minibatch.py 的功能是： Compute minibatch blobs for training a Fast R-CNN network. 与roidb不同的是， minibatch中存储的并不是完整的整张图像图像，而是从图像经过转换后得到的四维blob以及从图像中截取的proposals，以及与之对应的labels等

在整个faster rcnn训练中，有两处用到了minibatch.py，一处是rpn的开始数据输入，另一处自然是fast rcnn的数据输入。分别见stage1_rpn_train.pt与stage1_fast_rcnn_train.py的最前面，如下：
stage1_rpn_train.pt：

layer {name: 'input-data'type: 'Python'top: 'data'top: 'im_info'top: 'gt_boxes'python_param {module: 'roi_data_layer.layer'layer: 'RoIDataLayer'param_str: "'num_classes': 21"}
}

stage1_fast_rcnn_train.py：

name: "VGG_CNN_M_1024"
layer {name: 'data'type: 'Python'top: 'data'top: 'rois'top: 'labels'top: 'bbox_targets'top: 'bbox_inside_weights'top: 'bbox_outside_weights'python_param {module: 'roi_data_layer.layer'layer: 'RoIDataLayer'param_str: "'num_classes': 21"}
}

如上，共同的数据定义层为roi_data_layer.layer，在layer.py中，观察前向传播：

 def forward(self, bottom, top):"""Get blobs and copy them into this layer's top blob vector."""blobs = self._get_next_minibatch()for blob_name, blob in blobs.iteritems():top_ind = self._name_to_top_map[blob_name]# Reshape net's input blobstop[top_ind].reshape(*(blob.shape))# Copy data into net's input blobstop[top_ind].data[...] = blob.astype(np.float32, copy=False)def _get_next_minibatch(self):"""Return the blobs to be used for the next minibatch.If cfg.TRAIN.USE_PREFETCH is True, then blobs will be computed in aseparate process and made available through self._blob_queue."""if cfg.TRAIN.USE_PREFETCH:return self._blob_queue.get()else:db_inds = self._get_next_minibatch_inds()minibatch_db = [self._roidb[i] for i in db_inds]return get_minibatch(minibatch_db, self._num_classes)

这时我们发现了get_minibatch，此函数出现在minibatch.py中。

在看这份代码的时候，建议从get_minibatch开始。下面我们开始：
get_minibatch中，【输入】：roidb是一个list，list中的每个元素是一个字典，每个字典对应一张图片的信息，其中的主要信息有：

num_classes在pascal_voc中为21.

def get_minibatch(roidb, num_classes):"""Given a roidb, construct a minibatch sampled from it."""# 给定一个roidb，这个roidb中存储的可能是多张图片，也可能是单张或者多张图片，num_images = len(roidb) # Sample random scales to use for each image in this batchrandom_scale_inds = npr.randint(0, high=len(cfg.TRAIN.SCALES),size=num_images)assert(cfg.TRAIN.BATCH_SIZE % num_images == 0), \'num_images ({}) must divide BATCH_SIZE ({})'. \format(num_images, cfg.TRAIN.BATCH_SIZE)rois_per_image = cfg.TRAIN.BATCH_SIZE / num_images  #这里在fast rcnn中，为128/2=64fg_rois_per_image = np.round(cfg.TRAIN.FG_FRACTION * rois_per_image)#这里比例为0.25=1/4# Get the input image blob, formatted for caffe#将给定的roidb经过预处理（resize以及resize的scale），#然后再利用im_list_to_blob函数来将图像转换成caffe支持的数据结构，即 N * C * H * W的四维结构im_blob, im_scales = _get_image_blob(roidb, random_scale_inds)blobs = {'data': im_blob}if cfg.TRAIN.HAS_RPN:#用在rpnassert len(im_scales) == 1, "Single batch only"assert len(roidb) == 1, "Single batch only"# gt boxes: (x1, y1, x2, y2, cls)gt_inds = np.where(roidb[0]['gt_classes'] != 0)[0]gt_boxes = np.empty((len(gt_inds), 5), dtype=np.float32)gt_boxes[:, 0:4] = roidb[0]['boxes'][gt_inds, :] * im_scales[0]gt_boxes[:, 4] = roidb[0]['gt_classes'][gt_inds]blobs['gt_boxes'] = gt_boxes#首先解释im_info。对于一副任意大小PxQ图像，传入Faster RCNN前首先reshape到固定MxN，im_info=[M, N, scale_factor]则保存了此次缩放的所有信息。blobs['im_info'] = np.array( [[im_blob.shape[2], im_blob.shape[3], im_scales[0]]],dtype=np.float32)else: # not using RPN ，用在fast rcnn# Now, build the region of interest and label blobsrois_blob = np.zeros((0, 5), dtype=np.float32)labels_blob = np.zeros((0), dtype=np.float32)bbox_targets_blob = np.zeros((0, 4 * num_classes), dtype=np.float32)bbox_inside_blob = np.zeros(bbox_targets_blob.shape, dtype=np.float32)# all_overlaps = []for im_i in xrange(num_images):# 遍历给定的roidb中的每张图片，随机组合sample of RoIs， 来生成前景样本和背景样本。# 返回包括每张图片中的roi（proposal）的坐标，所属的类别，bbox回归目标，bbox的inside_weight等                        labels, overlaps, im_rois, bbox_targets, bbox_inside_weights \= _sample_rois(roidb[im_i], fg_rois_per_image, rois_per_image,num_classes)# Add to RoIs blob# _sample_rois返回的im_rois并没有缩放，所以这里要先缩放rois = _project_im_rois(im_rois, im_scales[im_i])batch_ind = im_i * np.ones((rois.shape[0], 1))rois_blob_this_image = np.hstack((batch_ind, rois))# 加上图片的序号，共5列（index,x1,y1,x2,y2）rois_blob = np.vstack((rois_blob, rois_blob_this_image))# 将所有的盒子竖着摆放，如下：# n  x1  y1  x2  y2# 0  ..  ..  ..  ..# 0  ..  ..  ..  ..# :   :   :   :   :# 1   ..  ..  ..  ..# 1   ..  ..  ..  ..# Add to labels, bbox targets, and bbox loss blobslabels_blob = np.hstack((labels_blob, labels))# 水平向量,一维向量bbox_targets_blob = np.vstack((bbox_targets_blob, bbox_targets))bbox_inside_blob = np.vstack((bbox_inside_blob, bbox_inside_weights))# 将所有的bbox_targets_blob竖着摆放，如下： N*4k ，只有对应的类非0#   tx1  ty1  wx1  wy1   tx2  ty2  wx2  wy2    tx3  ty3  wx3  wy3#    0     0    0   0     0     0    0   0       0     0    0   0#    0     0    0   0     0.2   0.3  1.0 0.5     0     0    0   0#    0     0    0   0     0     0    0   0       0     0    0   0#    0     0    0   0     0     0    0   0       0.5   0.5  1.0  1.0#    0     0    0   0     0     0    0   0       0     0    0   0# 对于bbox_inside_blob ，与bbox_targets_blob 规模相同，只不过把上面非0的元素换成1即可。# all_overlaps = np.hstack((all_overlaps, overlaps))# For debug visualizations# _vis_minibatch(im_blob, rois_blob, labels_blob, all_overlaps)blobs['rois'] = rois_blobblobs['labels'] = labels_blobif cfg.TRAIN.BBOX_REG:blobs['bbox_targets'] = bbox_targets_blobblobs['bbox_inside_weights'] = bbox_inside_blobblobs['bbox_outside_weights'] = \np.array(bbox_inside_blob > 0).astype(np.float32)
#对于bbox_outside_weights，此处看来与bbox_inside_blob 相同。return blobs

在 def get_minibatch(roidb, num_classes) 中调用此函数，传进来的实参为单张图像的roidb ，该函数主要功能是随机组合sample of RoIs，来生成前景样本和背景样本。这里很重要，
因为一般来说，生成的proposal背景类比较多，所以我们生成前景与背景的比例选择为1:3，所以
这里每张图片选取了1/4*64=16个前景，选取了3/4*64=48个背景box.

还有一个值得注意的是随机采样中，前景box可能会包含ground truth box.可能会参与分类，但是不会参加回归，因为其回归量为0. 是不是可以将

fg_inds = np.where(overlaps >= cfg.TRAIN.FG_THRESH)[0]

改为：

fg_inds = np.where(overlaps >= cfg.TRAIN.FG_THRESH && overlaps <1.0)[0]

会更合适呢，这样就可以提取的全部是rpn的 proposal。

def _sample_rois(roidb, fg_rois_per_image, rois_per_image, num_classes):"""Generate a random sample of RoIs comprising foreground and backgroundexamples."""# label = class RoI has max overlap withlabels = roidb['max_classes']overlaps = roidb['max_overlaps']rois = roidb['boxes']# Select foreground RoIs as those with >= FG_THRESH overlapfg_inds = np.where(overlaps >= cfg.TRAIN.FG_THRESH)[0]# Guard against the case when an image has fewer than fg_rois_per_image# foreground RoIsfg_rois_per_this_image = np.minimum(fg_rois_per_image, fg_inds.size)# Sample foreground regions without replacementif fg_inds.size > 0:fg_inds = npr.choice(fg_inds, size=fg_rois_per_this_image, replace=False)# Select background RoIs as those within [BG_THRESH_LO, BG_THRESH_HI)bg_inds = np.where((overlaps < cfg.TRAIN.BG_THRESH_HI) &(overlaps >= cfg.TRAIN.BG_THRESH_LO))[0]# Compute number of background RoIs to take from this image (guarding# against there being fewer than desired)bg_rois_per_this_image = rois_per_image - fg_rois_per_this_imagebg_rois_per_this_image = np.minimum(bg_rois_per_this_image,bg_inds.size)# Sample foreground regions without replacementif bg_inds.size > 0:bg_inds = npr.choice(bg_inds, size=bg_rois_per_this_image, replace=False)# The indices that we're selecting (both fg and bg)keep_inds = np.append(fg_inds, bg_inds)# Select sampled values from various arrays:labels = labels[keep_inds]# Clamp labels for the background RoIs to 0labels[fg_rois_per_this_image:] = 0overlaps = overlaps[keep_inds]rois = rois[keep_inds]# 调用_get_bbox_regression_labels函数，生成bbox_targets 和 bbox_inside_weights，#它们都是N * 4K 的ndarray，N表示keep_inds的size，也就是minibatch中样本的个数；bbox_inside_weights #也随之生成bbox_targets, bbox_inside_weights = _get_bbox_regression_labels(roidb['bbox_targets'][keep_inds, :], num_classes)return labels, overlaps, rois, bbox_targets, bbox_inside_weights

def _get_bbox_regression_labels(bbox_target_data, num_classes):
该函数主要是获取bbox_target_data中回归目标的的4个坐标编码作为bbox_targets，同时生成bbox_inside_weights，它们都是N * 4K 的ndarray，N表示keep_inds的size，也就是minibatch中样本的个数。
bbox_target_data: N*5 ,每一行为（c,tx,ty,tw,th）

def _get_bbox_regression_labels(bbox_target_data, num_classes):"""Bounding-box regression targets are stored in a compact form in theroidb.This function expands those targets into the 4-of-4*K representation usedby the network (i.e. only one class has non-zero targets). The loss weightsare similarly expanded.Returns:bbox_target_data (ndarray): N x 4K blob of regression targetsbbox_inside_weights (ndarray): N x 4K blob of loss weights"""clss = bbox_target_data[:, 0]bbox_targets = np.zeros((clss.size, 4 * num_classes), dtype=np.float32)bbox_inside_weights = np.zeros(bbox_targets.shape, dtype=np.float32)inds = np.where(clss > 0)[0] # 取前景框for ind in inds:cls = clss[ind]start = 4 * clsend = start + 4bbox_targets[ind, start:end] = bbox_target_data[ind, 1:]bbox_inside_weights[ind, start:end] = cfg.TRAIN.BBOX_INSIDE_WEIGHTSreturn bbox_targets, bbox_inside_weights

对于roidb的图像进行对应的缩放操作，并返回统一的blob数据，即 N * C * H * W（这里为2*3*600*1000）的四维结构

def _get_image_blob(roidb, scale_inds):"""Builds an input blob from the images in the roidb at the specifiedscales."""num_images = len(roidb)processed_ims = []im_scales = []for i in xrange(num_images):im = cv2.imread(roidb[i]['image'])  #shape:h*w*cif roidb[i]['flipped']:im = im[:, ::-1, :]   # 水平翻转target_size = cfg.TRAIN.SCALES[scale_inds[i]]im, im_scale = prep_im_for_blob(im, cfg.PIXEL_MEANS, target_size,cfg.TRAIN.MAX_SIZE)im_scales.append(im_scale)processed_ims.append(im)# Create a blob to hold the input imagesblob = im_list_to_blob(processed_ims)return blob, im_scales

以上im_list_to_blob中将一系列的图像转化为标准的4维矩阵，进行了填0的补全操作，使得所有的图片的大小相同。

prep_im_for_blob 进行尺寸变化，使得最小的边长为target_size，最大的边长不超过cfg.TRAIN.MAX_SIZE，并且返回缩放的比例。

def prep_im_for_blob(im, pixel_means, target_size, max_size):"""Mean subtract and scale an image for use in a blob."""im = im.astype(np.float32, copy=False)im -= pixel_meansim_shape = im.shapeim_size_min = np.min(im_shape[0:2])im_size_max = np.max(im_shape[0:2])im_scale = float(target_size) / float(im_size_min)# Prevent the biggest axis from being more than MAX_SIZEif np.round(im_scale * im_size_max) > max_size:im_scale = float(max_size) / float(im_size_max)im = cv2.resize(im, None, None, fx=im_scale, fy=im_scale,interpolation=cv2.INTER_LINEAR)return im, im_scale