gluoncv，faster rcnn 处理难样本
难样本，是我们在标注的时候，连肉眼都看不清的小像素物体，也可以说是既不是正样本，也不是负样本。
利用gluoncv时，这些标注框也实在存在，gluoncv会实在将他当做一个GT，但我们知道这是不好的。于是，我们在标注的时候，给他一个属性，他是一个难样本。之前用过lst文件训练自己的数据集，这里的Lst文件需要改一下格式；
这里的B就是6了，包括id,xmin,ymin,xmax,ymax,hard_flag，可以自己写，结构类似即可。
Faster rcnn：
https://github.com/dmlc/gluon-cv/blob/master/scripts/detection/faster_rcnn/train_faster_rcnn.py#L313
从这个大epoch里面，我们可以知道，将我们的难样本属性应该在train_data之前就弄好了。
也是就是说在：
from gluoncv.data.transforms.presets.rcnn import FasterRCNNDefaultTrainTransform from gluoncv.data.transforms.presets.rcnn import FasterRCNNDefaultValTransform
这两个变化里，注意import方式，pip 进来就修改site-package里面的文件。
class FasterRCNNDefaultTrainTransform(object): """Default Faster-RCNN training transform. Parameters ---------- short : int, default is 600 Resize image shorter side to ``short``. max_size : int, default is 1000 Make sure image longer side is smaller than ``max_size``. net : mxnet.gluon.HybridBlock, optional The faster-rcnn network. .. hint:: If net is ``None``, the transformation will not generate training targets. Otherwise it will generate training targets to accelerate the training phase since we push some workload to CPU workers instead of GPUs. mean : array-like of size 3 Mean pixel values to be subtracted from image tensor. Default is [0.485, 0.456, 0.406]. std : array-like of size 3 Standard deviation to be divided from image. Default is [0.229, 0.224, 0.225]. box_norm : array-like of size 4, default is (1., 1., 1., 1.) Std value to be divided from encoded values. num_sample : int, default is 256 Number of samples for RPN targets. pos_iou_thresh : float, default is 0.7 Anchors larger than ``pos_iou_thresh`` is regarded as positive samples. neg_iou_thresh : float, default is 0.3 Anchors smaller than ``neg_iou_thresh`` is regarded as negative samples. Anchors with IOU in between ``pos_iou_thresh`` and ``neg_iou_thresh`` are ignored. pos_ratio : float, default is 0.5 ``pos_ratio`` defines how many positive samples (``pos_ratio * num_sample``) is to be sampled. """ def __init__(self, short=600, max_size=1000, net=None, mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225), box_norm=(1., 1., 1., 1.), num_sample=256, pos_iou_thresh=0.7, neg_iou_thresh=0.3, pos_ratio=0.5, **kwargs): self._short = short self._max_size = max_size self._mean = mean self._std = std self._anchors = None if net is None: return # use fake data to generate fixed anchors for target generation ashape = 128 # in case network has reset_ctx to gpu anchor_generator = copy.deepcopy(net.rpn.anchor_generator) anchor_generator.collect_params().reset_ctx(None) anchors = anchor_generator( mx.nd.zeros((1, 3, ashape, ashape))).reshape((1, 1, ashape, ashape, -1)) self._anchors = anchors # record feature extractor for infer_shape if not hasattr(net, 'features'): raise ValueError("Cannot find features in network, it is a Faster-RCNN network?") self._feat_sym = net.features(mx.sym.var(name='data')) from ....model_zoo.rpn.rpn_target import RPNTargetGenerator self._target_generator = RPNTargetGenerator( num_sample=num_sample, pos_iou_thresh=pos_iou_thresh, neg_iou_thresh=neg_iou_thresh, pos_ratio=pos_ratio, stds=box_norm, **kwargs) def __call__(self, src, label): """Apply transform to training image/label.""" # resize shorter side but keep in max_size h, w, _ = src.shape img = timage.resize_short_within(src, self._short, self._max_size, interp=1) bbox = tbbox.resize(label, (w, h), (img.shape[1], img.shape[0])) # random horizontal flip h, w, _ = img.shape img, flips = timage.random_flip(img, px=0.5) bbox = tbbox.flip(bbox, (w, h), flip_x=flips[0]) # to tensor img = mx.nd.image.to_tensor(img) img = mx.nd.image.normalize(img, mean=self._mean, std=self._std) if self._anchors is None: return img, bbox.astype(img.dtype) # print 'xxxxxxxxxxxxxxxxxxxxxxxxxxxxlabal',label.shape # print label[:,5] # generate RPN target so cpu workers can help reduce the workload # feat_h, feat_w = (img.shape[1] // self._stride, img.shape[2] // self._stride) oshape = self._feat_sym.infer_shape(data=(1, 3, img.shape[1], img.shape[2]))[1][0] anchor = self._anchors[:, :, :oshape[2], :oshape[3], :].reshape((-1, 4)) gt_bboxes = mx.nd.array(bbox[:, :4]) cls_target, box_target, box_mask = self._target_generator( gt_bboxes, anchor, img.shape[2], img.shape[1],label[:,5]) return img, bbox.astype(img.dtype), cls_target, box_target, box_mask
可以检查一下这里的label的结构，和我们修改的一样
我们可以知道rpn处生成target的部分就是.target_generator，他是
from ....model_zoo.rpn.rpn_target import RPNTargetGenerator
添加上我们难样本属性label[:,5]
过来的，于是找到那里；
class RPNTargetGenerator(gluon.Block): """RPN target generator network. Parameters ---------- num_sample : int, default is 256 Number of samples for RPN targets. pos_iou_thresh : float, default is 0.7 Anchor with IOU larger than ``pos_iou_thresh`` is regarded as positive samples. neg_iou_thresh : float, default is 0.3 Anchor with IOU smaller than ``neg_iou_thresh`` is regarded as negative samples. Anchors with IOU in between ``pos_iou_thresh`` and ``neg_iou_thresh`` are ignored. pos_ratio : float, default is 0.5 ``pos_ratio`` defines how many positive samples (``pos_ratio * num_sample``) is to be sampled. stds : array-like of size 4, default is (1., 1., 1., 1.) Std value to be divided from encoded regression targets. allowed_border : int or float, default is 0 The allowed distance of anchors which are off the image border. This is used to clip out of border anchors. You can set it to very large value to keep all anchors. """ def __init__(self, num_sample=256, pos_iou_thresh=0.7, neg_iou_thresh=0.3, pos_ratio=0.5, stds=(1., 1., 1., 1.), allowed_border=0): super(RPNTargetGenerator, self).__init__() self._num_sample = num_sample self._pos_iou_thresh = pos_iou_thresh self._neg_iou_thresh = neg_iou_thresh self._pos_ratio = pos_ratio self._allowed_border = allowed_border self._bbox_split = BBoxSplit(axis=-1) self._sampler = RPNTargetSampler(num_sample, pos_iou_thresh, neg_iou_thresh, pos_ratio) self._cls_encoder = SigmoidClassEncoder() self._box_encoder = NormalizedBoxCenterEncoder(stds=stds) # pylint: disable=arguments-differ def forward(self, bbox, anchor, width, height,hard_label): """ RPNTargetGenerator is only used in data transform with no batch dimension. Be careful there's numpy operations inside Parameters ---------- bbox: (M, 4) ground truth boxes with corner encoding. anchor: (N, 4) anchor boxes with corner encoding. width: int width of input image height: int height of input image Returns ------- cls_target: (N,) value +1: pos, 0: neg, -1: ignore box_target: (N, 4) only anchors whose cls_target > 0 has nonzero box target box_mask: (N, 4) only anchors whose cls_target > 0 has nonzero mask """ F = mx.nd with autograd.pause(): # calculate ious between (N, 4) anchors and (M, 4) bbox ground-truths # ious is (N, M) ious = mx.nd.contrib.box_iou(anchor, bbox, format='corner') N, M = ious.shape # print '------------------ious.shape',ious.shape # print ious[0:5,] # print M==hard_label.size # mask out invalid anchors, (N, 4) a_xmin, a_ymin, a_xmax, a_ymax = F.split(anchor, num_outputs=4, axis=-1) invalid_mask = (a_xmin < 0) + (a_ymin < 0) + (a_xmax >= width) + (a_ymax >= height) invalid_mask = F.repeat(invalid_mask, repeats=bbox.shape[0], axis=-1) ious = F.where(invalid_mask, mx.nd.ones_like(ious) * -1, ious) samples, matches = self._sampler(ious,hard_label) # training targets for RPN cls_target, _ = self._cls_encoder(samples) box_target, box_mask = self._box_encoder( samples.expand_dims(axis=0), matches.expand_dims(0), anchor.expand_dims(axis=0), bbox.expand_dims(0)) return cls_target, box_target[0], box_mask[0]
可以看到生成sample的地方就是self._sampler，它是
self._sampler = RPNTargetSampler(num_sample, pos_iou_thresh, neg_iou_thresh, pos_ratio)
添加自己的属性hard_label，找到那里(同一个文件里)：
class RPNTargetSampler(gluon.Block): """A sampler to choose positive/negative samples from RPN anchors Parameters ---------- num_sample : int Number of samples for RCNN targets. pos_iou_thresh : float Proposal whose IOU larger than ``pos_iou_thresh`` is regarded as positive samples. neg_iou_thresh : float Proposal whose IOU smaller than ``neg_iou_thresh`` is regarded as negative samples. pos_ratio : float ``pos_ratio`` defines how many positive samples (``pos_ratio * num_sample``) is to be sampled. """ def __init__(self, num_sample, pos_iou_thresh, neg_iou_thresh, pos_ratio): super(RPNTargetSampler, self).__init__() self._num_sample = num_sample self._max_pos = int(round(num_sample * pos_ratio)) self._pos_iou_thresh = pos_iou_thresh self._neg_iou_thresh = neg_iou_thresh self._eps = np.spacing(np.float32(1.0)) # pylint: disable=arguments-differ def forward(self, ious,hard_label): """RPNTargetSampler is only used in data transform with no batch dimension. Parameters ---------- ious: (N, M) i.e. (num_anchors, num_gt). Returns ------- samples: (num_anchors,) value 1: pos, -1: neg, 0: ignore. matches: (num_anchors,) value [0, M). """ for i,hard in enumerate(hard_label): if hard == 1.0: ious[:,i] = 0.5 matches = mx.nd.argmax(ious, axis=1) # samples init with 0 (ignore) ious_max_per_anchor = mx.nd.max(ious, axis=1) samples = mx.nd.zeros_like(ious_max_per_anchor) # set argmax (1, num_gt) ious_max_per_gt = mx.nd.max(ious, axis=0, keepdims=True) # ious (num_anchor, num_gt) >= argmax (1, num_gt) -> mark row as positive mask = mx.nd.broadcast_greater(ious + self._eps, ious_max_per_gt) # reduce column (num_anchor, num_gt) -> (num_anchor) mask = mx.nd.sum(mask, axis=1) # row maybe sampled by 2 columns but still only matches to most overlapping gt samples = mx.nd.where(mask, mx.nd.ones_like(samples), samples) # set positive overlap to 1 samples = mx.nd.where(ious_max_per_anchor >= self._pos_iou_thresh, mx.nd.ones_like(samples), samples) # set negative overlap to -1 tmp = (ious_max_per_anchor < self._neg_iou_thresh) * (ious_max_per_anchor >= 0) samples = mx.nd.where(tmp, mx.nd.ones_like(samples) * -1, samples) # subsample fg labels samples = samples.asnumpy() num_pos = int((samples > 0).sum()) if num_pos > self._max_pos: disable_indices = np.random.choice( np.where(samples > 0)[0], size=(num_pos - self._max_pos), replace=False) samples[disable_indices] = 0 # use 0 to ignore # subsample bg labels num_neg = int((samples < 0).sum()) # if pos_sample is less than quota, we can have negative samples filling the gap max_neg = self._num_sample - min(num_pos, self._max_pos) if num_neg > max_neg: disable_indices = np.random.choice( np.where(samples < 0)[0], size=(num_neg - max_neg), replace=False) samples[disable_indices] = 0 # convert to ndarray samples = mx.nd.array(samples, ctx=matches.context) return samples, matches
加上自己的hard部分，默认0.7~0.3 既不是正样本，也不是负样本。
经过试验，直接改为既不是正样本，也不是负样本，效果不好，而他标记为难gt，而其他锚框与他交为正样本，则改为既不是正样本，也不是负样本，效果会好一点点。
很多教程一本正经说Faster rcnn 的网络结构如下：
不可否认的，但是绝对不是gluon cv 的faster rcnn网络结构，后面的全连接层是RCNN层，想要达到更好的效果处理难样本，那里同样需要修改。
