def _pool(
self, h_cls_seg, h_ag_loc, rois, roi_indices):
# PSROI Pooling
# shape: (n_roi, n_class, 2, roi_size, roi_size)
roi_cls_ag_seg_scores = psroi_pooling_2d(
h_cls_seg, rois, roi_indices,
self.n_class * 2, self.roi_size, self.roi_size,
self.spatial_scale, self.group_size)
roi_cls_ag_seg_scores = roi_cls_ag_seg_scores.reshape(
(-1, self.n_class, 2, self.roi_size, self.roi_size))
# shape: (n_roi, 2*4, roi_size, roi_size)
roi_ag_loc_scores = psroi_pooling_2d(
h_ag_loc, rois, roi_indices,
2 * 4, self.roi_size, self.roi_size,
self.spatial_scale, self.group_size)
# shape: (n_roi, n_class)
roi_cls_scores = _global_average_pooling_2d(
F.max(roi_cls_ag_seg_scores, axis=2))
# Bbox Regression
# shape: (n_roi, 2*4)
roi_ag_locs = _global_average_pooling_2d(roi_ag_loc_scores)
roi_ag_locs = roi_ag_locs.reshape((-1, 2, 4))
# Mask Regression
# shape: (n_roi, n_class, 2, roi_size, roi_size)
max_cls_indices = roi_cls_scores.array.argmax(axis=1)
# shape: (n_roi, 2, roi_size, roi_size)
roi_ag_seg_scores = roi_cls_ag_seg_scores[
self.xp.arange(len(max_cls_indices)), max_cls_indices]
return roi_ag_seg_scores, roi_ag_locs, roi_cls_scores
def apply_backward(self, x_data, roi_data, roi_index_data, y_grad_data):
x = chainer.Variable(x_data)
rois = chainer.Variable(roi_data)
roi_indices = chainer.Variable(roi_index_data)
y = functions.psroi_pooling_2d(x, rois, roi_indices, self.out_c,
self.out_h, self.out_w,
self.spatial_scale, self.group_size)
x.cleargrad()
y.grad = y_grad_data
y.backward()
return x, y
def check_forward(self, x_data, roi_data, roi_index_data):
x = chainer.Variable(x_data)
rois = chainer.Variable(roi_data)
roi_indices = chainer.Variable(roi_index_data)
y = functions.psroi_pooling_2d(x, rois, roi_indices, self.out_c,
self.out_h, self.out_w,
self.spatial_scale, self.group_size)
self.assertEqual(y.data.dtype, np.float32)
y_data = cuda.to_cpu(y.data)
self.assertEqual((self.n_roi, self.out_c, self.out_h, self.out_w),
y_data.shape)
def _pool(self, h_cls_seg, h_ag_loc, rois, roi_indices, gt_roi_labels):
# PSROI Pooling
# shape: (n_roi, n_class, 2, roi_size, roi_size)
roi_cls_ag_seg_scores = psroi_pooling_2d(h_cls_seg, rois, roi_indices,
self.n_class * 2,
self.roi_size, self.roi_size,
self.spatial_scale,
self.group_size)
roi_cls_ag_seg_scores = F.reshape(
roi_cls_ag_seg_scores,
(-1, self.n_class, 2, self.roi_size, self.roi_size))
# shape: (n_roi, 2*4, roi_size, roi_size)
roi_ag_loc_scores = psroi_pooling_2d(h_ag_loc, rois, roi_indices,
2 * 4, self.roi_size,
self.roi_size, self.spatial_scale,
self.group_size)
# shape: (n_roi, n_class)
roi_cls_scores = F.average(F.max(roi_cls_ag_seg_scores, axis=2),
axis=(2, 3))
# Bbox Regression
# shape: (n_roi, 2, 4)
roi_ag_locs = F.average(roi_ag_loc_scores, axis=(2, 3))
roi_ag_locs = F.reshape(roi_ag_locs, (-1, 2, 4))
# Mask Regression
# shape: (n_roi, n_class, 2, roi_size, roi_size)
if gt_roi_labels is None:
max_cls_indices = roi_cls_scores.array.argmax(axis=1)
else:
max_cls_indices = gt_roi_labels
# shape: (n_roi, 2, roi_size, roi_size)
roi_ag_seg_scores = roi_cls_ag_seg_scores[
self.xp.arange(len(max_cls_indices)), max_cls_indices]
return roi_ag_seg_scores, roi_ag_locs, roi_cls_scores