def activation_mask(self,
layer,
unit,
index,
shape=None,
use_fieldmap=True):
if shape is None:
record, shape = self.instance_data(index)
sw, sh = shape
# reduction = int(round(self.iw / float(sw)))
lp = self.layer[layer]
blobdata = lp.blobdata
fieldmap = lp.fieldmap
quantdata = lp.quantdata
threshold = quantdata[unit, int(round(quantdata.shape[1] * lp.level))]
if use_fieldmap:
up = upsample.upsampleL(fieldmap,
blobdata[index:index + 1, unit],
shape=(self.ih, self.iw),
scaleshape=(sh, sw))[0]
else:
upsample_f = nn.Upsample(size=(self.ih, self.iw), mode='bilinear')
up = numpy.squeeze(
upsample_f(
Variable(
torch.Tensor(
blobdata[index:index + 1,
unit]).unsqueeze(0))).data.cpu().numpy())
mask = up > threshold
return mask
def activation_mask(lp, unit, index, shape):
blobdata = lp.blobdata
fieldmap = lp.fieldmap
quantdata = lp.quantdata
threshold = quantdata[unit, int(round(quantdata.shape[1] * lp.level))]
up = upsample.upsampleL(fieldmap,
blobdata[index:index + 1, unit],
shape=lp.input_dim,
scaleshape=shape)[0]
mask = up > threshold
return mask
def activation_mask(self, layer, unit, index, shape=None):
if shape is None:
record, shape = self.instance_data(index)
sw, sh = shape
# reduction = int(round(self.iw / float(sw)))
lp = self.layer[layer]
blobdata = lp.blobdata
fieldmap = lp.fieldmap
quantdata = lp.quantdata
threshold = quantdata[unit, int(round(quantdata.shape[1] * lp.level))]
up = upsample.upsampleL(fieldmap,
blobdata[index:index + 1, unit],
shape=(self.ih, self.iw),
scaleshape=(sh, sw))[0]
mask = up > threshold
return mask
def process_data(fn_t, fn_read, shape, tally_depth, ds, iw, ih, categories,
fieldmap, thresh, labelcat, batch_size, ahead, verbose,
thread, start, end):
unit_size = len(thresh)
blobdata = cached_memmap(fn_read, mode='r', dtype='float32', shape=shape)
count_t = cached_memmap(fn_t,
mode='r+',
dtype='int32',
shape=(ds.size(), tally_depth, 3))
count_t[...] = 0
# The main loop
if verbose:
print 'Beginning work for evaluating', blob
pf = loadseg.SegmentationPrefetcher(ds,
categories=categories,
start=start,
end=end,
once=True,
batch_size=batch_size,
ahead=ahead,
thread=False)
index = start
start_time = time.time()
last_batch_time = start_time
batch_size = 0
for batch in pf.batches():
batch_time = time.time()
rate = (index - start) / (batch_time - start_time + 1e-15)
batch_rate = batch_size / (batch_time - last_batch_time + 1e-15)
last_batch_time = batch_time
if verbose:
print 'labelprobe index', index, 'items per sec', batch_rate, rate
sys.stdout.flush()
for rec in batch:
sw, sh = [rec[k] for k in ['sw', 'sh']]
reduction = int(round(iw / float(sw)))
up = upsample.upsampleL(fieldmap,
blobdata[index],
shape=(sh, sw),
reduction=reduction)
mask = up > thresh
accumulate_counts(mask, [rec[cat] for cat in categories],
count_t[index], unit_size, labelcat)
index += 1
batch_size = len(batch)
count_t.flush()