def __call__(self, images):
self.visual_backprop_anchors.clear()
with cuda.Device(images.data.device):
input_images = self.prepare_images(images.copy() * 255)
h = self.feature_extractor(input_images)
if self.train_imagenet:
return h
if images.shape[-2] > 224:
h = self.res6(h)
if images.shape[-2] > 300:
h = self.res7(h)
self.visual_backprop_anchors.append(h)
h = _global_average_pooling_2d(h)
transform_params = self.param_predictor(h)
transform_params = rotation_dropout(F.reshape(transform_params,
(-1, 2, 3)),
ratio=0.0)
points = F.spatial_transformer_grid(transform_params, self.out_size)
rois = F.spatial_transformer_sampler(images, points)
if self.transform_rois_to_grayscale:
assert rois.shape[
1] == 3, "rois are not in RGB, can not convert them to grayscale"
b, g, r = F.split_axis(rois, 3, axis=1)
rois = 0.299 * r + 0.587 * g + 0.114 * b
return rois, points
def __call__(self, x):
h = relu(self.conv1(x))
h = relu(self.bn2(self.conv2(h)))
h = relu(self.bn3(self.conv3(h)))
h = relu(self.bn4(self.conv4(h)))
h = R._global_average_pooling_2d(h)
h = self.fc(h)
return h
def get_transform_params(self, features):
h = _global_average_pooling_2d(features)
lstm_predictions = [
self.lstm(h) for _ in range(self.num_bboxes_to_localize)
]
lstm_predictions = F.stack(lstm_predictions, axis=1)
batch_size, num_boxes, _ = lstm_predictions.shape
lstm_predictions = F.reshape(lstm_predictions,
(-1, ) + lstm_predictions.shape[2:])
params = self.param_predictor(lstm_predictions)
transform_params = rotation_dropout(F.reshape(params, (-1, 2, 3)),
ratio=self.dropout_ratio)
return transform_params
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 __call__(self, x):
# self.visual_backprop_anchors.clear()
h = self.feature_extractor(x)
# self.visual_backprop_anchors.append(h)
h = _global_average_pooling_2d(h)
return h