
if use_gpu:
spatial_whitening = cmt.CUDAMatrix(spatial_whitening,
copy_on_host=False)
if do_temporal_whitening:
temporal_whitening = io.load_data(params, 'temporal_whitening')
if sign_peaks in ['negative', 'both']:
waveform_neg = io.load_data(params, 'waveform')
waveform_neg /= (numpy.abs(numpy.sum(waveform_neg)) *
len(waveform_neg))
if sign_peaks in ['positive', 'both']:
waveform_pos = io.load_data(params, 'waveform-pos')
waveform_pos /= (numpy.abs(numpy.sum(waveform_pos)) *
len(waveform_pos))
for gidx in [all_chunks[comm.rank]]:
local_chunk, t_offset = data_file.get_data(gidx,
chunk_size,
nodes=nodes)
local_shape = len(local_chunk)
if do_spatial_whitening:
if use_gpu:
local_chunk = cmt.CUDAMatrix(local_chunk,
copy_on_host=False)
local_chunk = local_chunk.dot(spatial_whitening).asarray()
else:
local_chunk = numpy.dot(local_chunk, spatial_whitening)
if do_temporal_whitening:
local_chunk = scipy.ndimage.filters.convolve1d(
local_chunk, temporal_whitening, axis=0, mode='constant')
if sign_peaks in ['negative', 'both']:
tmp_chunk = scipy.ndimage.filters.convolve1d(local_chunk,
waveform_neg,
axis=0,
mode='constant')
thresholds = numpy.zeros(N_e, dtype=numpy.float32)
for i in xrange(N_e):
u = numpy.median(tmp_chunk[:, i], 0)
thresholds[i] = numpy.median(
numpy.abs(tmp_chunk[:, i] - u), 0)
gdata = gather_array(thresholds, comm)
if comm.rank == 0:
gdata = gdata.reshape((comm.size, N_e))
thresholds = numpy.mean(gdata, 0)
bfile = h5py.File(file_out + '.basis.hdf5',
'r+',
libver='latest')
io.write_datasets(bfile, ['matched_thresholds'],
{'matched_thresholds': thresholds})
bfile.close()
comm.Barrier()
if sign_peaks in ['positive', 'both']:
tmp_chunk = scipy.ndimage.filters.convolve1d(local_chunk,
waveform_pos,
axis=0,
mode='constant')
thresholds = numpy.zeros(N_e, dtype=numpy.float32)
for i in xrange(N_e):
u = numpy.median(tmp_chunk[:, i], 0)
thresholds[i] = numpy.median(
numpy.abs(tmp_chunk[:, i] - u), 0)
gdata = gather_array(thresholds, comm)
if comm.rank == 0:
gdata = gdata.reshape((comm.size, N_e))
thresholds = numpy.mean(gdata, 0)
bfile = h5py.File(file_out + '.basis.hdf5',
'r+',
libver='latest')
io.write_datasets(bfile, ['matched_thresholds_pos'],
{'matched_thresholds_pos': thresholds})
bfile.close()
comm.Barrier()
data_file.close()
def main(params, nb_cpu, nb_gpu, use_gpu):
#################################################################
#params = detect_memory(params)
logger = init_logging(params.logfile)
SHARED_MEMORY = get_shared_memory_flag(params)
logger = logging.getLogger('circus.fitting')
data_file = params.data_file
data_file.open()
N_e = params.getint('data', 'N_e')
N_total = params.nb_channels
N_t = params.getint('detection', 'N_t')
template_shift = params.getint('detection', 'template_shift')
file_out = params.get('data', 'file_out')
file_out_suff = params.get('data', 'file_out_suff')
sign_peaks = params.get('detection', 'peaks')
matched_filter = params.getboolean('detection', 'matched-filter')
spike_thresh = params.getfloat('detection', 'spike_thresh')
do_temporal_whitening = params.getboolean('whitening', 'temporal')
do_spatial_whitening = params.getboolean('whitening', 'spatial')
chunk_size = params.getint('fitting', 'chunk_size')
gpu_only = params.getboolean('fitting', 'gpu_only')
nodes, edges = get_nodes_and_edges(params)
tmp_limits = params.get('fitting',
'amp_limits').replace('(',
'').replace(')',
'').split(',')
tmp_limits = map(float, tmp_limits)
amp_auto = params.getboolean('fitting', 'amp_auto')
nb_chances = params.getint('fitting', 'nb_chances')
max_chunk = params.getfloat('fitting', 'max_chunk')
noise_thr = params.getfloat('clustering', 'noise_thr')
collect_all = params.getboolean('fitting', 'collect_all')
ignore_dead_times = params.getboolean('triggers', 'ignore_times')
inv_nodes = numpy.zeros(N_total, dtype=numpy.int32)
inv_nodes[nodes] = numpy.argsort(nodes)
#################################################################
if use_gpu:
import cudamat as cmt
## Need to properly handle multi GPU per MPI nodes?
if nb_gpu > nb_cpu:
gpu_id = int(comm.rank // nb_cpu)
else:
gpu_id = 0
cmt.cuda_set_device(gpu_id)
cmt.init()
cmt.cuda_sync_threads()
if SHARED_MEMORY:
templates = io.load_data_memshared(params,
'templates',
normalize=True,
transpose=True)
N_tm, x = templates.shape
else:
templates = io.load_data(params, 'templates')
x, N_tm = templates.shape
temp_2_shift = 2 * template_shift
full_gpu = use_gpu and gpu_only
n_tm = N_tm // 2
n_scalar = N_e * N_t
temp_window = numpy.arange(-template_shift, template_shift + 1)
size_window = N_e * (2 * template_shift + 1)
if not amp_auto:
amp_limits = numpy.zeros((n_tm, 2))
amp_limits[:, 0] = tmp_limits[0]
amp_limits[:, 1] = tmp_limits[1]
else:
amp_limits = io.load_data(params, 'limits')
norm_templates = io.load_data(params, 'norm-templates')
if not SHARED_MEMORY:
for idx in xrange(templates.shape[1]):
myslice = numpy.arange(templates.indptr[idx],
templates.indptr[idx + 1])
templates.data[myslice] /= norm_templates[idx]
templates = templates.T
if matched_filter:
if sign_peaks in ['negative', 'both']:
waveform_neg = io.load_data(params, 'waveform')
waveform_neg /= (numpy.abs(numpy.sum(waveform_neg)) *
len(waveform_neg))
matched_tresholds_neg = io.load_data(params, 'matched-thresholds')
if sign_peaks in ['positive', 'both']:
waveform_pos = io.load_data(params, 'waveform-pos')
waveform_pos /= (numpy.abs(numpy.sum(waveform_pos)) *
len(waveform_pos))
matched_tresholds_pos = io.load_data(params,
'matched-thresholds-pos')
if ignore_dead_times:
all_dead_times = get_dead_times(params)
thresholds = io.load_data(params, 'thresholds')
if collect_all:
neighbors = {}
for i in xrange(n_tm):
tmp = templates[i, :].toarray().reshape(N_e,
N_t) * norm_templates[i]
neighbors[i] = numpy.where(numpy.sum(tmp, 1) != 0)[0]
if use_gpu:
templates = cmt.SparseCUDAMatrix(templates, copy_on_host=False)
info_string = ''
if comm.rank == 0:
if use_gpu:
info_string = "using %d GPUs" % (comm.size)
else:
info_string = "using %d CPUs" % (comm.size)
comm.Barrier()
c_overlap = io.get_overlaps(params,
nb_cpu=nb_cpu,
nb_gpu=nb_gpu,
use_gpu=use_gpu)
over_shape = c_overlap.get('over_shape')[:]
N_over = int(numpy.sqrt(over_shape[0]))
S_over = over_shape[1]
## If the number of overlaps is different from templates, we need to recompute them
if N_over != N_tm:
if comm.rank == 0:
print_and_log(
['Templates have been modified, recomputing the overlaps...'],
'default', logger)
c_overlap = io.get_overlaps(params,
erase=True,
nb_cpu=nb_cpu,
nb_gpu=nb_gpu,
use_gpu=use_gpu)
over_shape = c_overlap.get('over_shape')[:]
N_over = int(numpy.sqrt(over_shape[0]))
S_over = over_shape[1]
if SHARED_MEMORY:
c_overs = io.load_data_memshared(params, 'overlaps')
else:
c_overs = io.load_data(params, 'overlaps')
comm.Barrier()
if n_tm == 0:
if comm.rank == 0:
print_and_log(["No templates present. Redo clustering?"],
'default', logger)
sys.exit(0)
if comm.rank == 0:
print_and_log([
"Here comes the SpyKING CIRCUS %s and %d templates..." %
(info_string, n_tm)
], 'default', logger)
purge(file_out_suff, '.data')
if do_spatial_whitening:
spatial_whitening = io.load_data(params, 'spatial_whitening')
if do_temporal_whitening:
temporal_whitening = io.load_data(params, 'temporal_whitening')
if full_gpu:
try:
# If memory on the GPU is large enough, we load the overlaps onto it
for i in xrange(N_over):
c_overs[i] = cmt.SparseCUDAMatrix(c_overs[i],
copy_on_host=False)
except Exception:
if comm.rank == 0:
print_and_log([
"Not enough memory on GPUs: GPUs are used for projection only"
], 'info', logger)
for i in xrange(N_over):
if c_overs.has_key(i):
del c_overs[i]
full_gpu = False
nb_chunks, last_chunk_len = data_file.analyze(chunk_size)
processed_chunks = int(min(nb_chunks, max_chunk))
comm.Barrier()
spiketimes_file = open(file_out_suff + '.spiketimes-%d.data' % comm.rank,
'wb')
comm.Barrier()
amplitudes_file = open(file_out_suff + '.amplitudes-%d.data' % comm.rank,
'wb')
comm.Barrier()
templates_file = open(file_out_suff + '.templates-%d.data' % comm.rank,
'wb')
comm.Barrier()
if collect_all:
garbage_times_file = open(
file_out_suff + '.gspiketimes-%d.data' % comm.rank, 'wb')
comm.Barrier()
garbage_temp_file = open(
file_out_suff + '.gtemplates-%d.data' % comm.rank, 'wb')
comm.Barrier()
if use_gpu and do_spatial_whitening:
spatial_whitening = cmt.CUDAMatrix(spatial_whitening,
copy_on_host=False)
last_chunk_size = 0
to_explore = xrange(comm.rank, processed_chunks, comm.size)
if comm.rank == 0:
to_explore = get_tqdm_progressbar(to_explore)
for gcount, gidx in enumerate(to_explore):
#print "Node", comm.rank, "is analyzing chunk", gidx, "/", nb_chunks, " ..."
## We need to deal with the borders by taking chunks of size [0, chunck_size+template_shift]
is_first = data_file.is_first_chunk(gidx, nb_chunks)
is_last = data_file.is_last_chunk(gidx, nb_chunks)
if is_last:
padding = (-temp_2_shift, 0)
elif is_first:
padding = (0, temp_2_shift)
else:
padding = (-temp_2_shift, temp_2_shift)
result = {'spiketimes': [], 'amplitudes': [], 'templates': []}
local_chunk, t_offset = data_file.get_data(gidx,
chunk_size,
padding,
nodes=nodes)
len_chunk = len(local_chunk)
if do_spatial_whitening:
if use_gpu:
local_chunk = cmt.CUDAMatrix(local_chunk, copy_on_host=False)
local_chunk = local_chunk.dot(spatial_whitening).asarray()
else:
local_chunk = numpy.dot(local_chunk, spatial_whitening)
if do_temporal_whitening:
local_chunk = scipy.ndimage.filters.convolve1d(local_chunk,
temporal_whitening,
axis=0,
mode='constant')
#print "Extracting the peaks..."
if collect_all:
all_found_spikes = {}
for i in xrange(N_e):
all_found_spikes[i] = []
local_peaktimes = numpy.zeros(0, dtype=numpy.uint32)
if matched_filter:
if sign_peaks in ['positive', 'both']:
filter_chunk = scipy.ndimage.filters.convolve1d(
local_chunk, waveform_pos, axis=0, mode='constant')
for i in xrange(N_e):
peaktimes = algo.detect_peaks(filter_chunk[:, i],
matched_tresholds_pos[i])
local_peaktimes = numpy.concatenate(
(local_peaktimes, peaktimes))
if collect_all:
all_found_spikes[i] += peaktimes.tolist()
if sign_peaks in ['negative', 'both']:
filter_chunk = scipy.ndimage.filters.convolve1d(
local_chunk, waveform_neg, axis=0, mode='constant')
for i in xrange(N_e):
peaktimes = algo.detect_peaks(filter_chunk[:, i],
matched_tresholds_neg[i])
local_peaktimes = numpy.concatenate(
(local_peaktimes, peaktimes))
if collect_all:
all_found_spikes[i] += peaktimes.tolist()
else:
for i in xrange(N_e):
if sign_peaks == 'negative':
peaktimes = algo.detect_peaks(local_chunk[:, i],
thresholds[i],
valley=True)
elif sign_peaks == 'positive':
peaktimes = algo.detect_peaks(local_chunk[:, i],
thresholds[i],
valley=False)
elif sign_peaks == 'both':
peaktimes = algo.detect_peaks(numpy.abs(local_chunk[:, i]),
thresholds[i],
valley=False)
local_peaktimes = numpy.concatenate(
(local_peaktimes, peaktimes))
if collect_all:
all_found_spikes[i] += peaktimes.tolist()
local_peaktimes = numpy.unique(local_peaktimes)
g_offset = t_offset + padding[0]
if ignore_dead_times:
dead_indices = numpy.searchsorted(
all_dead_times, [t_offset, t_offset + chunk_size])
if dead_indices[0] != dead_indices[1]:
local_peaktimes = numpy.array(list(
set(local_peaktimes + g_offset).difference(
all_dead_times[dead_indices[0]:dead_indices[1]])),
dtype=numpy.uint32) - g_offset
local_peaktimes = numpy.sort(local_peaktimes)
#print "Removing the useless borders..."
local_borders = (template_shift, len_chunk - template_shift)
idx = (local_peaktimes >= local_borders[0]) & (local_peaktimes <
local_borders[1])
local_peaktimes = numpy.compress(idx, local_peaktimes)
if collect_all:
for i in xrange(N_e):
all_found_spikes[i] = numpy.array(all_found_spikes[i],
dtype=numpy.uint32)
if ignore_dead_times:
if dead_indices[0] != dead_indices[1]:
all_found_spikes[i] = numpy.array(
list(
set(all_found_spikes[i] + g_offset).difference(
all_dead_times[
dead_indices[0]:dead_indices[1]])),
dtype=numpy.uint32) - g_offset
all_found_spikes[i] = numpy.sort(all_found_spikes[i])
idx = (all_found_spikes[i] >= local_borders[0]) & (
all_found_spikes[i] < local_borders[1])
all_found_spikes[i] = numpy.compress(idx, all_found_spikes[i])
n_t = len(local_peaktimes)
if full_gpu:
# all_indices = cmt.CUDAMatrix(all_indices)
tmp_gpu = cmt.CUDAMatrix(local_peaktimes.reshape((1, n_t)),
copy_on_host=False)
if n_t > 0:
#print "Computing the b (should full_gpu by putting all chunks on GPU if possible?)..."
if collect_all:
c_local_chunk = local_chunk.copy()
local_chunk = local_chunk.T.ravel()
sub_mat = numpy.zeros((size_window, n_t), dtype=numpy.float32)
if len_chunk != last_chunk_size:
slice_indices = numpy.zeros(0, dtype=numpy.int32)
for idx in xrange(N_e):
slice_indices = numpy.concatenate(
(slice_indices, len_chunk * idx + temp_window))
last_chunk_size = len_chunk
for count, idx in enumerate(local_peaktimes):
sub_mat[:, count] = numpy.take(local_chunk,
slice_indices + idx)
del local_chunk
if use_gpu:
sub_mat = cmt.CUDAMatrix(sub_mat, copy_on_host=False)
b = cmt.sparse_dot(templates, sub_mat)
else:
b = templates.dot(sub_mat)
del sub_mat
local_restriction = (t_offset, t_offset + chunk_size)
all_spikes = local_peaktimes + g_offset
# Because for GPU, slicing by columns is more efficient, we need to transpose b
#b = b.transpose()
if use_gpu and not full_gpu:
b = b.asarray()
failure = numpy.zeros(n_t, dtype=numpy.int32)
if full_gpu:
mask = numpy.zeros((2 * n_tm, n_t), dtype=numpy.float32)
mask[:n_tm, :] = 1
data = cmt.empty(mask.shape)
patch_gpu = b.shape[1] == 1
else:
mask = numpy.ones((n_tm, n_t), dtype=numpy.float32)
sub_b = b[:n_tm, :]
if collect_all:
c_all_times = numpy.zeros((len_chunk, N_e), dtype=numpy.bool)
c_min_times = numpy.maximum(
numpy.arange(len_chunk) - template_shift, 0)
c_max_times = numpy.minimum(
numpy.arange(len_chunk) + template_shift + 1, len_chunk)
for i in xrange(N_e):
c_all_times[all_found_spikes[i], i] = True
while (numpy.mean(failure) < nb_chances):
if full_gpu:
gpu_mask = cmt.CUDAMatrix(mask, copy_on_host=False)
b.mult(gpu_mask, data)
tmp_mat = data.max(0)
argmax_bi = numpy.argsort(tmp_mat.asarray()[0, :])[::-1]
del tmp_mat
else:
data = sub_b * mask
argmax_bi = numpy.argsort(numpy.max(data, 0))[::-1]
for peak_index in argmax_bi:
if full_gpu:
b_array = b.asarray()
sub_b = b_array[:n_tm, :]
peak_scalar_products = np.take(sub_b, peak_index, axis=1)
best_template_index = np.argmax(peak_scalar_products,
axis=0)
best_template2_index = best_template_index + n_tm
if full_gpu:
best_amp = sub_b[best_template_index,
peak_index] / n_scalar
best_amp2 = b_array[best_template_index,
peak_index] / n_scalar
else:
best_amp = sub_b[best_template_index,
peak_index] / n_scalar
best_amp2 = b[best_template2_index,
peak_index] / n_scalar
best_amp_n = best_amp / norm_templates[best_template_index]
best_amp2_n = best_amp2 / norm_templates[
best_template2_index]
# Verify amplitude constraint.
a_min = amp_limits[best_template_index, 0]
a_max = amp_limits[best_template_index, 1]
if (a_min <= best_amp_n) & (best_amp_n <= a_max):
# Keep the matching.
peak_time_step = local_peaktimes[peak_index]
data = (local_peaktimes - peak_time_step).astype(
np.int32)
is_neighbor = np.where(np.abs(data) <= temp_2_shift)[0]
idx_neighbor = data[is_neighbor] + temp_2_shift
nb_neighbors = len(is_neighbor)
indices = np.zeros((S_over, nb_neighbors),
dtype=np.int32)
indices[idx_neighbor, np.arange(nb_neighbors)] = 1
if full_gpu:
indices = cmt.CUDAMatrix(indices,
copy_on_host=False)
if patch_gpu:
b_lines = b.get_col_slice(0, b.shape[0])
else:
b_lines = b.get_col_slice(
is_neighbor[0], is_neighbor[-1] + 1)
tmp1 = cmt.sparse_dot(c_overs[best_template_index],
indices,
mult=-best_amp[keep])
tmp2 = cmt.sparse_dot(
c_overs[best_template2_index],
indices,
mult=-best_amp2[keep])
b_lines.add(tmp1.add(tmp2))
del tmp1, tmp2
else:
tmp1 = c_overs[best_template_index].multiply(
-best_amp)
tmp2 = c_overs[best_template2_index].multiply(
-best_amp2)
b[:, is_neighbor] += (tmp1 + tmp2).dot(indices)
# Add matching to the result.
t_spike = all_spikes[peak_index]
if (t_spike >= local_restriction[0]) and (
t_spike < local_restriction[1]):
#print "Accept spikes", t_spike, local_restriction, type(t_spike), t_spike > local_restriction[0], t_spike < local_restriction[1]
result['spiketimes'] += [t_spike]
result['amplitudes'] += [(best_amp_n, best_amp2_n)]
result['templates'] += [best_template_index]
# Mark current matching as tried.
mask[best_template_index, peak_index] = 0
else:
# Reject the matching.
# Update failure counter of the peak.
failure[peak_index] += 1
# If the maximal number of failures is reached then mark peak as solved (i.e. not fitted).
if failure[peak_index] == nb_chances:
mask[:, peak_index] = 0
else:
mask[best_template_index, peak_index] = 0
spikes_to_write = numpy.array(result['spiketimes'],
dtype=numpy.uint32)
amplitudes_to_write = numpy.array(result['amplitudes'],
dtype=numpy.float32)
templates_to_write = numpy.array(result['templates'],
dtype=numpy.uint32)
spiketimes_file.write(spikes_to_write.tostring())
amplitudes_file.write(amplitudes_to_write.tostring())
templates_file.write(templates_to_write.tostring())
if collect_all:
for temp, spike in zip(templates_to_write,
spikes_to_write - g_offset):
c_all_times[c_min_times[spike]:c_max_times[spike],
neighbors[temp]] = False
gspikes = numpy.where(numpy.sum(c_all_times, 1) > 0)[0]
c_all_times = numpy.take(c_all_times, gspikes, axis=0)
c_local_chunk = numpy.take(c_local_chunk, gspikes,
axis=0) * c_all_times
if sign_peaks == 'negative':
bestlecs = numpy.argmin(c_local_chunk, 1)
if matched_filter:
threshs = -matched_tresholds_neg[bestlecs]
else:
threshs = -thresholds[bestlecs]
idx = numpy.where(numpy.min(c_local_chunk, 1) < threshs)[0]
elif sign_peaks == 'positive':
bestlecs = numpy.argmax(c_local_chunk, 1)
if matched_filter:
threshs = matched_tresholds_pos[bestlecs]
else:
threshs = thresholds[bestlecs]
idx = numpy.where(numpy.max(c_local_chunk, 1) > threshs)[0]
elif sign_peaks == 'both':
c_local_chunk = numpy.abs(c_local_chunk)
bestlecs = numpy.argmax(c_local_chunk, 1)
if matched_filter:
threshs = numpy.minimum(
matched_tresholds_neg[bestlecs],
matched_tresholds_pos[bestlecs])
else:
threshs = thresholds[bestlecs]
idx = numpy.where(numpy.max(c_local_chunk, 1) > threshs)[0]
gspikes = numpy.take(gspikes, idx)
bestlecs = numpy.take(bestlecs, idx)
gspikes_to_write = numpy.array(gspikes + g_offset,
dtype=numpy.uint32)
gtemplates_to_write = numpy.array(bestlecs, dtype=numpy.uint32)
garbage_times_file.write(gspikes_to_write.tostring())
garbage_temp_file.write(gtemplates_to_write.tostring())
if full_gpu:
del gpu_mask, b, data
sys.stderr.flush()
spiketimes_file.flush()
os.fsync(spiketimes_file.fileno())
spiketimes_file.close()
amplitudes_file.flush()
os.fsync(amplitudes_file.fileno())
amplitudes_file.close()
templates_file.flush()
os.fsync(templates_file.fileno())
templates_file.close()
if collect_all:
garbage_temp_file.flush()
os.fsync(garbage_temp_file.fileno())
garbage_temp_file.close()
garbage_times_file.flush()
os.fsync(garbage_times_file.fileno())
garbage_times_file.close()
comm.Barrier()
if comm.rank == 0:
io.collect_data(comm.size, params, erase=True)
data_file.close()
def extract_juxta_spikes_(params):
'''Detect spikes from the extracellular traces'''
file_out_suff = params.get('data', 'file_out_suff')
sampling_rate = params.getint('data', 'sampling_rate')
dist_peaks = params.getint('detection', 'dist_peaks')
template_shift = params.getint('detection', 'template_shift')
juxta_dtype = params.get('validating', 'juxta_dtype')
juxta_thresh = params.getfloat('validating', 'juxta_thresh')
juxta_valley = params.getboolean('validating', 'juxta_valley')
juxta_spikes = params.get('validating', 'juxta_spikes')
juxta_filename = "{}.juxta.dat".format(file_out_suff)
beer_path = "{}.beer.hdf5".format(file_out_suff)
if juxta_spikes == '':
# Read juxtacellular trace.
juxta_data = numpy.fromfile(juxta_filename, dtype=juxta_dtype)
#juxta_data = juxta_data.astype(numpy.float32)
# juxta_data = juxta_data - dtype_offset
juxta_data = numpy.ascontiguousarray(juxta_data)
# Filter juxtacellular trace.
juxta_data = highpass(juxta_data, sampling_rate=sampling_rate)
juxta_data -= numpy.median(juxta_data)
# Compute median and median absolute deviation.
juxta_median = numpy.median(juxta_data)
juxta_ad = numpy.abs(juxta_data - juxta_median)
juxta_mad = numpy.median(juxta_ad, axis=0)
# Save medians and median absolute deviations to BEER file.
beer_file = h5py.File(beer_path, 'a', libver='latest')
if "juxta_median" in beer_file.keys():
beer_file.pop("juxta_median")
beer_file.create_dataset("juxta_median", data=juxta_median)
if "juxta_mad" in beer_file.keys():
beer_file.pop("juxta_mad")
beer_file.create_dataset("juxta_mad", data=juxta_mad)
beer_file.close()
if comm.rank == 0:
print_and_log(["Extract juxtacellular spikes"],
level='debug',
logger=logger)
# Detect juxta spike times.
threshold = juxta_thresh * juxta_mad
juxta_spike_times = algo.detect_peaks(juxta_data,
threshold,
valley=juxta_valley,
mpd=dist_peaks)
# Remove juxta spike times in the borders.
juxta_spike_times = juxta_spike_times[
template_shift <= juxta_spike_times]
juxta_spike_times = juxta_spike_times[
juxta_spike_times < juxta_data.size - template_shift]
else:
juxta_spike_times = numpy.load(juxta_spikes)
# Save juxta spike times to BEER file.
beer_file = h5py.File(beer_path, 'a', libver='latest')
group_name = "juxta_spiketimes"
if group_name in beer_file.keys():
beer_file.pop(group_name)
beer_file.create_group(group_name)
key = "{}/elec_0".format(group_name)
beer_file.create_dataset(key, data=juxta_spike_times)
beer_file.close()
# juxta_spike_values = numpy.zeros_like(juxta_spike_times, dtype='float')
# for i, t in enumerate(juxta_spike_times):
# if juxta_valley:
# juxta_spike_values[i] = - juxta_data[t]
# else:
# juxta_spike_values[i] = + juxta_data[t]
if juxta_spikes == '':
# Find juxta spike values of juxta spike times.
juxta_spike_values = juxta_data[juxta_spike_times]
if juxta_valley:
juxta_spike_values *= -1
# Save juxta spike values to BEER file.
beer_file = h5py.File(beer_path, 'a', libver='latest')
group_name = "juxta_spike_values"
if group_name in beer_file.keys():
beer_file.pop(group_name)
beer_file.create_group(group_name)
key = "{}/elec_0".format(group_name)
beer_file.create_dataset(key, data=juxta_spike_values)
beer_file.close()
return
def extract_chunk_spikes(gidx, extra_thresh, valley=True):
"""Detect spikes from a chunk of the extracellular traces"""
loc_chunk, t_offset = data_file.get_data(gidx, chunk_size, nodes=nodes)
loc_shape = len(loc_chunk)
# Whiten signal.
if do_spatial_whitening:
loc_chunk = numpy.dot(loc_chunk, spatial_whitening)
if do_temporal_whitening:
loc_chunk = scipy.ndimage.filters.convolve1d(loc_chunk,
temporal_whitening,
axis=0,
mode='constant')
##### TODO: uncomment or remove temporary zone
# # For each electrode, center traces by removing the medians.
# extra_medians = numpy.median(loc_chunk, axis=0)
# loc_chunk = loc_chunk - extra_medians
##### end temporary zone
# Preallocation for results.
peak_times = N_elec * [None]
peak_channels = N_elec * [None]
# For each electrode.
for e in xrange(N_elec):
# Extract the peaks of the current chunk.
threshold = extra_thresh * extra_mads[e]
peak_times[e] = algo.detect_peaks(loc_chunk[:, e],
threshold,
valley=valley,
mpd=dist_peaks)
peak_channels[e] = e * numpy.ones(peak_times[e].size, dtype='int')
peak_values = loc_chunk[peak_times[e], e]
if valley:
peak_indices = numpy.where(-10.0 * threshold <= peak_values)[0]
else:
peak_indices = numpy.where(peak_values <= +10.0 * threshold)[0]
peak_times[e] = peak_times[e][peak_indices]
peak_channels[e] = peak_channels[e][peak_indices]
peak_times = numpy.concatenate(peak_times)
peak_channels = numpy.concatenate(peak_channels)
# Remove the useless borders.
if alignment:
loc_borders = (2 * template_shift, loc_shape - 2 * template_shift)
else:
loc_borders = (template_shift, loc_shape - template_shift)
peak_flags = (loc_borders[0] <= peak_times) & (peak_times <
loc_borders[1])
peak_times = numpy.compress(peak_flags, peak_times)
peak_channels = numpy.compress(peak_flags, peak_channels)
# Filter unique peak times.
loc_peak_times = numpy.unique(peak_times)
##### TODO: remove debug zone
# if gidx < 1:
# numpy.save("tmp/loc_peak_times_{}_{}_.npy".format(gidx, int(extra_thresh)), loc_peak_times)
##### end debug zone
n_times = len(loc_peak_times)
loc_peak_flags = numpy.zeros(n_times, dtype='bool')
loc_peak_elecs = numpy.zeros(n_times, dtype='int')
loc_peak_values = numpy.zeros(n_times, dtype='float')
if 0 < len(loc_peak_times):
diff_times = loc_peak_times[-1] - loc_peak_times[0]
all_times = numpy.zeros((N_elec, diff_times + 1), dtype='bool')
min_times = numpy.maximum(
loc_peak_times - loc_peak_times[0] - safety_time, 0)
max_times = numpy.minimum(
loc_peak_times - loc_peak_times[0] + safety_time + 1,
diff_times)
# Shuffle peaks.
##### TODO: clean temporary zone
# argmax_peak = numpy.random.permutation(numpy.arange(n_times))
if valley:
for i, loc_peak_time in enumerate(loc_peak_times):
loc_peak_values[i] = numpy.amin(
loc_chunk[loc_peak_time, :])
argmax_peak = numpy.argsort(loc_peak_values)
else:
for i, loc_peak_time in enumerate(loc_peak_times):
loc_peak_values[i] = numpy.amax(
loc_chunk[loc_peak_time, :])
argmax_peak = numpy.argsort(loc_peak_values)
argmes_peak = argmax_peak[::-1]
##### end temporary zone
all_indices = loc_peak_times[argmax_peak]
# Select peaks with spatio-temporal masks.
for peak_index, peak_time in zip(argmax_peak, all_indices):
# Select electrode showing lowest amplitude.
if valley:
elec = numpy.argmin(loc_chunk[peak_time, :])
else:
elec = numpy.argmax(loc_chunk[peak_time, :])
_, neighs = get_neighbors(params, chan=elec)
if safety_space:
mslice = all_times[
neighs, min_times[peak_index]:max_times[peak_index]]
else:
mslice = all_times[
elec, min_times[peak_index]:max_times[peak_index]]
is_local_min = (elec in peak_channels[peak_times == peak_time])
if is_local_min and not mslice.any():
loc_peak_flags[peak_index] = True
loc_peak_elecs[peak_index] = elec
if valley:
loc_peak_values[peak_index] = -loc_chunk[peak_time,
elec]
else:
loc_peak_values[peak_index] = loc_chunk[peak_time,
elec]
if safety_space:
all_times[
neighs,
min_times[peak_index]:max_times[peak_index]] = True
# all_times[elec, min_times[peak_index]:max_times[peak_index]] = True
else:
all_times[
elec,
min_times[peak_index]:max_times[peak_index]] = True
loc_peak_times = numpy.compress(loc_peak_flags, loc_peak_times)
loc_peak_elecs = numpy.compress(loc_peak_flags, loc_peak_elecs)
loc_peak_values = numpy.compress(loc_peak_flags, loc_peak_values)
##### TODO: remove debug zone
# if gidx < 1:
# numpy.save("tmp/loc_peak_times_{}_{}.npy".format(gidx, int(extra_thresh)), loc_peak_times)
# numpy.save("tmp/loc_peak_elecs_{}_{}.npy".format(gidx, int(extra_thresh)), loc_peak_elecs)
# numpy.save("tmp/loc_peak_values_{}_{}.npy".format(gidx, int(extra_thresh)), loc_peak_values)
# numpy.save("tmp/loc_chunk_{}_{}.npy".format(gidx, int(extra_thresh)), loc_chunk)
##### end debug zone
return loc_peak_times + t_offset, loc_peak_elecs, loc_peak_values
def main(filename, params, nb_cpu, nb_gpu, use_gpu):
try:
SHARED_MEMORY = True
MPI.Win.Allocate_shared(1, 1, MPI.INFO_NULL, MPI.COMM_SELF).Free()
except NotImplementedError:
SHARED_MEMORY = False
#################################################################
sampling_rate = params.getint('data', 'sampling_rate')
N_e = params.getint('data', 'N_e')
N_t = params.getint('data', 'N_t')
N_total = params.getint('data', 'N_total')
template_shift = params.getint('data', 'template_shift')
file_out = params.get('data', 'file_out')
file_out_suff = params.get('data', 'file_out_suff')
sign_peaks = params.get('detection', 'peaks')
matched_filter = params.getboolean('detection', 'matched-filter')
spike_thresh = params.getfloat('detection', 'spike_thresh')
do_temporal_whitening = params.getboolean('whitening', 'temporal')
do_spatial_whitening = params.getboolean('whitening', 'spatial')
chunk_size = int(params.getfloat('fitting', 'chunk') * sampling_rate)
gpu_only = params.getboolean('fitting', 'gpu_only')
nodes, edges = io.get_nodes_and_edges(params)
tmp_limits = params.get('fitting',
'amp_limits').replace('(',
'').replace(')',
'').split(',')
tmp_limits = map(float, tmp_limits)
amp_auto = params.getboolean('fitting', 'amp_auto')
space_explo = params.getfloat('fitting', 'space_explo')
nb_chances = params.getint('fitting', 'nb_chances')
max_chunk = params.getfloat('fitting', 'max_chunk')
inv_nodes = numpy.zeros(N_total, dtype=numpy.int32)
inv_nodes[nodes] = numpy.argsort(nodes)
#################################################################
if use_gpu:
import cudamat as cmt
## Need to properly handle multi GPU per MPI nodes?
if nb_gpu > nb_cpu:
gpu_id = int(comm.rank // nb_cpu)
else:
gpu_id = 0
cmt.cuda_set_device(gpu_id)
cmt.init()
cmt.cuda_sync_threads()
if SHARED_MEMORY:
templates = io.load_data_memshared(params,
comm,
'templates',
normalize=True,
transpose=True)
N_tm, x = templates.shape
else:
templates = io.load_data(params, 'templates')
x, N_tm = templates.shape
N_e = params.getint('data', 'N_e')
N_t = params.getint('data', 'N_t')
template_shift = int((N_t - 1) // 2)
temp_2_shift = 2 * template_shift
full_gpu = use_gpu and gpu_only
n_tm = N_tm // 2
n_scalar = N_e * N_t
last_spikes = numpy.zeros((n_tm, 1), dtype=numpy.int32)
temp_window = numpy.arange(-template_shift, template_shift + 1)
if not amp_auto:
amp_limits = numpy.zeros((n_tm, 2))
amp_limits[:, 0] = tmp_limits[0]
amp_limits[:, 1] = tmp_limits[1]
else:
amp_limits = io.load_data(params, 'limits')
norm_templates = io.load_data(params, 'norm-templates')
if not SHARED_MEMORY:
for idx in xrange(templates.shape[1]):
myslice = numpy.arange(templates.indptr[idx],
templates.indptr[idx + 1])
templates.data[myslice] /= norm_templates[idx]
templates = templates.T
if use_gpu:
templates = cmt.SparseCUDAMatrix(templates)
info_string = ''
if matched_filter:
if sign_peaks in ['negative', 'both']:
waveform_neg = io.load_data(params, 'waveform')
waveform_neg /= (numpy.abs(numpy.sum(waveform_neg)) *
len(waveform_neg))
matched_tresholds_neg = io.load_data(params, 'matched-thresholds')
if sign_peaks in ['positive', 'both']:
waveform_pos = io.load_data(params, 'waveform-pos')
waveform_pos /= (numpy.abs(numpy.sum(waveform_pos)) *
len(waveform_pos))
matched_tresholds_pos = io.load_data(params,
'matched-thresholds-pos')
if comm.rank == 0:
if use_gpu:
info_string = "using %d GPUs" % (comm.size)
else:
info_string = "using %d CPUs" % (comm.size)
comm.Barrier()
thresholds = io.load_data(params, 'thresholds')
if SHARED_MEMORY:
c_overs = io.load_data_memshared(params,
comm,
'overlaps',
nb_cpu=nb_cpu,
nb_gpu=nb_gpu,
use_gpu=use_gpu)
c_overlap = io.get_overlaps(comm,
params,
nb_cpu=nb_cpu,
nb_gpu=nb_gpu,
use_gpu=use_gpu)
over_shape = c_overlap.get('over_shape')[:]
N_over = int(numpy.sqrt(over_shape[0]))
S_over = over_shape[1]
else:
c_overlap = io.get_overlaps(comm,
params,
nb_cpu=nb_cpu,
nb_gpu=nb_gpu,
use_gpu=use_gpu)
over_x = c_overlap.get('over_x')[:]
over_y = c_overlap.get('over_y')[:]
over_data = c_overlap.get('over_data')[:]
over_shape = c_overlap.get('over_shape')[:]
N_over = int(numpy.sqrt(over_shape[0]))
S_over = over_shape[1]
c_overlap.close()
# To be faster, we rearrange the overlaps into a dictionnary. This has a cost: twice the memory usage for
# a short period of time
c_overs = {}
overlaps = scipy.sparse.csr_matrix(
(over_data, (over_x, over_y)),
shape=(over_shape[0], over_shape[1]))
del over_x, over_y, over_data
for i in xrange(N_over):
c_overs[i] = overlaps[i * N_over:(i + 1) * N_over]
del overlaps
comm.Barrier()
if comm.rank == 0:
io.print_and_log([
"Here comes the SpyKING CIRCUS %s and %d templates..." %
(info_string, n_tm)
], 'default', params)
io.purge(file_out_suff, '.data')
if do_spatial_whitening:
spatial_whitening = io.load_data(params, 'spatial_whitening')
if do_temporal_whitening:
temporal_whitening = io.load_data(params, 'temporal_whitening')
if full_gpu:
try:
# If memory on the GPU is large enough, we load the overlaps onto it
for i in xrange(N_over):
c_overs[i] = cmt.SparseCUDAMatrix(c_overs[i])
except Exception:
if comm.rank == 0:
io.print_and_log([
"Not enough memory on GPUs: GPUs are used for projection only"
], 'info', params)
for i in xrange(N_over):
if c_overs.has_key(i):
del c_overs[i]
full_gpu = False
borders, nb_chunks, chunk_len, last_chunk_len = io.analyze_data(
params, chunk_size)
nb_chunks = int(min(nb_chunks, max_chunk))
if comm.rank == 0:
pbar = get_progressbar(int(nb_chunks // comm.size))
spiketimes_file = open(file_out_suff + '.spiketimes-%d.data' % comm.rank,
'wb')
amplitudes_file = open(file_out_suff + '.amplitudes-%d.data' % comm.rank,
'wb')
templates_file = open(file_out_suff + '.templates-%d.data' % comm.rank,
'wb')
comm.Barrier()
if use_gpu and do_spatial_whitening:
spatial_whitening = cmt.CUDAMatrix(spatial_whitening,
copy_on_host=False)
last_chunk_size = 0
for gcount, gidx in enumerate(xrange(comm.rank, nb_chunks, comm.size)):
#print "Node", comm.rank, "is analyzing chunk", gidx, "/", nb_chunks, " ..."
## We need to deal with the borders by taking chunks of size [0, chunck_size+template_shift]
if gidx == (nb_chunks - 1):
padding = (-2 * borders, 0)
elif gidx == 0:
padding = (0, 2 * borders)
else:
padding = (-2 * borders, 2 * borders)
result = {'spiketimes': [], 'amplitudes': [], 'templates': []}
local_chunk, local_shape = io.load_chunk(params,
gidx,
chunk_len,
chunk_size,
padding,
nodes=nodes)
if do_spatial_whitening:
if use_gpu:
local_chunk = cmt.CUDAMatrix(local_chunk, copy_on_host=False)
local_chunk = local_chunk.dot(spatial_whitening).asarray()
else:
local_chunk = numpy.dot(local_chunk, spatial_whitening)
if do_temporal_whitening:
local_chunk = scipy.ndimage.filters.convolve1d(local_chunk,
temporal_whitening,
axis=0,
mode='constant')
#print "Extracting the peaks..."
local_peaktimes = numpy.zeros(0, dtype=numpy.int32)
if matched_filter:
if sign_peaks in ['positive', 'both']:
filter_chunk = scipy.ndimage.filters.convolve1d(
local_chunk, waveform_pos, axis=0, mode='constant')
for i in xrange(N_e):
peaktimes = algo.detect_peaks(filter_chunk[:, i],
matched_tresholds_pos[i])
local_peaktimes = numpy.concatenate(
(local_peaktimes, peaktimes))
if sign_peaks in ['negative', 'both']:
filter_chunk = scipy.ndimage.filters.convolve1d(
local_chunk, waveform_neg, axis=0, mode='constant')
for i in xrange(N_e):
peaktimes = algo.detect_peaks(filter_chunk[:, i],
matched_tresholds_neg[i])
local_peaktimes = numpy.concatenate(
(local_peaktimes, peaktimes))
else:
for i in xrange(N_e):
if sign_peaks == 'negative':
peaktimes = algo.detect_peaks(local_chunk[:, i],
thresholds[i],
valley=True)
elif sign_peaks == 'positive':
peaktimes = algo.detect_peaks(local_chunk[:, i],
thresholds[i],
valley=False)
elif sign_peaks == 'both':
peaktimes = algo.detect_peaks(numpy.abs(local_chunk[:, i]),
thresholds[i],
valley=False)
local_peaktimes = numpy.concatenate(
(local_peaktimes, peaktimes))
local_peaktimes = numpy.unique(local_peaktimes)
#print "Removing the useless borders..."
local_borders = (template_shift, local_shape - template_shift)
idx = (local_peaktimes >= local_borders[0]) & (local_peaktimes <
local_borders[1])
local_peaktimes = numpy.compress(idx, local_peaktimes)
n_t = len(local_peaktimes)
len_chunk = local_chunk.shape[0]
all_indices = numpy.arange(n_t)
if full_gpu:
# all_indices = cmt.CUDAMatrix(all_indices)
tmp_gpu = cmt.CUDAMatrix(local_peaktimes.reshape((1, n_t)),
copy_on_host=False)
if n_t > 0:
#print "Computing the b (should full_gpu by putting all chunks on GPU if possible?)..."
local_chunk = local_chunk.T.ravel()
sub_mat = numpy.zeros((N_e * (2 * template_shift + 1), n_t),
dtype=numpy.float32)
if len_chunk != last_chunk_size:
slice_indices = numpy.zeros(0, dtype=numpy.int32)
for idx in xrange(N_e):
slice_indices = numpy.concatenate(
(slice_indices, len_chunk * idx + temp_window))
last_chunk_size = len_chunk
for count, idx in enumerate(local_peaktimes):
sub_mat[:, count] = numpy.take(local_chunk,
slice_indices + idx)
del local_chunk
if use_gpu:
sub_mat = cmt.CUDAMatrix(sub_mat, copy_on_host=False)
b = cmt.sparse_dot(templates, sub_mat)
else:
b = templates.dot(sub_mat)
del sub_mat
local_offset = gidx * chunk_size + padding[0] // N_total
local_bounds = (temp_2_shift, local_shape - temp_2_shift)
all_spikes = local_peaktimes + local_offset
penalty = numpy.ones((n_tm, n_t), dtype=numpy.float32)
# Because for GPU, slicing by columns is more efficient, we need to transpose b
#b = b.transpose()
if use_gpu and not full_gpu:
b = b.asarray()
failure = numpy.zeros(n_t, dtype=numpy.int32)
if full_gpu:
mask = cmt.CUDAMatrix(penalty, copy_on_host=False)
data = cmt.empty(mask.shape)
cm_zeros = cmt.CUDAMatrix(numpy.zeros(mask.shape),
copy_on_host=False)
patch_gpu = b.shape[1] == 1
else:
mask = penalty
sub_b = b[:n_tm, :]
min_time = local_peaktimes.min()
max_time = local_peaktimes.max()
local_len = max_time - min_time + 1
min_times = numpy.maximum(
local_peaktimes - min_time - temp_2_shift, 0)
max_times = numpy.minimum(
local_peaktimes - min_time + temp_2_shift + 1,
max_time - min_time)
max_n_t = int(space_explo * (max_time - min_time + 1) //
(2 * temp_2_shift + 1))
while (numpy.mean(failure) < nb_chances):
if full_gpu:
sub_b = b.get_row_slice(0, n_tm)
sub_b.mult(mask, data)
tmp_mat = data.max(0)
argmax_bi = numpy.argsort(tmp_mat.asarray()[0, :])[::-1]
del tmp_mat, sub_b
else:
data = sub_b * mask
argmax_bi = numpy.argsort(numpy.max(data, 0))[::-1]
while (len(argmax_bi) > 0):
subset = []
indices = []
all_times = numpy.zeros(local_len, dtype=numpy.bool)
for count, idx in enumerate(argmax_bi):
myslice = all_times[min_times[idx]:max_times[idx]]
if not myslice.any():
subset += [idx]
indices += [count]
all_times[min_times[idx]:max_times[idx]] = True
if len(subset) > max_n_t:
break
subset = numpy.array(subset, dtype=numpy.int32)
argmax_bi = numpy.delete(argmax_bi, indices)
if full_gpu:
sub_b = b.get_row_slice(0, n_tm)
tmp_mat = sub_b.argmax(0)
inds_t, inds_temp = subset, tmp_mat.asarray()[
0, :][subset].astype(numpy.int32)
del tmp_mat
else:
inds_t, inds_temp = subset, numpy.argmax(
numpy.take(sub_b, subset, axis=1), 0)
if full_gpu:
best_amp = sub_b.asarray()[inds_temp,
inds_t] / n_scalar
best_amp2 = b.asarray()[inds_temp + n_tm,
inds_t] / n_scalar
sub_mask = numpy.ones((sub_b.shape),
dtype=numpy.float32)
sub_mask[inds_temp, inds_t] = 0
sub_mask = cmt.CUDAMatrix(sub_mask, copy_on_host=False)
mask.mult(sub_mask)
del sub_mask
else:
mask[inds_temp, inds_t] = 0
best_amp = sub_b[inds_temp, inds_t] / n_scalar
best_amp2 = b[inds_temp + n_tm, inds_t] / n_scalar
best_amp_n = best_amp / numpy.take(norm_templates,
inds_temp)
best_amp2_n = best_amp2 / numpy.take(
norm_templates, inds_temp + n_tm)
all_idx = ((best_amp_n >= amp_limits[inds_temp, 0]) &
(best_amp_n <= amp_limits[inds_temp, 1]))
to_keep = numpy.where(all_idx == True)[0]
to_reject = numpy.where(all_idx == False)[0]
ts = numpy.take(local_peaktimes, inds_t[to_keep])
good = (ts >= local_bounds[0]) & (ts < local_bounds[1])
# We reduce to only the good times that will be kept
#to_keep = to_keep[good]
#ts = ts[good]
if len(ts) > 0:
if full_gpu:
tmp = cmt.CUDAMatrix(numpy.ones((len(ts), 1)),
copy_on_host=False)
tmp3 = cmt.CUDAMatrix(-ts.reshape((len(ts), 1)),
copy_on_host=False)
tmp = tmp.dot(tmp_gpu)
tmp.add_col_vec(tmp3)
condition = cmt.empty(tmp.shape)
cmt.abs(tmp, condition).less_than(temp_2_shift + 1)
condition = condition.asarray().astype(numpy.bool)
tmp = tmp.asarray().astype(numpy.int32)
else:
tmp = numpy.dot(
numpy.ones((len(ts), 1), dtype=numpy.int32),
local_peaktimes.reshape((1, n_t)))
tmp -= ts.reshape((len(ts), 1))
condition = numpy.abs(tmp) <= temp_2_shift
for count, keep in enumerate(to_keep):
idx_b = numpy.compress(condition[count, :],
all_indices)
ytmp = tmp[count,
condition[count, :]] + temp_2_shift
indices = numpy.zeros((S_over, len(ytmp)),
dtype=numpy.float32)
indices[ytmp, numpy.arange(len(ytmp))] = 1
if full_gpu:
indices = cmt.CUDAMatrix(indices,
copy_on_host=False)
if patch_gpu:
b_lines = b.get_col_slice(0, b.shape[0])
else:
b_lines = b.get_col_slice(
idx_b[0], idx_b[-1] + 1)
tmp1 = cmt.sparse_dot(c_overs[inds_temp[keep]],
indices,
mult=-best_amp[keep])
tmp2 = cmt.sparse_dot(c_overs[inds_temp[keep] +
n_tm],
indices,
mult=-best_amp2[keep])
b_lines.add(tmp1)
b_lines.add(tmp2)
del tmp1, tmp2
else:
tmp1 = c_overs[inds_temp[keep]].multiply(
-best_amp[keep]).dot(indices)
tmp2 = c_overs[inds_temp[keep] +
n_tm].multiply(-best_amp2[keep]
).dot(indices)
b[:, idx_b] += tmp1 + tmp2
if good[count]:
t_spike = ts[count] + local_offset
result['spiketimes'] += [t_spike]
result['amplitudes'] += [(best_amp_n[keep],
best_amp2_n[keep])]
result['templates'] += [inds_temp[keep]]
myslice = numpy.take(inds_t, to_reject)
failure[myslice] += 1
sub_idx = (numpy.take(failure, myslice) >= nb_chances)
if full_gpu:
N = numpy.sum(sub_idx)
if N > 0:
cu_slice = cmt.CUDAMatrix(numpy.compress(
sub_idx, myslice).reshape(1, N),
copy_on_host=False)
mask.set_selected_columns(cu_slice, cm_zeros)
del cu_slice
else:
mask[:, numpy.compress(sub_idx, myslice)] = 0
if full_gpu:
del sub_b
spikes_to_write = numpy.array(result['spiketimes'],
dtype=numpy.uint32)
amplitudes_to_write = numpy.array(result['amplitudes'],
dtype=numpy.float32)
templates_to_write = numpy.array(result['templates'],
dtype=numpy.int32)
spiketimes_file.write(spikes_to_write.tostring())
amplitudes_file.write(amplitudes_to_write.tostring())
templates_file.write(templates_to_write.tostring())
if full_gpu:
del mask, b, cm_zeros, data
if comm.rank == 0:
pbar.update(gcount)
spiketimes_file.flush()
os.fsync(spiketimes_file.fileno())
spiketimes_file.close()
amplitudes_file.flush()
os.fsync(amplitudes_file.fileno())
amplitudes_file.close()
templates_file.flush()
os.fsync(templates_file.fileno())
templates_file.close()
comm.Barrier()
if comm.rank == 0:
pbar.finish()
if comm.rank == 0:
io.collect_data(comm.size, params, erase=True)
def main(params, nb_cpu, nb_gpu, use_gpu):
#################################################################
logger = init_logging(params.logfile)
logger = logging.getLogger('circus.fitting')
data_file = params.data_file
data_file.open()
N_e = params.getint('data', 'N_e')
N_total = params.nb_channels
N_t = params.getint('detection', 'N_t')
template_shift = params.getint('detection', 'template_shift')
file_out = params.get('data', 'file_out')
file_out_suff = params.get('data', 'file_out_suff')
sign_peaks = params.get('detection', 'peaks')
matched_filter = params.getboolean('detection', 'matched-filter')
spike_thresh = params.getfloat('detection', 'spike_thresh')
do_temporal_whitening = params.getboolean('whitening', 'temporal')
do_spatial_whitening = params.getboolean('whitening', 'spatial')
chunk_size = params.getint('fitting', 'chunk_size')
gpu_only = params.getboolean('fitting', 'gpu_only')
nodes, edges = get_nodes_and_edges(params)
tmp_limits = params.get('fitting', 'amp_limits').replace('(', '').replace(')', '').split(',')
tmp_limits = map(float, tmp_limits)
amp_auto = params.getboolean('fitting', 'amp_auto')
space_explo = params.getfloat('fitting', 'space_explo')
nb_chances = params.getint('fitting', 'nb_chances')
max_chunk = params.getfloat('fitting', 'max_chunk')
noise_thr = params.getfloat('clustering', 'noise_thr')
collect_all = params.getboolean('fitting', 'collect_all')
ignore_dead_times = params.getboolean('triggers', 'ignore_times')
inv_nodes = numpy.zeros(N_total, dtype=numpy.int32)
inv_nodes[nodes] = numpy.argsort(nodes)
#################################################################
if use_gpu:
import cudamat as cmt
## Need to properly handle multi GPU per MPI nodes?
if nb_gpu > nb_cpu:
gpu_id = int(comm.rank//nb_cpu)
else:
gpu_id = 0
cmt.cuda_set_device(gpu_id)
cmt.init()
cmt.cuda_sync_threads()
if SHARED_MEMORY:
templates = io.load_data_memshared(params, 'templates', normalize=True, transpose=True)
N_tm, x = templates.shape
else:
templates = io.load_data(params, 'templates')
x, N_tm = templates.shape
temp_2_shift = 2*template_shift
full_gpu = use_gpu and gpu_only
n_tm = N_tm//2
n_scalar = N_e*N_t
last_spikes = numpy.zeros((n_tm, 1), dtype=numpy.int32)
temp_window = numpy.arange(-template_shift, template_shift+1)
if not amp_auto:
amp_limits = numpy.zeros((n_tm, 2))
amp_limits[:, 0] = tmp_limits[0]
amp_limits[:, 1] = tmp_limits[1]
else:
amp_limits = io.load_data(params, 'limits')
norm_templates = io.load_data(params, 'norm-templates')
if not SHARED_MEMORY:
for idx in xrange(templates.shape[1]):
myslice = numpy.arange(templates.indptr[idx], templates.indptr[idx+1])
templates.data[myslice] /= norm_templates[idx]
templates = templates.T
if matched_filter:
if sign_peaks in ['negative', 'both']:
waveform_neg = io.load_data(params, 'waveform')
waveform_neg /= (numpy.abs(numpy.sum(waveform_neg))* len(waveform_neg))
matched_tresholds_neg = io.load_data(params, 'matched-thresholds')
if sign_peaks in ['positive', 'both']:
waveform_pos = io.load_data(params, 'waveform-pos')
waveform_pos /= (numpy.abs(numpy.sum(waveform_pos))* len(waveform_pos))
matched_tresholds_pos = io.load_data(params, 'matched-thresholds-pos')
if ignore_dead_times:
dead_times = numpy.loadtxt(params.get('triggers', 'dead_file'))
if len(dead_times.shape) == 1:
dead_times = dead_times.reshape(1, 2)
dead_in_ms = params.getboolean('triggers', 'dead_in_ms')
if dead_in_ms:
dead_times *= numpy.int64(data_file.sampling_rate*1e-3)
dead_times = dead_times.astype(numpy.int64)
all_dead_times = []
for i in xrange(len(dead_times)):
all_dead_times += range(dead_times[i, 0], dead_times[i, 1])
thresholds = io.load_data(params, 'thresholds')
if collect_all:
neighbors = {}
for i in xrange(n_tm):
tmp = templates[i, :].toarray().reshape(N_e, N_t) * norm_templates[i]
neighbors[i] = numpy.where(numpy.sum(tmp, 1) != 0)[0]
if use_gpu:
templates = cmt.SparseCUDAMatrix(templates, copy_on_host=False)
info_string = ''
if comm.rank == 0:
if use_gpu:
info_string = "using %d GPUs" %(comm.size)
else:
info_string = "using %d CPUs" %(comm.size)
comm.Barrier()
c_overlap = io.get_overlaps(params, nb_cpu=nb_cpu, nb_gpu=nb_gpu, use_gpu=use_gpu)
over_shape = c_overlap.get('over_shape')[:]
N_over = int(numpy.sqrt(over_shape[0]))
S_over = over_shape[1]
## If the number of overlaps is different from templates, we need to recompute them
if N_over != N_tm:
if comm.rank == 0:
print_and_log(['Templates have been modified, recomputing the overlaps...'], 'default', logger)
c_overlap = io.get_overlaps(params, erase=True, nb_cpu=nb_cpu, nb_gpu=nb_gpu, use_gpu=use_gpu)
over_shape = c_overlap.get('over_shape')[:]
N_over = int(numpy.sqrt(over_shape[0]))
S_over = over_shape[1]
if SHARED_MEMORY:
c_overs = io.load_data_memshared(params, 'overlaps', nb_cpu=nb_cpu, nb_gpu=nb_gpu, use_gpu=use_gpu)
else:
c_overlap = io.get_overlaps(params, nb_cpu=nb_cpu, nb_gpu=nb_gpu, use_gpu=use_gpu)
over_x = c_overlap.get('over_x')[:]
over_y = c_overlap.get('over_y')[:]
over_data = c_overlap.get('over_data')[:]
over_shape = c_overlap.get('over_shape')[:]
c_overlap.close()
# To be faster, we rearrange the overlaps into a dictionnary. This has a cost: twice the memory usage for
# a short period of time
c_overs = {}
overlaps = scipy.sparse.csr_matrix((over_data, (over_x, over_y)), shape=(over_shape[0], over_shape[1]))
del over_x, over_y, over_data
for i in xrange(N_over):
c_overs[i] = overlaps[i*N_over:(i+1)*N_over]
del overlaps
comm.Barrier()
if comm.rank == 0:
print_and_log(["Here comes the SpyKING CIRCUS %s and %d templates..." %(info_string, n_tm)], 'default', logger)
purge(file_out_suff, '.data')
if do_spatial_whitening:
spatial_whitening = io.load_data(params, 'spatial_whitening')
if do_temporal_whitening:
temporal_whitening = io.load_data(params, 'temporal_whitening')
if full_gpu:
try:
# If memory on the GPU is large enough, we load the overlaps onto it
for i in xrange(N_over):
c_overs[i] = cmt.SparseCUDAMatrix(c_overs[i], copy_on_host=False)
except Exception:
if comm.rank == 0:
print_and_log(["Not enough memory on GPUs: GPUs are used for projection only"], 'info', logger)
for i in xrange(N_over):
if c_overs.has_key(i):
del c_overs[i]
full_gpu = False
nb_chunks, last_chunk_len = data_file.analyze(chunk_size)
processed_chunks = int(min(nb_chunks, max_chunk))
comm.Barrier()
spiketimes_file = open(file_out_suff + '.spiketimes-%d.data' %comm.rank, 'wb')
comm.Barrier()
amplitudes_file = open(file_out_suff + '.amplitudes-%d.data' %comm.rank, 'wb')
comm.Barrier()
templates_file = open(file_out_suff + '.templates-%d.data' %comm.rank, 'wb')
comm.Barrier()
if collect_all:
garbage_times_file = open(file_out_suff + '.gspiketimes-%d.data' %comm.rank, 'wb')
comm.Barrier()
garbage_temp_file = open(file_out_suff + '.gtemplates-%d.data' %comm.rank, 'wb')
comm.Barrier()
if use_gpu and do_spatial_whitening:
spatial_whitening = cmt.CUDAMatrix(spatial_whitening, copy_on_host=False)
last_chunk_size = 0
to_explore = xrange(comm.rank, processed_chunks, comm.size)
if comm.rank == 0:
to_explore = get_tqdm_progressbar(to_explore)
for gcount, gidx in enumerate(to_explore):
#print "Node", comm.rank, "is analyzing chunk", gidx, "/", nb_chunks, " ..."
## We need to deal with the borders by taking chunks of size [0, chunck_size+template_shift]
is_first = data_file.is_first_chunk(gidx, nb_chunks)
is_last = data_file.is_last_chunk(gidx, nb_chunks)
if is_last:
padding = (-2*template_shift, 0)
elif is_first:
padding = (0, 2*template_shift)
else:
padding = (-2*template_shift, 2*template_shift)
result = {'spiketimes' : [], 'amplitudes' : [], 'templates' : []}
local_chunk, t_offset = data_file.get_data(gidx, chunk_size, padding, nodes=nodes)
len_chunk = len(local_chunk)
if do_spatial_whitening:
if use_gpu:
local_chunk = cmt.CUDAMatrix(local_chunk, copy_on_host=False)
local_chunk = local_chunk.dot(spatial_whitening).asarray()
else:
local_chunk = numpy.dot(local_chunk, spatial_whitening)
if do_temporal_whitening:
local_chunk = scipy.ndimage.filters.convolve1d(local_chunk, temporal_whitening, axis=0, mode='constant')
#print "Extracting the peaks..."
if collect_all:
all_found_spikes = {}
for i in xrange(N_e):
all_found_spikes[i] = []
local_peaktimes = numpy.zeros(0, dtype=numpy.int32)
if matched_filter:
if sign_peaks in ['positive', 'both']:
filter_chunk = scipy.ndimage.filters.convolve1d(local_chunk, waveform_pos, axis=0, mode='constant')
for i in xrange(N_e):
peaktimes = algo.detect_peaks(filter_chunk[:, i], matched_tresholds_pos[i])
local_peaktimes = numpy.concatenate((local_peaktimes, peaktimes))
if collect_all:
all_found_spikes[i] += peaktimes.tolist()
if sign_peaks in ['negative', 'both']:
filter_chunk = scipy.ndimage.filters.convolve1d(local_chunk, waveform_neg, axis=0, mode='constant')
for i in xrange(N_e):
peaktimes = algo.detect_peaks(filter_chunk[:, i], matched_tresholds_neg[i])
local_peaktimes = numpy.concatenate((local_peaktimes, peaktimes))
if collect_all:
all_found_spikes[i] += peaktimes.tolist()
else:
for i in xrange(N_e):
if sign_peaks == 'negative':
peaktimes = algo.detect_peaks(local_chunk[:, i], thresholds[i], valley=True)
elif sign_peaks == 'positive':
peaktimes = algo.detect_peaks(local_chunk[:, i], thresholds[i], valley=False)
elif sign_peaks == 'both':
peaktimes = algo.detect_peaks(numpy.abs(local_chunk[:, i]), thresholds[i], valley=False)
local_peaktimes = numpy.concatenate((local_peaktimes, peaktimes))
if collect_all:
all_found_spikes[i] += peaktimes.tolist()
local_peaktimes = numpy.unique(local_peaktimes)
if ignore_dead_times:
local_peaktimes = numpy.array(list(set(local_peaktimes + t_offset).difference(all_dead_times)), dtype=numpy.int32) - t_offset
local_peaktimes = numpy.sort(local_peaktimes)
#print "Removing the useless borders..."
local_borders = (template_shift, len_chunk - template_shift)
idx = (local_peaktimes >= local_borders[0]) & (local_peaktimes < local_borders[1])
local_peaktimes = numpy.compress(idx, local_peaktimes)
if collect_all:
for i in xrange(N_e):
all_found_spikes[i] = numpy.array(all_found_spikes[i], dtype=numpy.int32)
if ignore_dead_times:
all_found_spikes[i] = numpy.array(list(set(all_found_spikes[i] + t_offset).difference(all_dead_times)), dtype=numpy.int32) - t_offset
all_found_spikes[i] = numpy.sort(all_found_spikes[i])
idx = (all_found_spikes[i] >= local_borders[0]) & (all_found_spikes[i] < local_borders[1])
all_found_spikes[i] = numpy.compress(idx, all_found_spikes[i])
n_t = len(local_peaktimes)
all_indices = numpy.arange(n_t)
if full_gpu:
# all_indices = cmt.CUDAMatrix(all_indices)
tmp_gpu = cmt.CUDAMatrix(local_peaktimes.reshape((1, n_t)), copy_on_host=False)
if n_t > 0:
#print "Computing the b (should full_gpu by putting all chunks on GPU if possible?)..."
if collect_all:
c_local_chunk = local_chunk.copy()
local_chunk = local_chunk.T.ravel()
sub_mat = numpy.zeros((N_e*(2*template_shift+1), n_t), dtype=numpy.float32)
if len_chunk != last_chunk_size:
slice_indices = numpy.zeros(0, dtype=numpy.int32)
for idx in xrange(N_e):
slice_indices = numpy.concatenate((slice_indices, len_chunk*idx + temp_window))
last_chunk_size = len_chunk
for count, idx in enumerate(local_peaktimes):
sub_mat[:, count] = numpy.take(local_chunk, slice_indices + idx)
del local_chunk
if use_gpu:
sub_mat = cmt.CUDAMatrix(sub_mat, copy_on_host=False)
b = cmt.sparse_dot(templates, sub_mat)
else:
b = templates.dot(sub_mat)
del sub_mat
local_offset = padding[0] + t_offset
local_bounds = (temp_2_shift, len_chunk - temp_2_shift)
all_spikes = local_peaktimes + local_offset
# Because for GPU, slicing by columns is more efficient, we need to transpose b
#b = b.transpose()
if use_gpu and not full_gpu:
b = b.asarray()
failure = numpy.zeros(n_t, dtype=numpy.int32)
if full_gpu:
mask = numpy.zeros((2*n_tm, n_t), dtype=numpy.float32)
mask[:n_tm, :] = 1
data = cmt.empty(mask.shape)
patch_gpu= b.shape[1] == 1
else:
mask = numpy.ones((n_tm, n_t), dtype=numpy.float32)
sub_b = b[:n_tm, :]
min_time = local_peaktimes.min()
max_time = local_peaktimes.max()
local_len = max_time - min_time + 1
min_times = numpy.maximum(local_peaktimes - min_time - temp_2_shift, 0)
max_times = numpy.minimum(local_peaktimes - min_time + temp_2_shift + 1, max_time - min_time)
max_n_t = int(space_explo*(max_time-min_time+1)//(2*temp_2_shift + 1))
if collect_all:
c_all_times = numpy.zeros((len_chunk, N_e), dtype=numpy.bool)
c_min_times = numpy.maximum(numpy.arange(len_chunk) - template_shift, 0)
c_max_times = numpy.minimum(numpy.arange(len_chunk) + template_shift + 1, len_chunk)
for i in xrange(N_e):
c_all_times[all_found_spikes[i], i] = True
while (numpy.mean(failure) < nb_chances):
if full_gpu:
gpu_mask = cmt.CUDAMatrix(mask, copy_on_host=False)
b.mult(gpu_mask, data)
tmp_mat = data.max(0)
argmax_bi = numpy.argsort(tmp_mat.asarray()[0, :])[::-1]
del tmp_mat
else:
data = sub_b * mask
argmax_bi = numpy.argsort(numpy.max(data, 0))[::-1]
while (len(argmax_bi) > 0):
subset = []
indices = []
all_times = numpy.zeros(local_len, dtype=numpy.bool)
for count, idx in enumerate(argmax_bi):
myslice = all_times[min_times[idx]:max_times[idx]]
if not myslice.any():
subset += [idx]
indices += [count]
all_times[min_times[idx]:max_times[idx]] = True
if len(subset) > max_n_t:
break
subset = numpy.array(subset, dtype=numpy.int32)
argmax_bi = numpy.delete(argmax_bi, indices)
if full_gpu:
b_array = b.asarray()
sub_b = b_array[:n_tm, :]
inds_t, inds_temp = subset, numpy.argmax(numpy.take(sub_b, subset, axis=1), 0)
if full_gpu:
best_amp = sub_b[inds_temp, inds_t]/n_scalar
best_amp2 = b_array[inds_temp + n_tm, inds_t]/n_scalar
else:
best_amp = sub_b[inds_temp, inds_t]/n_scalar
best_amp2 = b[inds_temp + n_tm, inds_t]/n_scalar
mask[inds_temp, inds_t] = 0
best_amp_n = best_amp/numpy.take(norm_templates, inds_temp)
best_amp2_n = best_amp2/numpy.take(norm_templates, inds_temp + n_tm)
all_idx = ((best_amp_n >= amp_limits[inds_temp, 0]) & (best_amp_n <= amp_limits[inds_temp, 1]))
to_keep = numpy.where(all_idx == True)[0]
to_reject = numpy.where(all_idx == False)[0]
ts = numpy.take(local_peaktimes, inds_t[to_keep])
good = (ts >= local_bounds[0]) & (ts < local_bounds[1])
# We reduce to only the good times that will be kept
#to_keep = to_keep[good]
#ts = ts[good]
if len(ts) > 0:
if full_gpu:
tmp = cmt.CUDAMatrix(numpy.ones((len(ts), 1)), copy_on_host=False)
tmp3 = cmt.CUDAMatrix(-ts.reshape((len(ts), 1)), copy_on_host=False)
tmp = tmp.dot(tmp_gpu)
tmp.add_col_vec(tmp3)
condition = cmt.empty(tmp.shape)
cmt.abs(tmp, condition).less_than(temp_2_shift + 1)
condition = condition.asarray().astype(numpy.bool)
tmp = tmp.asarray().astype(numpy.int32)
else:
tmp = numpy.dot(numpy.ones((len(ts), 1), dtype=numpy.int32), local_peaktimes.reshape((1, n_t)))
tmp -= ts.reshape((len(ts), 1))
condition = numpy.abs(tmp) <= temp_2_shift
for count, keep in enumerate(to_keep):
idx_b = numpy.compress(condition[count, :], all_indices)
ytmp = tmp[count, condition[count, :]] + temp_2_shift
indices = numpy.zeros((S_over, len(ytmp)), dtype=numpy.float32)
indices[ytmp, numpy.arange(len(ytmp))] = 1
if full_gpu:
indices = cmt.CUDAMatrix(indices, copy_on_host=False)
if patch_gpu:
b_lines = b.get_col_slice(0, b.shape[0])
else:
b_lines = b.get_col_slice(idx_b[0], idx_b[-1]+1)
tmp1 = cmt.sparse_dot(c_overs[inds_temp[keep]], indices, mult=-best_amp[keep])
tmp2 = cmt.sparse_dot(c_overs[inds_temp[keep] + n_tm], indices, mult=-best_amp2[keep])
b_lines.add(tmp1.add(tmp2))
del tmp1, tmp2
else:
tmp1 = c_overs[inds_temp[keep]].multiply(-best_amp[keep]).dot(indices)
tmp2 = c_overs[inds_temp[keep] + n_tm].multiply(-best_amp2[keep]).dot(indices)
b[:, idx_b] += tmp1 + tmp2
if good[count]:
t_spike = ts[count] + local_offset
result['spiketimes'] += [t_spike]
result['amplitudes'] += [(best_amp_n[keep], best_amp2_n[keep])]
result['templates'] += [inds_temp[keep]]
myslice = numpy.take(inds_t, to_reject)
failure[myslice] += 1
sub_idx = (numpy.take(failure, myslice) >= nb_chances)
mask[:, numpy.compress(sub_idx, myslice)] = 0
spikes_to_write = numpy.array(result['spiketimes'], dtype=numpy.uint32)
amplitudes_to_write = numpy.array(result['amplitudes'], dtype=numpy.float32)
templates_to_write = numpy.array(result['templates'], dtype=numpy.int32)
spiketimes_file.write(spikes_to_write.tostring())
amplitudes_file.write(amplitudes_to_write.tostring())
templates_file.write(templates_to_write.tostring())
if collect_all:
for temp, spike in zip(templates_to_write, spikes_to_write - local_offset):
c_all_times[c_min_times[spike]:c_max_times[spike], neighbors[temp]] = False
gspikes = numpy.where(numpy.sum(c_all_times, 1) > 0)[0]
c_all_times = numpy.take(c_all_times, gspikes, axis=0)
c_local_chunk = numpy.take(c_local_chunk, gspikes, axis=0) * c_all_times
if sign_peaks == 'negative':
bestlecs = numpy.argmin(c_local_chunk, 1)
if matched_filter:
threshs = -matched_tresholds_neg[bestlecs]
else:
threshs = -thresholds[bestlecs]
idx = numpy.where(numpy.min(c_local_chunk, 1) < threshs)[0]
elif sign_peaks == 'positive':
bestlecs = numpy.argmax(c_local_chunk, 1)
if matched_filter:
threshs = matched_tresholds_pos[bestlecs]
else:
threshs = thresholds[bestlecs]
idx = numpy.where(numpy.max(c_local_chunk, 1) > threshs)[0]
elif sign_peaks == 'both':
c_local_chunk = numpy.abs(c_local_chunk)
bestlecs = numpy.argmax(c_local_chunk, 1)
if matched_filter:
threshs = numpy.minimum(matched_tresholds_neg[bestlecs], matched_tresholds_pos[bestlecs])
else:
threshs = thresholds[bestlecs]
idx = numpy.where(numpy.max(c_local_chunk, 1) > threshs)[0]
gspikes = numpy.take(gspikes, idx)
bestlecs = numpy.take(bestlecs, idx)
gspikes_to_write = numpy.array(gspikes + local_offset, dtype=numpy.uint32)
gtemplates_to_write = numpy.array(bestlecs, dtype=numpy.int32)
garbage_times_file.write(gspikes_to_write.tostring())
garbage_temp_file.write(gtemplates_to_write.tostring())
if full_gpu:
del gpu_mask, b, data
spiketimes_file.flush()
os.fsync(spiketimes_file.fileno())
spiketimes_file.close()
amplitudes_file.flush()
os.fsync(amplitudes_file.fileno())
amplitudes_file.close()
templates_file.flush()
os.fsync(templates_file.fileno())
templates_file.close()
if collect_all:
garbage_temp_file.flush()
os.fsync(garbage_temp_file.fileno())
garbage_temp_file.close()
garbage_times_file.flush()
os.fsync(garbage_times_file.fileno())
garbage_times_file.close()
comm.Barrier()
if comm.rank == 0:
io.collect_data(comm.size, params, erase=True)
data_file.close()
def main(params, nb_cpu, nb_gpu, use_gpu):
#################################################################
#params = detect_memory(params)
logger = init_logging(params.logfile)
SHARED_MEMORY = get_shared_memory_flag(params)
logger = logging.getLogger('circus.fitting')
data_file = params.data_file
N_e = params.getint('data', 'N_e')
N_total = params.nb_channels
N_t = params.getint('detection', 'N_t')
template_shift = params.getint('detection', 'template_shift')
file_out = params.get('data', 'file_out')
file_out_suff = params.get('data', 'file_out_suff')
sign_peaks = params.get('detection', 'peaks')
dist_peaks = params.getint('detection', 'dist_peaks')
matched_filter = params.getboolean('detection', 'matched-filter')
spike_thresh = params.getfloat('detection', 'spike_thresh')
do_temporal_whitening = params.getboolean('whitening', 'temporal')
do_spatial_whitening = params.getboolean('whitening', 'spatial')
chunk_size = detect_memory(params)
gpu_only = params.getboolean('fitting', 'gpu_only')
nodes, edges = get_nodes_and_edges(params)
tmp_limits = params.get('fitting', 'amp_limits').replace('(', '').replace(')', '').split(',')
tmp_limits = map(float, tmp_limits)
amp_auto = params.getboolean('fitting', 'amp_auto')
nb_chances = params.getint('fitting', 'nb_chances')
max_chunk = params.getfloat('fitting', 'max_chunk')
noise_thr = params.getfloat('clustering', 'noise_thr')
collect_all = params.getboolean('fitting', 'collect_all')
ignore_dead_times = params.getboolean('triggers', 'ignore_times')
inv_nodes = numpy.zeros(N_total, dtype=numpy.int32)
inv_nodes[nodes] = numpy.arange(len(nodes))
data_file.open()
#################################################################
if use_gpu:
import cudamat as cmt
## Need to properly handle multi GPU per MPI nodes?
if nb_gpu > nb_cpu:
gpu_id = int(comm.rank//nb_cpu)
else:
gpu_id = 0
cmt.cuda_set_device(gpu_id)
cmt.init()
cmt.cuda_sync_threads()
if matched_filter:
if sign_peaks in ['negative', 'both']:
waveform_neg = io.load_data(params, 'waveform')
waveform_neg /= (numpy.abs(numpy.sum(waveform_neg))* len(waveform_neg))
matched_tresholds_neg = io.load_data(params, 'matched-thresholds')
if sign_peaks in ['positive', 'both']:
waveform_pos = io.load_data(params, 'waveform-pos')
waveform_pos /= (numpy.abs(numpy.sum(waveform_pos))* len(waveform_pos))
matched_tresholds_pos = io.load_data(params, 'matched-thresholds-pos')
if ignore_dead_times:
all_dead_times = get_dead_times(params)
thresholds = io.load_data(params, 'thresholds')
comm.Barrier()
if comm.rank == 0:
print_and_log(["Extracting MUA activity..."], 'default', logger)
purge(file_out_suff, '.data')
if do_spatial_whitening:
spatial_whitening = io.load_data(params, 'spatial_whitening')
if do_temporal_whitening:
temporal_whitening = io.load_data(params, 'temporal_whitening')
nb_chunks, last_chunk_len = data_file.analyze(chunk_size)
processed_chunks = int(min(nb_chunks, max_chunk))
comm.Barrier()
spiketimes_file = open(file_out_suff + '.mua-%d.data' %comm.rank, 'wb')
comm.Barrier()
electrodes_file = open(file_out_suff + '.elec-%d.data' %comm.rank, 'wb')
comm.Barrier()
amplitudes_file = open(file_out_suff + '.amp-%d.data' %comm.rank, 'wb')
comm.Barrier()
if use_gpu and do_spatial_whitening:
spatial_whitening = cmt.CUDAMatrix(spatial_whitening, copy_on_host=False)
to_explore = xrange(comm.rank, processed_chunks, comm.size)
if comm.rank == 0:
to_explore = get_tqdm_progressbar(to_explore)
for gcount, gidx in enumerate(to_explore):
#print "Node", comm.rank, "is analyzing chunk", gidx, "/", nb_chunks, " ..."
## We need to deal with the borders by taking chunks of size [0, chunck_size+template_shift]
is_first = data_file.is_first_chunk(gidx, nb_chunks)
is_last = data_file.is_last_chunk(gidx, nb_chunks)
if is_last:
padding = (-dist_peaks, 0)
elif is_first:
padding = (0, dist_peaks)
else:
padding = (-dist_peaks, dist_peaks)
result = {'spiketimes' : [], 'amplitudes' : [], 'templates' : []}
local_chunk, t_offset = data_file.get_data(gidx, chunk_size, padding, nodes=nodes)
len_chunk = len(local_chunk)
if do_spatial_whitening:
if use_gpu:
local_chunk = cmt.CUDAMatrix(local_chunk, copy_on_host=False)
local_chunk = local_chunk.dot(spatial_whitening).asarray()
else:
local_chunk = numpy.dot(local_chunk, spatial_whitening)
if do_temporal_whitening:
local_chunk = scipy.ndimage.filters.convolve1d(local_chunk, temporal_whitening, axis=0, mode='constant')
#print "Extracting the peaks..."
local_peaktimes = numpy.zeros(0, dtype=numpy.uint32)
local_elecs = numpy.zeros(0, dtype=numpy.uint32)
local_amps = numpy.zeros(0, dtype=numpy.float32)
if matched_filter:
if sign_peaks in ['positive', 'both']:
filter_chunk = scipy.ndimage.filters.convolve1d(local_chunk, waveform_pos, axis=0, mode='constant')
for i in xrange(N_e):
peaktimes = algo.detect_peaks(filter_chunk[:, i], matched_tresholds_pos[i], mpd=dist_peaks)
local_peaktimes = numpy.concatenate((local_peaktimes, peaktimes))
local_elecs = numpy.concatenate((local_elecs, i*numpy.ones(len(peaktimes), dtype='uint32')))
local_amps = numpy.concatenate((local_amps, filter_chunk[peaktimes, i]))
if sign_peaks in ['negative', 'both']:
filter_chunk = scipy.ndimage.filters.convolve1d(local_chunk, waveform_neg, axis=0, mode='constant')
for i in xrange(N_e):
peaktimes = algo.detect_peaks(filter_chunk[:, i], matched_tresholds_neg[i], mpd=dist_peaks)
local_peaktimes = numpy.concatenate((local_peaktimes, peaktimes))
local_elecs = numpy.concatenate((local_elecs, i*numpy.ones(len(peaktimes), dtype='uint32')))
local_amps = numpy.concatenate((local_amps, filter_chunk[peaktimes, i]))
else:
for i in xrange(N_e):
if sign_peaks == 'negative':
peaktimes = algo.detect_peaks(local_chunk[:, i], thresholds[i], valley=True, mpd=dist_peaks)
elif sign_peaks == 'positive':
peaktimes = algo.detect_peaks(local_chunk[:, i], thresholds[i], valley=False, mpd=dist_peaks)
elif sign_peaks == 'both':
peaktimes = algo.detect_peaks(numpy.abs(local_chunk[:, i]), thresholds[i], valley=False, mpd=dist_peaks)
local_peaktimes = numpy.concatenate((local_peaktimes, peaktimes))
local_elecs = numpy.concatenate((local_elecs, i*numpy.ones(len(peaktimes), dtype='uint32')))
local_amps = numpy.concatenate((local_amps, local_chunk[peaktimes, i]))
g_offset = t_offset + padding[0]
if ignore_dead_times:
dead_indices = numpy.searchsorted(all_dead_times, [t_offset, t_offset + chunk_size])
if dead_indices[0] != dead_indices[1]:
local_peaktimes = numpy.array(list(set(local_peaktimes + g_offset).difference(all_dead_times[dead_indices[0]:dead_indices[1]])), dtype=numpy.uint32) - g_offset
local_peaktimes = numpy.sort(local_peaktimes)
#print "Removing the useless borders..."
local_borders = (0, len_chunk - dist_peaks)
idx = (local_peaktimes >= local_borders[0]) & (local_peaktimes < local_borders[1])
local_peaktimes = numpy.compress(idx, local_peaktimes) + g_offset
local_elecs = numpy.compress(idx, local_elecs)
local_amps = numpy.compress(idx, local_amps)
spiketimes_file.write(local_peaktimes.tostring())
electrodes_file.write(local_elecs.tostring())
amplitudes_file.write(local_amps.tostring())
sys.stderr.flush()
spiketimes_file.flush()
os.fsync(spiketimes_file.fileno())
spiketimes_file.close()
electrodes_file.flush()
os.fsync(electrodes_file.fileno())
electrodes_file.close()
amplitudes_file.flush()
os.fsync(amplitudes_file.fileno())
amplitudes_file.close()
comm.Barrier()
if comm.rank == 0:
io.collect_mua(comm.size, params, erase=True)
data_file.close()
def plot_extracted_extra_spikes(loc_all_chunks, data_len, mpi_input,
data_dtype, chunk_len, chunk_size, N_total,
nodes, extra_means, extra_stds, k, params,
safety_space, safety_time):
"""Temporary function to see if the computed thresholds for a given dataset are valid"""
count = 0
gidx = loc_all_chunks[0]
loc_chunk, loc_shape = io.load_chunk(params,
gidx,
chunk_len,
chunk_size,
nodes=nodes)
sampling_rate = params.getint('data', 'sampling_rate')
dist_peaks = params.getint('data', 'dist_peaks')
skip_artefact = params.getboolean('data', 'skip_artefact')
template_shift = params.getint('data', 'template_shift')
alignment = params.getboolean('detection', 'alignment')
nodes, _ = io.get_nodes_and_edges(params)
N_elec = nodes.size
do_temporal_whitening = params.getboolean('whitening', 'temporal')
do_spatial_whitening = params.getboolean('whitening', 'spatial')
if do_spatial_whitening:
spatial_whitening = io.load_data(params, 'spatial_whitening')
if do_temporal_whitening:
temporal_whitening = io.load_data(params, 'temporal_whitening')
loc_chunk = numpy.zeros(data_len, dtype=data_dtype)
mpi_input.Read_at(gidx * chunk_len, loc_chunk)
loc_shape = chunk_size
loc_chunk = loc_chunk.reshape(loc_shape, N_total)
# Consider only the valid channels.
loc_chunk = loc_chunk[:, nodes]
# extra_means_ = extra_means[nodes]
# extra_stds_ = extra_stds[nodes]
extra_means_ = extra_means
extra_stds_ = extra_stds
# Whiten signal.
if do_spatial_whitening:
loc_chunk = numpy.dot(loc_chunk, spatial_whitening)
if do_temporal_whitening:
loc_chunk = scipy.ndimage.filters.convolve1d(loc_chunk,
temporal_whitening,
axis=0,
mode='constant')
# Preallocation for results.
peak_times = N_elec * [None]
peak_channels = N_elec * [None]
# For each electrode.
for e in xrange(0, N_elec):
# Extract the peaks of the current chunk.
threshold = k * extra_stds_[e]
peak_times[e] = algo.detect_peaks(loc_chunk[:, e],
threshold,
valley=True,
mpd=dist_peaks)
peak_channels[e] = e * numpy.ones(peak_times[e].size, dtype='int')
if skip_artefact:
# Remove strong artifacts.
peak_values = loc_chunk[peak_times[e], e]
peak_indices = numpy.where(-10.0 * threshold <= peak_values)[0]
peak_times[e] = peak_times[e][peak_indices]
peak_times = numpy.concatenate(peak_times)
peak_channels = numpy.concatenate(peak_channels)
# Remove the useless borders.
if alignment:
loc_borders = (2 * template_shift, loc_shape - 2 * template_shift)
else:
loc_borders = (template_shift, loc_shape - template_shift)
peak_flags = (loc_borders[0] <= peak_times) & (peak_times < loc_borders[1])
peak_times = peak_times[peak_flags]
peak_channels = peak_channels[peak_flags]
# Filter unique peak times.
loc_peak_times = numpy.unique(peak_times)
n_times = len(loc_peak_times)
loc_peak_flags = numpy.zeros(n_times, dtype='bool')
loc_peak_elecs = numpy.zeros(n_times, dtype='int')
if 0 < len(loc_peak_times):
diff_times = loc_peak_times[-1] - loc_peak_times[0]
all_times = numpy.zeros((N_elec, diff_times + 1), dtype='bool')
min_times = numpy.maximum(
loc_peak_times - loc_peak_times[0] - safety_time, 0)
max_times = numpy.minimum(
loc_peak_times - loc_peak_times[0] + safety_time + 1, diff_times)
##### TODO: remove temporary zone
numpy.random.seed(42)
##### end temporary zone
argmax_peak = numpy.random.permutation(numpy.arange(n_times))
all_indices = loc_peak_times[argmax_peak]
# Select peaks with spatio-temporal masks.
for peak_index, peak_time in zip(argmax_peak, all_indices):
# Select electrode showing lowest amplitude.
elec = numpy.argmin(loc_chunk[peak_time, :])
_, neighs = get_neighbors(params, chan=elec)
if safety_space:
mslice = all_times[neighs,
min_times[peak_index]:max_times[peak_index]]
else:
mslice = all_times[elec,
min_times[peak_index]:max_times[peak_index]]
is_local_min = (elec in peak_channels[peak_times == peak_time])
if is_local_min and not mslice.any():
loc_peak_flags[peak_index] = True
loc_peak_elecs[peak_index] = elec
if safety_space:
all_times[
neighs,
min_times[peak_index]:max_times[peak_index]] = True
else:
all_times[
elec,
min_times[peak_index]:max_times[peak_index]] = True
loc_peak_times = numpy.compress(loc_peak_flags, loc_peak_times)
loc_peak_elecs = numpy.compress(loc_peak_flags, loc_peak_elecs)
time = loc_peak_times
channel = loc_peak_elecs
pos = numpy.random.rand(time.size) - 0.5
fig = plt.figure()
ax = fig.gca()
# For each first chunk plot one channel per figure.
for j in xrange(0, loc_chunk.shape[1]):
ax = fig.gca()
ax.plot(loc_chunk[:, j])
ax.plot([0, loc_chunk.shape[0] - 1], 2 * [extra_means_[j]], 'k--')
ax.plot([0, loc_chunk.shape[0] - 1],
2 * [extra_means_[j] + extra_stds_[j]], 'k--')
ax.plot([0, loc_chunk.shape[0] - 1],
2 * [extra_means_[j] - extra_stds_[j]], 'k--')
ax.plot([0, loc_chunk.shape[0] - 1],
2 * [extra_means_[j] + k * extra_stds_[j]], 'k--')
ax.plot([0, loc_chunk.shape[0] - 1],
2 * [extra_means_[j] - k * extra_stds_[j]], 'k--')
idx, = numpy.where(channel == j)
# y = + 250.0 * numpy.ones(idx.size) + 100.0 * pos[idx]
y = -9.5 * numpy.ones(idx.size) + 1.0 * pos[idx]
ax.scatter(time[idx], y, c='r')
# new_idx, = numpy.where(new_channel == j)
# new_y = - 350.0 * numpy.ones(new_idx.size) + 100.0 * new_pos[new_idx]
# ax.scatter(new_time[new_idx], new_y, c='g')
ax.set_xlim(0, loc_chunk.shape[0] - 1)
# ax.set_ylim(- 400.0, 400.0)
ax.set_ylim(-10.0, 10.0)
plt.savefig("/tmp/check-{}-{}.png".format(j, comm.rank))
fig.clear()
return
