【detectron】FPN网络输入

在detectron训练网络的过程中，给网络送的blob在下面的函数中生成：（位于minibatch.py）

def get_minibatch(roidb): """Given a roidb, construct a minibatch sampled from it.""" # We collect blobs from each image onto a list and then concat them into a # single tensor, hence we initialize each blob to an empty list blobs = {k: [] for k in get_minibatch_blob_names()} # Get the input image blob, formatted for caffe2 im_blob, im_scales = _get_image_blob(roidb) #对输入的图像处理程网络需要的形式（batch，channel，height，width），im_scales是变换的尺度 blobs['data'] = im_blob if cfg.RPN.RPN_ON: # RPN-only or end-to-end Faster/Mask R-CNN valid = rpn_roi_data.add_rpn_blobs(blobs, im_scales, roidb) elif cfg.RETINANET.RETINANET_ON: im_width, im_height = im_blob.shape[3], im_blob.shape[2] # im_width, im_height corresponds to the network input: padded image # (if needed) width and height. We pass it as input and slice the data # accordingly so that we don't need to use SampleAsOp valid = retinanet_roi_data.add_retinanet_blobs( blobs, im_scales, roidb, im_width, im_height ) else: # Fast R-CNN like models trained on precomputed proposals valid = fast_rcnn_roi_data.add_fast_rcnn_blobs(blobs, im_scales, roidb) return blobs, valid

其中给FPN网络输送blob的函数为valid = rpn_roi_data.add_rpn_blobs(blobs, im_scales, roidb)，具体来分析这个函数。

1.该函数首先完成的工作是对送进来的图片(roidb)生成anchor

def add_rpn_blobs(blobs, im_scales, roidb): """Add blobs needed training RPN-only and end-to-end Faster R-CNN models.""" if cfg.FPN.FPN_ON and cfg.FPN.MULTILEVEL_RPN: # RPN applied to many feature levels, as in the FPN paper k_max = cfg.FPN.RPN_MAX_LEVEL k_min = cfg.FPN.RPN_MIN_LEVEL foas = [] for lvl in range(k_min, k_max + 1): #对于每一层FPN field_stride = 2.lvl #元anchor的base_size,依次为4,8,16,32,64 anchor_sizes = (cfg.FPN.RPN_ANCHOR_START_SIZE * 2.(lvl - k_min), ) #每一层相应的的anchor size，依次为32 64 128 256 512（default下） anchor_aspect_ratios = cfg.FPN.RPN_ASPECT_RATIOS #(0.5,1,2) foa = data_utils.get_field_of_anchors( field_stride, anchor_sizes, anchor_aspect_ratios ) foas.append(foa) all_anchors = np.concatenate([f.field_of_anchors for f in foas]) #将每一层FPN产生的anchors合并在一起 else: foa = data_utils.get_field_of_anchors( cfg.RPN.STRIDE, cfg.RPN.SIZES, cfg.RPN.ASPECT_RATIOS ) all_anchors = foa.field_of_anchors

这里以P2-P6的FPN网络为例（训练图片的大小为1024×1024，其余参数皆为默认）：

• field_stride：表示以stride为步长，每隔一个stride生成一个anchor，同时在在生成anchor的过程中扮演元anchor的base_size的角色。见博客
• anchor_sizes：在同一个FPN层生成的anchor面积为anchor_sizes×anchor_sizes，其中对于P2层，其anchor_sizes的大小是由参数C.FPN.RPN_ANCHOR_START_SIZE决定的，后一层的anchor_sizes是前一层的2倍。
• anchor_aspect_ratios：在同以FPN层，在满足面积相同的情况下，生成三种长宽比例的anchor

每一个foa都代表着对应FPN层产生的anchor，下右二图就是foas，也就是P2层产生的anchor，以及相关的参数。

对后将foas中所有的anchor concatenate到一起，形成下面的形式，记为all_anchors，如下。为什么要这样的形式呢，是为了方便计算每一个anhcor与gt的重叠度，进而进行fg与bg的标记。

2.构成输入blob 

对一张图产生anchor之后，就要构建blob

for im_i, entry in enumerate(roidb): scale = im_scales[im_i] im_height = np.round(entry['height'] * scale) im_width = np.round(entry['width'] * scale) gt_inds = np.where( (entry['gt_classes'] > 0) & (entry['is_crowd'] == 0) )[0] gt_rois = entry['boxes'][gt_inds, :] * scale im_info = np.array([[im_height, im_width, scale]], dtype=np.float32) blobs['im_info'].append(im_info) # Add RPN targets if cfg.FPN.FPN_ON and cfg.FPN.MULTILEVEL_RPN: # RPN applied to many feature levels, as in the FPN paper rpn_blobs = _get_rpn_blobs( im_height, im_width, foas, all_anchors, gt_rois ) for i, lvl in enumerate(range(k_min, k_max + 1)): for k, v in rpn_blobs[i].items(): blobs[k + '_fpn' + str(lvl)].append(v) else: # Classical RPN, applied to a single feature level rpn_blobs = _get_rpn_blobs( im_height, im_width, [foa], all_anchors, gt_rois ) for k, v in rpn_blobs.items(): blobs[k].append(v)

这段代码是针对每一张送入的样本图片，获取其gt信息，主要包括：

• gt_rois：gt的box
• gt_classes ：gt的label
• im_height ：图片的高度（放缩后）
• im_width ：图片的宽度（放缩后）

获取到了上述的信息后，再调用_get_rpn_blobs（见下面），获取针对该样本图片的blob

def _get_rpn_blobs(im_height, im_width, foas, all_anchors, gt_boxes): total_anchors = all_anchors.shape[0] straddle_thresh = cfg.TRAIN.RPN_STRADDLE_THRESH if straddle_thresh >= 0: #保留在图片内部的anchors # Only keep anchors inside the image by a margin of straddle_thresh # Set TRAIN.RPN_STRADDLE_THRESH to -1 (or a large value) to keep all # anchors inds_inside = np.where( (all_anchors[:, 0] >= -straddle_thresh) & (all_anchors[:, 1] >= -straddle_thresh) & (all_anchors[:, 2] < im_width + straddle_thresh) & (all_anchors[:, 3] < im_height + straddle_thresh) )[0] # keep only inside anchors anchors = all_anchors[inds_inside, :] else: inds_inside = np.arange(all_anchors.shape[0]) anchors = all_anchors num_inside = len(inds_inside) logger.debug('total_anchors: {}'.format(total_anchors)) logger.debug('inds_inside: {}'.format(num_inside)) logger.debug('anchors.shape: {}'.format(anchors.shape)) # Compute anchor labels: # label=1 is positive, 0 is negative, -1 is don't care (ignore) labels = np.empty((num_inside, ), dtype=np.int32) #np.empty创建无意义的数组 labels.fill(-1) #将数组全都填补为-1 if len(gt_boxes) > 0: # Compute overlaps between the anchors and the gt boxes overlaps anchor_by_gt_overlap = box_utils.bbox_overlaps(anchors, gt_boxes) #计算每一个anchor与gt重叠率，anchor_by_gt_overlap.shape = [anchors_num, gt_num] # Map from anchor to gt box that has highest overlap anchor_to_gt_argmax = anchor_by_gt_overlap.argmax(axis=1) #返回每一个anchor与哪一个gt重叠率最大，anchor_to_gt_argmax.shape = [anchors_num, 1] # For each anchor, amount of overlap with most overlapping gt box anchor_to_gt_max = anchor_by_gt_overlap[np.arange(num_inside), #上述的重叠率是多少 anchor_to_gt_max.shape = [anchors_num, 1] anchor_to_gt_argmax] # Map from gt box to an anchor that has highest overlap gt_to_anchor_argmax = anchor_by_gt_overlap.argmax(axis=0) #返回与每一个gt重叠最大的anchor的index。gt_to_anchor_argmax.shape = (3,).axis=0表示就是对于每一列找出最大值，刚好每一列代表的就所有anchor与该gt的重叠率 # For each gt box, amount of overlap with most overlapping anchor gt_to_anchor_max = anchor_by_gt_overlap[ gt_to_anchor_argmax, np.arange(anchor_by_gt_overlap.shape[1]) #返回与每个gt重叠最大的重叠率 ] # Find all anchors that share the max overlap amount # (this includes many ties) anchors_with_max_overlap = np.where( anchor_by_gt_overlap == gt_to_anchor_max )[0] #找到所有共享这个最大重叠率的anchors # Fg label: for each gt use anchors with highest overlap # (including ties) labels[anchors_with_max_overlap] = 1 #1.首先将这些重叠最大的anchor，label设置为1 # Fg label: above threshold IOU labels[anchor_to_gt_max >= cfg.TRAIN.RPN_POSITIVE_OVERLAP] = 1 #2.其次将大于0.7的重叠率的anchor的label设置为1 # subsample positive labels if we have too many num_fg = int(cfg.TRAIN.RPN_FG_FRACTION * cfg.TRAIN.RPN_BATCH_SIZE_PER_IM) #设置的前景数量，default：256×0.5 fg_inds = np.where(labels == 1)[0] if len(fg_inds) > num_fg: disable_inds = npr.choice( fg_inds, size=(len(fg_inds) - num_fg), replace=False ) #多出来的数量为size=(len(fg_inds) - num_fg)，随机地从fg_inds选出来，设置为False labels[disable_inds] = -1 fg_inds = np.where(labels == 1)[0] # subsample negative labels if we have too many # (samples with replacement, but since the set of bg inds is large most # samples will not have repeats) num_bg = cfg.TRAIN.RPN_BATCH_SIZE_PER_IM - np.sum(labels == 1) #需要的bg数量 bg_inds = np.where(anchor_to_gt_max < cfg.TRAIN.RPN_NEGATIVE_OVERLAP)[0] #实际的bg数量 if len(bg_inds) > num_bg: enable_inds = bg_inds[npr.randint(len(bg_inds), size=num_bg)] else: enable_inds = bg_inds labels[enable_inds] = 0 bg_inds = np.where(labels == 0)[0] bbox_targets = np.zeros((num_inside, 4), dtype=np.float32) bbox_targets[fg_inds, :] = data_utils.compute_targets( anchors[fg_inds, :], gt_boxes[anchor_to_gt_argmax[fg_inds], :] #根据fg_inds，取出对应的gt的index（anchor_to_gt_argmax[fg_inds]），再得到对应的gt（gt_boxes[anchor_to_gt_argmax[fg_inds], :]） ) # Bbox regression loss has the form: # loss(x) = weight_outside * L(weight_inside * x) # Inside weights allow us to set zero loss on an element-wise basis # Bbox regression is only trained on positive examples so we set their # weights to 1.0 (or otherwise if config is different) and 0 otherwise bbox_inside_weights = np.zeros((num_inside, 4), dtype=np.float32) bbox_inside_weights[labels == 1, :] = (1.0, 1.0, 1.0, 1.0) # The bbox regression loss only averages by the number of images in the # mini-batch, whereas we need to average by the total number of example # anchors selected # Outside weights are used to scale each element-wise loss so the final # average over the mini-batch is correct bbox_outside_weights = np.zeros((num_inside, 4), dtype=np.float32) # uniform weighting of examples (given non-uniform sampling) num_examples = np.sum(labels >= 0) #其实就是batch_size的数量 256 bbox_outside_weights[labels == 1, :] = 1.0 / num_examples bbox_outside_weights[labels == 0, :] = 1.0 / num_examples # Map up to original set of anchors labels = data_utils.unmap(labels, total_anchors, inds_inside, fill=-1) bbox_targets = data_utils.unmap( bbox_targets, total_anchors, inds_inside, fill=0 ) bbox_inside_weights = data_utils.unmap( bbox_inside_weights, total_anchors, inds_inside, fill=0 ) bbox_outside_weights = data_utils.unmap( bbox_outside_weights, total_anchors, inds_inside, fill=0 ) #利用合并的all_anchors对所有anchor贴标签和生成bbox_targets，bbox_inside_weights，bbox_outside_weights #但是foas中是没有上述的标签以及参数的，所以就要Split the generated labels, etc. into labels per each field of anchors blobs_out = [] start_idx = 0 for foa in foas: H = foa.field_size W = foa.field_size A = foa.num_cell_anchors end_idx = start_idx + H * W * A #也就是anchor的数量 _labels = labels[start_idx:end_idx] #因为lebels是按顺序（P2-P6依次排列的），所以取前面 _bbox_targets = bbox_targets[start_idx:end_idx, :] _bbox_inside_weights = bbox_inside_weights[start_idx:end_idx, :] _bbox_outside_weights = bbox_outside_weights[start_idx:end_idx, :] start_idx = end_idx # labels output with shape (1, A, height, width) _labels = _labels.reshape((1, H, W, A)).transpose(0, 3, 1, 2) # bbox_targets output with shape (1, 4 * A, height, width) _bbox_targets = _bbox_targets.reshape( (1, H, W, A * 4)).transpose(0, 3, 1, 2) # bbox_inside_weights output with shape (1, 4 * A, height, width) _bbox_inside_weights = _bbox_inside_weights.reshape( (1, H, W, A * 4)).transpose(0, 3, 1, 2) # bbox_outside_weights output with shape (1, 4 * A, height, width) _bbox_outside_weights = _bbox_outside_weights.reshape( (1, H, W, A * 4)).transpose(0, 3, 1, 2) blobs_out.append( dict( rpn_labels_int32_wide=_labels, rpn_bbox_targets_wide=_bbox_targets, rpn_bbox_inside_weights_wide=_bbox_inside_weights, rpn_bbox_outside_weights_wide=_bbox_outside_weights ) ) return blobs_out[0] if len(blobs_out) == 1 else blobs_out

上述代码完成的内容相当于anchor_target_layer，可见博客。最后一个for循环完成的任务是将与all_anchor同维度（行数一致）的labels，bbox_targets，inside_weights，outside_weights重新分配成foas的形式，见下图。

最后该函数返回的rpn_blobs形式如下，右图表示P2层。

由于blob是如下形式，所以还要将rpn_blobs中每一层的值对应的付给