def forward(self, imgs, gt_mb_locs, gt_mb_labels):
mb_locs, mb_confs = self.model(imgs)
loc_loss, conf_loss = multibox_loss(mb_locs, mb_confs, gt_mb_locs, gt_mb_labels, self.k)
loss = loc_loss * self.alpha + conf_loss
chainer.reporter.report({'loss': loss, 'loss/loc': loc_loss, 'loss/conf': conf_loss}, self)
return loss
def _check_forward(self, mb_locs_local, mb_confs_local, gt_mb_locs_local,
gt_mb_labels_local, k):
loc_loss_local, conf_loss_local = multibox_loss(
mb_locs_local, mb_confs_local, gt_mb_locs_local,
gt_mb_labels_local, k, self.comm)
loc_loss_local = cuda.to_cpu(loc_loss_local.array)
conf_loss_local = cuda.to_cpu(conf_loss_local.array)
loc_loss = self.comm.allreduce_obj(loc_loss_local) / self.comm.size
conf_loss = self.comm.allreduce_obj(conf_loss_local) / self.comm.size
expect_loc_loss, expect_conf_loss = multibox_loss(
self.mb_locs, self.mb_confs, self.gt_mb_locs, self.gt_mb_labels, k)
np.testing.assert_almost_equal(loc_loss,
expect_loc_loss.array,
decimal=2)
np.testing.assert_almost_equal(conf_loss,
expect_conf_loss.array,
decimal=2)
def __call__(self, imgs, gt_mb_locs, gt_mb_labs):
mb_locs, mb_confs = self.model(imgs)
loc_loss, conf_loss = multibox_loss(
mb_locs, mb_confs, gt_mb_locs, gt_mb_labs, self.k)
loss = loc_loss * self.alpha + conf_loss
chainer.reporter.report(
{'loss': loss, 'loss/loc': loc_loss, 'loss/conf': conf_loss},
self)
return loss
def _check_forward(self, mb_locs_local, mb_confs_local, gt_mb_locs_local,
gt_mb_labels_local, k):
loc_loss_local, conf_loss_local = multibox_loss(
mb_locs_local, mb_confs_local, gt_mb_locs_local,
gt_mb_labels_local, k, self.comm)
loc_loss_local = cuda.to_cpu(loc_loss_local.array)
conf_loss_local = cuda.to_cpu(conf_loss_local.array)
from mpi4py import MPI
self.comm.mpi_comm.Allreduce(MPI.IN_PLACE, loc_loss_local)
self.comm.mpi_comm.Allreduce(MPI.IN_PLACE, conf_loss_local)
loc_loss, conf_loss = multibox_loss(self.mb_locs, self.mb_confs,
self.gt_mb_locs, self.gt_mb_labels,
k)
np.testing.assert_almost_equal(loc_loss_local,
loc_loss.array,
decimal=2)
np.testing.assert_almost_equal(conf_loss_local,
conf_loss.array,
decimal=2)
def __call__(self, imgs, gt_mb_locs, gt_mb_labels):
mb_locs, mb_confs = self.model(imgs)
loc_loss, conf_loss = multibox_loss(mb_locs, mb_confs, gt_mb_locs,
gt_mb_labels, self.k)
loss = loc_loss * self.alpha + conf_loss
chainer.reporter.report(
{
self.loss_labels[0]: loss,
self.loss_labels[1]: loc_loss,
self.loss_labels[2]: conf_loss
}, self)
return loss
def __call__(self, params_detection_loss, params_segmentation_loss):
mb_locs, mb_confs, gt_mb_locs, gt_mb_labels, k = params_detection_loss
score, mask_gt = params_segmentation_loss
loc_loss, conf_loss = multibox_loss(mb_locs, mb_confs, gt_mb_locs,
gt_mb_labels, k)
detection_loss = loc_loss * self.alpha + conf_loss
segmentation_loss = mask_loss(score, mask_gt)
loss_pre = detection_loss + segmentation_loss
if self.param1.array is None:
self.param1.initialize([1])
if self.param2.array is None:
self.param2.initialize([1])
param1_precision = chainer.functions.exp(-self.param1)
param2_precision = chainer.functions.exp(-self.param2)
loss = detection_loss * param1_precision + 2 * self.param1 \
+ segmentation_loss * param2_precision + self.param2
return loss, (loss, loss_pre, loc_loss, conf_loss, segmentation_loss,
self.param1, self.param2)
def __call__(self, imgs, gt_mb_locs, gt_mb_labels, gt_mask):
detection_result, segmentation_result = self.model(imgs)
# mb_locs, mb_confs = self.model(imgs)
detection = False
segmentation = False
if detection_result is not None:
detection = True
if segmentation_result is not None:
segmentation = True
if segmentation and detection:
self.multi_task_loss = self.multi_task_loss
else:
self.multi_task_loss = False
if detection:
mb_locs, mb_confs = detection_result
if segmentation:
score = segmentation_result
if not self.multi_task_loss:
if detection:
loc_loss, conf_loss = multibox_loss(mb_locs, mb_confs,
gt_mb_locs, gt_mb_labels,
self.k)
loss_detection = loc_loss * self.alpha + conf_loss
if segmentation:
loss_segmentation = F.softmax_cross_entropy(score, gt_mask)
if segmentation and detection:
if self.use_dynamic_loss:
pass
else:
loss = 2 * self.loss_split * loss_detection + 2 * (
1 - self.loss_split) * loss_segmentation
chainer.reporter.report(
{
'loss': loss,
'loss/loc': loc_loss,
'loss/conf': conf_loss,
'loss/mask': loss_segmentation,
'loss/split': self.loss_split
}, self)
elif segmentation:
loss = loss_segmentation
chainer.reporter.report(
{
'loss': loss,
'loss/mask': loss_segmentation
}, self)
elif detection:
loss = loss_detection
chainer.reporter.report(
{
'loss': loss,
'loss/loc': loc_loss,
'loss/conf': conf_loss
}, self)
else:
loss, report_values = self.multi_loss(
(mb_locs, mb_confs, gt_mb_locs, gt_mb_labels, self.k),
(score, gt_mask))
# loss, (loss_pre, loc_loss, conf_loss, segmentation_loss)
chainer.reporter.report(
{
'loss': report_values[0].data[0],
'loss/pre': report_values[1],
'loss/loc': report_values[2],
'loss/conf': report_values[3],
'loss/mask': report_values[4],
'param_det': report_values[5].data[0],
'param_seg': report_values[6].data[0]
}, self)
return loss
def _check_forward(self, mb_locs, mb_confs, gt_mb_locs, gt_mb_labels, k):
if self.variable:
mb_locs = chainer.Variable(mb_locs)
mb_confs = chainer.Variable(mb_confs)
gt_mb_locs = chainer.Variable(gt_mb_locs)
gt_mb_labels = chainer.Variable(gt_mb_labels)
loc_loss, conf_loss = multibox_loss(mb_locs, mb_confs, gt_mb_locs,
gt_mb_labels, k)
self.assertIsInstance(loc_loss, chainer.Variable)
self.assertEqual(loc_loss.shape, ())
self.assertEqual(loc_loss.dtype, mb_locs.dtype)
self.assertIsInstance(conf_loss, chainer.Variable)
self.assertEqual(conf_loss.shape, ())
self.assertEqual(conf_loss.dtype, mb_confs.dtype)
if self.variable:
mb_locs = mb_locs.array
mb_confs = mb_confs.array
gt_mb_locs = gt_mb_locs.array
gt_mb_labels = gt_mb_labels.array
mb_locs = cuda.to_cpu(mb_locs)
mb_confs = cuda.to_cpu(mb_confs)
gt_mb_locs = cuda.to_cpu(gt_mb_locs)
gt_mb_labels = cuda.to_cpu(gt_mb_labels)
loc_loss = cuda.to_cpu(loc_loss.array)
conf_loss = cuda.to_cpu(conf_loss.array)
n_positive_total = 0
expect_loc_loss = 0
expect_conf_loss = 0
for i in six.moves.xrange(gt_mb_labels.shape[0]):
n_positive = 0
negatives = []
for j in six.moves.xrange(gt_mb_labels.shape[1]):
loc = F.huber_loss(mb_locs[np.newaxis, i, j],
gt_mb_locs[np.newaxis, i, j], 1).array
conf = F.softmax_cross_entropy(mb_confs[np.newaxis, i, j],
gt_mb_labels[np.newaxis, i,
j]).array
if gt_mb_labels[i, j] > 0:
n_positive += 1
expect_loc_loss += loc
expect_conf_loss += conf
else:
negatives.append(conf)
n_positive_total += n_positive
if n_positive > 0:
expect_conf_loss += sum(sorted(negatives)[-n_positive * k:])
if n_positive_total == 0:
expect_loc_loss = 0
expect_conf_loss = 0
else:
expect_loc_loss /= n_positive_total
expect_conf_loss /= n_positive_total
np.testing.assert_almost_equal(loc_loss, expect_loc_loss, decimal=2)
np.testing.assert_almost_equal(conf_loss, expect_conf_loss, decimal=2)
def _check_forward(self, mb_locs, mb_confs, gt_mb_locs, gt_mb_labels, k):
if self.variable:
mb_locs = chainer.Variable(mb_locs)
mb_confs = chainer.Variable(mb_confs)
gt_mb_locs = chainer.Variable(gt_mb_locs)
gt_mb_labels = chainer.Variable(gt_mb_labels)
loc_loss, conf_loss = multibox_loss(
mb_locs, mb_confs, gt_mb_locs, gt_mb_labels, k)
self.assertIsInstance(loc_loss, chainer.Variable)
self.assertEqual(loc_loss.shape, ())
self.assertEqual(loc_loss.dtype, mb_locs.dtype)
self.assertIsInstance(conf_loss, chainer.Variable)
self.assertEqual(conf_loss.shape, ())
self.assertEqual(conf_loss.dtype, mb_confs.dtype)
if self.variable:
mb_locs = mb_locs.array
mb_confs = mb_confs.array
gt_mb_locs = gt_mb_locs.array
gt_mb_labels = gt_mb_labels.array
mb_locs = cuda.to_cpu(mb_locs)
mb_confs = cuda.to_cpu(mb_confs)
gt_mb_locs = cuda.to_cpu(gt_mb_locs)
gt_mb_labels = cuda.to_cpu(gt_mb_labels)
loc_loss = cuda.to_cpu(loc_loss.array)
conf_loss = cuda.to_cpu(conf_loss.array)
n_positive_total = 0
expect_loc_loss = 0
expect_conf_loss = 0
for i in six.moves.xrange(gt_mb_labels.shape[0]):
n_positive = 0
negatives = []
for j in six.moves.xrange(gt_mb_labels.shape[1]):
loc = F.huber_loss(
mb_locs[np.newaxis, i, j],
gt_mb_locs[np.newaxis, i, j], 1).array
conf = F.softmax_cross_entropy(
mb_confs[np.newaxis, i, j],
gt_mb_labels[np.newaxis, i, j]).array
if gt_mb_labels[i, j] > 0:
n_positive += 1
expect_loc_loss += loc
expect_conf_loss += conf
else:
negatives.append(conf)
n_positive_total += n_positive
if n_positive > 0:
expect_conf_loss += sum(sorted(negatives)[-n_positive * k:])
if n_positive_total == 0:
expect_loc_loss = 0
expect_conf_loss = 0
else:
expect_loc_loss /= n_positive_total
expect_conf_loss /= n_positive_total
np.testing.assert_almost_equal(
loc_loss, expect_loc_loss, decimal=2)
np.testing.assert_almost_equal(
conf_loss, expect_conf_loss, decimal=2)
