def __call__(self, x, rois, roi_indices, spatial_scale):
roi_indices = roi_indices.astype(np.float32)
indices_and_rois = self.xp.concatenate((roi_indices[:, None], rois),
axis=1)
# roi poolingをする前に、thin feature mapに変換します
# activationしないっぽいことが書いてあるんだよなー
left_path = self.conv_bl(self.conv_ul(x))
right_path = self.conv_br(self.conv_ur(x))
tfp = left_path + right_path
pool = _roi_align_2d_yx(tfp, indices_and_rois, self.roi_size,
self.roi_size, spatial_scale)
h = F.relu(self.fc(pool))
roi_cls_locs = self.cls_loc(h)
roi_scores = self.score(h)
# at prediction time, we use two pass method.
# at first path, we predict box location and class
# at second path, we predict mask with accurate location from first path
if chainer.config.train:
mask = F.relu(self.conv2(pool))
mask = F.relu(self.conv3_(mask))
mask = F.relu(self.conv4(mask))
mask = self.deconv1_(pool)
#mask = self.conv2(self.deconv1(pool))
return roi_cls_locs, roi_scores, mask
else:
# cache tfp for second path
self.tfp = tfp
return roi_cls_locs, roi_scores
def __call__(self, x, indices_and_rois, levels, spatial_scales):
pool_box = list()
levels = chainer.cuda.to_cpu(levels).astype(np.int32)
for l, i in zip(levels, indices_and_rois):
pool_box.append(
_roi_align_2d_yx(x[l], i[None], self.roi_size_box,
self.roi_size_box, spatial_scales[l]))
pool_box = F.concat(pool_box, axis=0)
h = F.relu(self.conv1(pool_box))
h = F.relu(self.fc1(h))
h = F.relu(self.fc2(h))
roi_cls_locs = self.cls_loc(h)
roi_scores = self.score(h)
# at prediction time, we use two pass method.
# at first path, we predict box location and class
# at second path, we predict mask with accurate location from first path
if chainer.config.train:
pool_mask = list()
for l, i in zip(levels, indices_and_rois):
pool_mask.append(
_roi_align_2d_yx(x[l], i[None], self.roi_size_mask,
self.roi_size_mask, spatial_scales[l]))
mask = F.concat(pool_mask, axis=0)
for l in self.mask_convs.children():
mask = F.relu(l(mask))
#mask = F.relu(self.mask1(pool_mask))
#mask = F.relu(self.mask2(mask))
#mask = F.relu(self.mask3(mask))
#mask = F.relu(self.mask4(mask))
mask = self.conv2(self.deconv1(mask))
*_, h, w = mask.shape
mask = F.resize_images(mask, output_shape=(2 * h, 2 * w))
return roi_cls_locs, roi_scores, mask
else:
# cache
self.x = x
return roi_cls_locs, roi_scores
def predict_mask(self, levels, indices_and_rois, spatial_scales):
pool_mask = list()
for l, i in zip(levels, indices_and_rois):
pool_mask.append(
_roi_align_2d_yx(self.x[l], i[None], self.roi_size_mask,
self.roi_size_mask, spatial_scales[l]))
pool_mask = F.concat(pool_mask, axis=0)
mask = F.relu(self.mask1(pool_mask))
mask = F.relu(self.mask2(mask))
mask = F.relu(self.mask3(mask))
mask = F.relu(self.mask4(mask))
mask = self.conv2(self.deconv1(mask))
return mask
def predict_mask(self, rois, roi_indices, spatial_scale):
roi_indices = roi_indices.astype(np.float32)
indices_and_rois = self.xp.concatenate((roi_indices[:, None], rois),
axis=1)
pool = _roi_align_2d_yx(self.tfp, indices_and_rois, self.roi_size,
self.roi_size, spatial_scale)
mask = F.relu(self.conv2(pool))
mask = F.relu(self.conv3_(mask))
mask = F.relu(self.conv4(mask))
mask = self.deconv1_(pool)
# mask = self.deconv1(pool)
# mask = self.conv2_(mask)
# mask = self.conv3(mask)
# mask = self.conv2(self.deconv1(pool))
return mask
def __call__(self, x, rois, roi_indices, spatial_scale):
roi_indices = roi_indices.astype(np.float32)
indices_and_rois = self.xp.concatenate(
(roi_indices[:, None], rois), axis=1)
pool = _roi_align_2d_yx(x, indices_and_rois, self.roi_size,
self.roi_size, spatial_scale)
# h: 分岐する直前まで
h = F.relu(self.res5(pool))
h = F.relu(self.conv1(h))
# global average pooling
gap = _global_average_pooling_2d(h)
roi_cls_locs = self.cls_loc(gap)
roi_scores = self.score(gap)
# mask
mask = self.conv2(F.relu(self.deconv1(h)))
return roi_cls_locs, roi_scores, mask
def predict_mask(self, levels, indices_and_rois, spatial_scales):
pool_mask = list()
for l, i in zip(levels, indices_and_rois):
pool_mask.append(
_roi_align_2d_yx(self.x[l], i[None], self.roi_size_mask,
self.roi_size_mask, spatial_scales[l]))
mask = F.concat(pool_mask, axis=0)
for l in self.mask_convs:
mask = F.relu(l(mask))
#mask = F.relu(self.mask1(pool_mask))
#mask = F.relu(self.mask2(mask))
#mask = F.relu(self.mask3(mask))
#mask = F.relu(self.mask4(mask))
mask = self.conv2(self.deconv1(mask))
*_, h, w = mask.shape
mask = F.resize_images(mask, output_shape=(2 * h, 2 * w))
return mask