上一节我们已经谈到了计算节点,但是即使是官方文档介绍里面相关内容也过于简略,我们使用Faster-RCNN代码中的新建节点为例,重新介绍一下新建节点的调用栈。
1、调用新建节点
参数分为三部分,op_type是节点名称,对应于辅助class的装饰器的输入;其他参数一部分传递给辅助class的初始化函数(这部分参数的虚参名和初始化函数的需参名要对应上),一部分直接作为一个list传给节点定义class的forward函数的in_data参数。
group = mx.symbol.Custom(rois=rois, # 2000*5的roi信息,CustomOpProp无此参数 gt_boxes=gt_boxes, # n*5的ground truth信息,n表示object数量,CustomOpProp无此参数 op_type='proposal_target', # <-----对应辅助节点装饰器参数 # CustomOp,CustomOpProp初始化参数 num_classes=num_classes, # num_classes是实际要分类的类别数加上背景类 batch_images=rcnn_batch_size, # 1 batch_rois=rcnn_batch_rois, # 128 fg_fraction=rcnn_fg_fraction, #s 0.25,正样本所占的比例 fg_overlap=rcnn_fg_overlap, # 0.5 box_stds=rcnn_bbox_stds # (0.1, 0.1, 0.2, 0.2) )
2、建立辅助class:CustomOpProp
本部分的class方法参数固定,不可以随意修改
@mx.operator.register('proposal_target') # <-----对应调用的op_type class ProposalTargetProp(mx.operator.CustomOpProp): def __init__(self, num_classes='21', batch_images='1', batch_rois='128', fg_fraction='0.25', # 接受上面调用的后5个参数 fg_overlap='0.5', box_stds='(0.1, 0.1, 0.2, 0.2)'): super(ProposalTargetProp, self).__init__(need_top_grad=False) # <-----本节点是否需要后面的梯度 self._num_classes = int(num_classes) # num_classes是实际要分类的类别数加上背景类 self._batch_images = int(batch_images) # 1 self._batch_rois = int(batch_rois) # 128 self._fg_fraction = float(fg_fraction) # 0.25,正样本所占的比例 self._fg_overlap = float(fg_overlap) # 0.5 self._box_stds = tuple(np.fromstring(box_stds[1:-1], dtype=float, sep=',')) # (0.1, 0.1, 0.2, 0.2) def list_arguments(self): return ['rois', 'gt_boxes'] # 向前传播需要的参数 def list_outputs(self): return ['rois_output', 'label', 'bbox_target', 'bbox_weight'] # 向前传播输出参数名 def infer_shape(self, in_shape): assert self._batch_rois % self._batch_images == 0, 'BATCHIMAGES {} must devide BATCH_ROIS {}'.format(self._batch_images, self._batch_rois) rpn_rois_shape = in_shape[0] gt_boxes_shape = in_shape[1] output_rois_shape = (self._batch_rois, 5) label_shape = (self._batch_rois, ) bbox_target_shape = (self._batch_rois, self._num_classes * 4) bbox_weight_shape = (self._batch_rois, self._num_classes * 4) return [rpn_rois_shape, gt_boxes_shape], [output_rois_shape, label_shape, bbox_target_shape, bbox_weight_shape] def create_operator(self, ctx, shapes, dtypes): # 返回初始化了的自定义节点类 return ProposalTargetOperator(self._num_classes, self._batch_images, self._batch_rois, self._fg_fraction, self._fg_overlap, self._box_stds) def declare_backward_dependency(self, out_grad, in_data, out_data): return []
__init__
初始化方法会首先从调用位置接收参数,调用位置的op_type参数用于指定选用哪个辅助class,然后其他参数优先传入本初的__init__方法,剩下的没有和本方法参数对应上的参数会做为节点class的forward方法的in_data参数。
下面一行表示本节点不需要接收后面层的梯度,对应到节点定义class的backward方法,out_grad参数就不应在函数体内调用了,in_grad(向前传播回去的梯度)计算完全依赖本层的参数。
super(ProposalTargetProp, self).__init__(need_top_grad=False)
list_arguments
使用list_arguments()方法返回时,输出并不直接就是上述代码list_arguments方法的return列表,实际上会将整个net结构截至的本节点为止,全部的variable变量名称输入。
对于本节点group,定义输入有两个:
rois=rois # 2000*5的roi信息,CustomOpProp无此参数
gt_boxes=gt_boxes
其中gt_boxes变量本身是个variable(定义为gt_boxes = mx.symbol.Variable(name="gt_boxes")),但是rois为symbol,是之前网络的输出symbol,所以实际输出的arguments为gt_boxes本身,以及rois所依赖的全部variables,包含认为定义的占位符variable和网络层自带的参数variable。
['data', 'conv1_1_weight', 'conv1_1_bias', 'conv1_2_weight', 'conv1_2_bias', 'conv2_1_weight', 'conv2_1_bias', 'conv2_2_weight', 'conv2_2_bias', 'conv3_1_weight', 'conv3_1_bias', 'conv3_2_weight', 'conv3_2_bias', 'conv3_3_weight', 'conv3_3_bias', 'conv4_1_weight', 'conv4_1_bias', 'conv4_2_weight', 'conv4_2_bias', 'conv4_3_weight', 'conv4_3_bias', 'conv5_1_weight', 'conv5_1_bias', 'conv5_2_weight', 'conv5_2_bias', 'conv5_3_weight', 'conv5_3_bias', 'rpn_conv_3x3_weight', 'rpn_conv_3x3_bias', 'rpn_cls_score_weight', 'rpn_cls_score_bias', 'rpn_bbox_pred_weight', 'rpn_bbox_pred_bias', 'im_info',
'gt_boxes']
infer_shape
其输入参数in_shape就是list_arguments中return的那几个变量的shape,对于本例,就是rois和gtboxes的shape,本方法用于推断输出symbol和梯度symbol的shape是否正确。
3、实现节点class
节点class的初始化和调用部分的参数完全无关,是由辅助节点来进行传参调用的。但是其forward方法的in_data参数其值接收是从调用初进行的,in_data中的参数就是上面list_arguments方法的return结果(['rois', 'gt_boxes']),实际传参可以有空缺(例如第一小节可以删掉gt_boxes),缺省参数视为定义了一个Variable占位。
class ProposalTargetOperator(mx.operator.CustomOp): def __init__(self, num_classes, batch_images, batch_rois, fg_fraction, fg_overlap, box_stds): super(ProposalTargetOperator, self).__init__() self._num_classes = num_classes # num_classes是实际要分类的类别数加上背景类 self._batch_images = batch_images # 1 self._batch_rois = batch_rois # 128 self._rois_per_image = int(batch_rois / batch_images) self._fg_rois_per_image = int(round(fg_fraction * self._rois_per_image)) self._fg_overlap = fg_overlap # 0.5 self._box_stds = box_stds # (0.1, 0.1, 0.2, 0.2) def forward(self, is_train, req, in_data, out_data, aux): """Forward interface. Can override when creating new operators. Parameters ---------- is_train : bool whether this is for training req : list of str how to assign to out_data. can be 'null', 'write', or 'add'. You can optionally use self.assign(dst, req, src) to handle this. in_data, out_data, aux: list of NDArrays input, output, and auxiliary states for forward. See document for corresponding arguments of Operator::Forward """ assert self._batch_images == in_data[1].shape[0], 'check batch size of gt_boxes' all_rois = in_data[0].asnumpy() # [2000, 5] all_gt_boxes = in_data[1].asnumpy() # [n, 5] rois = np.empty((0, 5), dtype=np.float32) labels = np.empty((0, ), dtype=np.float32) bbox_targets = np.empty((0, 4 * self._num_classes), dtype=np.float32) bbox_weights = np.empty((0, 4 * self._num_classes), dtype=np.float32) for batch_idx in range(self._batch_images): b_rois = all_rois[np.where(all_rois[:, 0] == batch_idx)[0]] b_gt_boxes = all_gt_boxes[batch_idx] b_gt_boxes = b_gt_boxes[np.where(b_gt_boxes[:, -1] > 0)[0]] # Include ground-truth boxes in the set of candidate rois batch_pad = batch_idx * np.ones((b_gt_boxes.shape[0], 1), dtype=b_gt_boxes.dtype) b_rois = np.vstack((b_rois, np.hstack((batch_pad, b_gt_boxes[:, :-1])))) b_rois, b_labels, b_bbox_targets, b_bbox_weights = sample_rois(b_rois, b_gt_boxes, num_classes=self._num_classes, rois_per_image=self._rois_per_image, fg_rois_per_image=self._fg_rois_per_image, fg_overlap=self._fg_overlap, box_stds=self._box_stds) rois = np.vstack((rois, b_rois)) labels = np.hstack((labels, b_labels)) bbox_targets = np.vstack((bbox_targets, b_bbox_targets)) bbox_weights = np.vstack((bbox_weights, b_bbox_weights)) self.assign(out_data[0], req[0], rois) self.assign(out_data[1], req[1], labels) self.assign(out_data[2], req[2], bbox_targets) self.assign(out_data[3], req[3], bbox_weights) def backward(self, req, out_grad, in_data, out_data, in_grad, aux): """Backward interface. Can override when creating new operators. Parameters ---------- req : list of str how to assign to in_grad. can be 'null', 'write', or 'add'. You can optionally use self.assign(dst, req, src) to handle this. out_grad, in_data, out_data, in_grad, aux : list of NDArrays input and output for backward. See document for corresponding arguments of Operator::Backward """ self.assign(in_grad[0], req[0], 0) self.assign(in_grad[1], req[1], 0)
介绍完辅助节点,本部分的介绍就不太多了,注意的就是向前向后两个方法没有返回值,使用assign来给symbol赋值,数量顺序要和辅助class的argument、output对应上,具体实现因为没有研究C++源码,没办法更详细介绍了,不过不影响使用(大概)。
另外,辅助节点class是会在python解释器里直接执行的,也就是说我们添加在函数体中的print什么的能够得到输出,但是在本class中,添加的中间输出不会被print出来,应该是建立符号图时被略去了(有关C++优化计算图的机理,李沐博士有介绍,不过我的C++功底不够,没有看过源码,仍旧觉得符号式编程的运行过程很神奇……),另外,我们在bind等操作做检查时,仅仅会运行辅助节点,不到真实的数据流入,这个class是不会运行乃至报错的,所以辅助节点的设计真的很重要。