Faster RCNN算法训练代码解析（2）

接着上篇的博客，我们获取imdb和roidb的数据后，就可以搭建网络进行训练了。

我们回到trian_rpn()函数里面，此时运行完了roidb, imdb = get_roidb(imdb_name)，取得了imdb和roidb数据。

先进入第一阶段的训练：

    print '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'
    print 'Stage 1 RPN, init from ImageNet model'
    print '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'

    cfg.TRAIN.SNAPSHOT_INFIX = 'stage1'
    mp_kwargs = dict(
            queue=mp_queue,
            imdb_name=args.imdb_name,
            init_model=args.pretrained_model,
            solver=solvers[0],
            max_iters=max_iters[0],
            cfg=cfg)
    p = mp.Process(target=train_rpn, kwargs=mp_kwargs)  ##创建线程对象
    p.start() ##开始进程
    rpn_stage1_out = mp_queue.get()  ##从进程里面获取运行结果
    p.join()  ##等待子进程结束

进入子进程train_rpn：

def train_rpn(queue=None, imdb_name=None, init_model=None, solver=None,
              max_iters=None, cfg=None):
    """Train a Region Proposal Network in a separate training process.
    """

    # Not using any proposals, just ground-truth boxes
    cfg.TRAIN.HAS_RPN = True
    cfg.TRAIN.BBOX_REG = False  # applies only to Fast R-CNN bbox regression
    cfg.TRAIN.PROPOSAL_METHOD = 'gt'
    cfg.TRAIN.IMS_PER_BATCH = 1
    print 'Init model: {}'.format(init_model)
    print('Using config:')
    pprint.pprint(cfg)

    import caffe
    _init_caffe(cfg)

    roidb, imdb = get_roidb(imdb_name)  ##获取roidb和imdb格式的数据集
    print 'roidb len: {}'.format(len(roidb))
    output_dir = get_output_dir(imdb)   ##训练的输出路径:'py-faster-rcnn/output/faster_rcnn_alt_opt/voc_2007_trainval'
    print 'Output will be saved to `{:s}`'.format(output_dir)

    model_paths = train_net(solver, roidb, output_dir,  ##进入训练网络，其实里面还包裹了一层solver，定义该对象之后才进行训练。
                            pretrained_model=init_model,
                            max_iters=max_iters)
    # Cleanup all but the final model
    for i in model_paths[:-1]:   ##除了最后一个快照，把所有其他的快照都清除掉
        os.remove(i)
    rpn_model_path = model_paths[-1]
    # Send final model path through the multiprocessing queue
    queue.put({'model_path': rpn_model_path})  ##将最后的快照保存到线程里面，进行线程通信。

接着我们运行到了model_paths = train_net(solver, roidb, output_dir, pretrained_model=init_model, max_iters=max_iters)函数，我们进入该函数里面看看：

def train_net(solver_prototxt, roidb, output_dir,
              pretrained_model=None, max_iters=40000):
    """Train a Fast R-CNN network."""

    roidb = filter_roidb(roidb)  ##过滤部分roidb，具体判断一个图片的roidb是否合格：前景大于某个值，背景在某个范围内，不符合则过滤掉
    sw = SolverWrapper(solver_prototxt, roidb, output_dir,
                       pretrained_model=pretrained_model)

    print 'Solving...'
    model_paths = sw.train_model(max_iters)  ##开始训练函数
    print 'done solving'
    return model_paths

接着我们进入sw = SolverWrapper(solver_prototxt, roidb, output_dir, pretrained_model=pretrained_model) 函数：

class SolverWrapper(object):
    """A simple wrapper around Caffe's solver.
    This wrapper gives us control over he snapshotting process, which we
    use to unnormalize the learned bounding-box regression weights.
    """

    def __init__(self, solver_prototxt, roidb, output_dir,
                 pretrained_model=None):
        """Initialize the SolverWrapper."""
        self.output_dir = output_dir

        if (cfg.TRAIN.HAS_RPN and cfg.TRAIN.BBOX_REG and
            cfg.TRAIN.BBOX_NORMALIZE_TARGETS):
            # RPN can only use precomputed normalization because there are no
            # fixed statistics to compute a priori
            assert cfg.TRAIN.BBOX_NORMALIZE_TARGETS_PRECOMPUTED

        if cfg.TRAIN.BBOX_REG:
            print 'Computing bounding-box regression targets...'
            self.bbox_means, self.bbox_stds = 
                    rdl_roidb.add_bbox_regression_targets(roidb)
            print 'done'

        self.solver = caffe.SGDSolver(solver_prototxt)   ##该函数主要是用来建造层的，这里建立ROI层和anchors层
        if pretrained_model is not None:   ##如果存在'data/imagenet_models/ZF.v2.caffemodel'则加载进来
            print ('Loading pretrained model '
                   'weights from {:s}').format(pretrained_model)
            self.solver.net.copy_from(pretrained_model)  

        self.solver_param = caffe_pb2.SolverParameter()
        with open(solver_prototxt, 'rt') as f:   ##加载py-faster-rcnn/models/pascal_voc/ZF/faster_rcnn_alt_opt/stage1_rpn_solver60k80k.pt文件参数
            pb2.text_format.Merge(f.read(), self.solver_param)

        self.solver.net.layers[0].set_roidb(roidb)  ##创建该包裹对象的时候把外部的roidb设置进包裹函数的self.solver.net.layers[0]里面，用以训练

    def snapshot(self):
        """Take a snapshot of the network after unnormalizing the learned
        bounding-box regression weights. This enables easy use at test-time.
        """
        net = self.solver.net

        scale_bbox_params = (cfg.TRAIN.BBOX_REG and
                             cfg.TRAIN.BBOX_NORMALIZE_TARGETS and
                             net.params.has_key('bbox_pred'))

        if scale_bbox_params:
            # save original values
            orig_0 = net.params['bbox_pred'][0].data.copy()
            orig_1 = net.params['bbox_pred'][1].data.copy()

            # scale and shift with bbox reg unnormalization; then save snapshot
            net.params['bbox_pred'][0].data[...] = 
                    (net.params['bbox_pred'][0].data *
                     self.bbox_stds[:, np.newaxis])
            net.params['bbox_pred'][1].data[...] = 
                    (net.params['bbox_pred'][1].data *
                     self.bbox_stds + self.bbox_means)

        infix = ('_' + cfg.TRAIN.SNAPSHOT_INFIX
                 if cfg.TRAIN.SNAPSHOT_INFIX != '' else '')
        filename = (self.solver_param.snapshot_prefix + infix +
                    '_iter_{:d}'.format(self.solver.iter) + '.caffemodel')
        filename = os.path.join(self.output_dir, filename)

        net.save(str(filename))
        print 'Wrote snapshot to: {:s}'.format(filename)

        if scale_bbox_params:
            # restore net to original state
            net.params['bbox_pred'][0].data[...] = orig_0
            net.params['bbox_pred'][1].data[...] = orig_1
        return filename

    def train_model(self, max_iters):...
def get_training_roidb(imdb):...
def filter_roidb(roidb): ...
def train_net(solver_prototxt, roidb, output_dir,
              pretrained_model=None, max_iters=40000):...
    

这里我们可以发现SolverWrapper是一个类，作者这里使用新的solverwrapper来包裹原来的solver，这样就能在原来的基础上添加部分功能。以便于控制了网络的快照过程，以用来对bounding-box 回归的权重进行非归一化。

该类的核心语句其实就是self.solver = caffe.SGDSolver(solver_prototxt)，这里的solver_prototxt=py-faster-rcnn/tools/../lib/roi_data_layer;

上面的SGDSolver函数里面创建类class RoIDataLayer(caffe.Layer)，该类是caffe layer的一个扩展实现，用于fast rcnn训练。

进入该层class RoIDataLayer(caffe.Layer)看看：

class RoIDataLayer(caffe.Layer):
    """Fast R-CNN data layer used for training."""

    def _shuffle_roidb_inds(self):...def _get_next_minibatch_inds(self):...
    def _get_next_minibatch(self):...def set_roidb(self, roidb):...
        def setup(self, bottom, top):
        """Setup the RoIDataLayer."""

        # parse the layer parameter string, which must be valid YAML
        layer_params = yaml.load(self.param_str_)  ##获取该层的参数 layer_params={'num_classes': 21}

        self._num_classes = layer_params['num_classes']

        self._name_to_top_map = {}

        # data blob: holds a batch of N images, each with 3 channels
        idx = 0
        top[idx].reshape(cfg.TRAIN.IMS_PER_BATCH, 3,
            max(cfg.TRAIN.SCALES), cfg.TRAIN.MAX_SIZE)
        self._name_to_top_map['data'] = idx
        idx += 1

        if cfg.TRAIN.HAS_RPN:
            top[idx].reshape(1, 3)
            self._name_to_top_map['im_info'] = idx
            idx += 1

            top[idx].reshape(1, 4)
            self._name_to_top_map['gt_boxes'] = idx
            idx += 1
        else: # not using RPN
            # rois blob: holds R regions of interest, each is a 5-tuple
            # (n, x1, y1, x2, y2) specifying an image batch index n and a
            # rectangle (x1, y1, x2, y2)
            top[idx].reshape(1, 5)
            self._name_to_top_map['rois'] = idx
            idx += 1

            # labels blob: R categorical labels in [0, ..., K] for K foreground
            # classes plus background
            top[idx].reshape(1)
            self._name_to_top_map['labels'] = idx
            idx += 1

            if cfg.TRAIN.BBOX_REG:
                # bbox_targets blob: R bounding-box regression targets with 4
                # targets per class
                top[idx].reshape(1, self._num_classes * 4)
                self._name_to_top_map['bbox_targets'] = idx
                idx += 1

                # bbox_inside_weights blob: At most 4 targets per roi are active;
                # thisbinary vector sepcifies the subset of active targets
                top[idx].reshape(1, self._num_classes * 4)
                self._name_to_top_map['bbox_inside_weights'] = idx
                idx += 1

                top[idx].reshape(1, self._num_classes * 4)
                self._name_to_top_map['bbox_outside_weights'] = idx
                idx += 1

        print 'RoiDataLayer: name_to_top:', self._name_to_top_map
        assert len(top) == len(self._name_to_top_map)

def forward(self, bottom, top):...
def backward(self, top, propagate_down, bottom):...    
def reshape(self, bottom, top):...
       

对于上面我们主要分析下setup函数，我们这里设置了RPN层，所以结构上有三个输入：data(1 3 600 1000)、im_info(1 3)、gt_boxes(1 4)

RoiDataLayer: name_to_top: {'gt_boxes': 2, 'data': 0, 'im_info': 1}

rpn_cls_score_rpn_cls_score_0_split：1 18 39 64   ##9个anchors  ×  2个前后景

rpn_cls_score_reshape：1 2 351 64   ##reshape成前后景两类的概率

rpn_bbox_pred：1 36 39 64    ##9个anchors  ×  4个坐标值

此时系统开始构造该roi网络结构，然后进入类class AnchorTargetLayer(caffe.Layer):

class AnchorTargetLayer(caffe.Layer):
    """
    Assign anchors to ground-truth targets. Produces anchor classification
    labels and bounding-box regression targets.
    """

    def setup(self, bottom, top):
        layer_params = yaml.load(self.param_str_)   ##获取该层参数layers_params={'feat_stride': 16}，这里是anchors的其中一个比例（8，16，32）
        anchor_scales = layer_params.get('scales', (8, 16, 32))
        self._anchors = generate_anchors(scales=np.array(anchor_scales))
        self._num_anchors = self._anchors.shape[0]   ##9
        self._feat_stride = layer_params['feat_stride']  ##16

        if DEBUG:
            print 'anchors:'
            print self._anchors
            print 'anchor shapes:'
            print np.hstack((
                self._anchors[:, 2::4] - self._anchors[:, 0::4],
                self._anchors[:, 3::4] - self._anchors[:, 1::4],
            ))
            self._counts = cfg.EPS
            self._sums = np.zeros((1, 4))
            self._squared_sums = np.zeros((1, 4))
            self._fg_sum = 0
            self._bg_sum = 0
            self._count = 0

        #设为0，则取出任何超过图像边界的proposals，只要超出一点点，都要去除
        self._allowed_border = layer_params.get('allowed_border', 0)

        height, width = bottom[0].data.shape[-2:]   ## 39，64
        if DEBUG:
            print 'AnchorTargetLayer: height', height, 'width', width

        A = self._num_anchors   ## 9
        # labels
        top[0].reshape(1, 1, A * height, width)  ##将labels输出层reshape成（1，1，9×39，64）
        # bbox_targets
        top[1].reshape(1, A * 4, height, width)  ##将bbox_targets输出层reshape成（1，9*4，39，64）
        # bbox_inside_weights
        top[2].reshape(1, A * 4, height, width)  ##将bbox_inside_weights输出层reshape成（1，9*4，39，64）
        # bbox_outside_weights
        top[3].reshape(1, A * 4, height, width)  ##将bbox_outside_weights输出层reshape成（1，9*4，39，64）

    def forward(self, bottom, top):...def backward(self, top, propagate_down, bottom):...
    def reshape(self, bottom, top)：...

我们进入self._anchors = generate_anchors(scales=np.array(anchor_scales))函数来产生anchors，进入该函数。该函数主要是产生9个anchors：

def generate_anchors(base_size=16, ratios=[0.5, 1, 2],
                     scales=2**np.arange(3, 6)):
    """
    Generate anchor (reference) windows by enumerating aspect ratios X
    scales wrt a reference (0, 0, 15, 15) window.
    """

    base_anchor = np.array([1, 1, base_size, base_size]) - 1  ##设置基础anchor，左上坐标为（1，1），右下坐标为（16，16），即宽为15
    ratio_anchors = _ratio_enum(base_anchor, ratios)
    anchors = np.vstack([_scale_enum(ratio_anchors[i, :], scales)
                         for i in xrange(ratio_anchors.shape[0])])
    return anchors

进入ratio_anchors = _ratio_enum(base_anchor, ratios)函数：

def _ratio_enum(anchor, ratios):
    """
    Enumerate a set of anchors for each aspect ratio wrt an anchor.
    """

    w, h, x_ctr, y_ctr = _whctrs(anchor)  ##返回一个anchor的宽，高， xy 中心点  （16，16，7.5，7.5）
    size = w * h   ##size=256
    size_ratios = size / ratios  ##size_ratios =[ 512.  256.  128.]，此时ratios=[0.5, 1, 2]
    ws = np.round(np.sqrt(size_ratios))  ##ws=[23. 16. 11.]
    hs = np.round(ws * ratios)   ##hs=[ 12.  16.  22.]
    anchors = _mkanchors(ws, hs, x_ctr, y_ctr)
    return anchors

def _whctrs(anchor):  ##返回一个anchor的宽，高， xy 中心点 
    """
    Return width, height, x center, and y center for an anchor (window).
    """

    w = anchor[2] - anchor[0] + 1
    h = anchor[3] - anchor[1] + 1
    x_ctr = anchor[0] + 0.5 * (w - 1)
    y_ctr = anchor[1] + 0.5 * (h - 1)
    return w, h, x_ctr, y_ctr

def _mkanchors(ws, hs, x_ctr, y_ctr):  ##根据中心点和尺度ws和hs组成的组合，来创建anchors，返回四个坐标点
    """
    Given a vector of widths (ws) and heights (hs) around a center
    (x_ctr, y_ctr), output a set of anchors (windows).
    """

    ws = ws[:, np.newaxis]
    hs = hs[:, np.newaxis]
    anchors = np.hstack((x_ctr - 0.5 * (ws - 1),
                         y_ctr - 0.5 * (hs - 1),
                         x_ctr + 0.5 * (ws - 1),
                         y_ctr + 0.5 * (hs - 1)))
    return anchors

 建造的anchors如下，行代表三个anchors，列代表2个坐标点（左上xy，右下xy）

这里得到三个anchors，然后取其中一个anchor为基础建立另三个anchors，另一个anchor为基础建立另三个anchors，这样我们加上原来的3个就得到了9个anchors。

然后我们回到self.solver = caffe.SGDSolver(solver_prototxt)，到这里为止我们建立完RoIDataLayer层、AnchorTargetLayer层。代码看上面的class SolverWrapper(object)。

继续后面的程序运行，接着加载预训练模型ZF.v2.caffemodel和加载参数文件stage1_rpn_solver60k80k.pt，接着进入set_roidb(self, roidb)函数。该函数主要对roidb进行顺序打乱。

    def set_roidb(self, roidb):
        """Set the roidb to be used by this layer during training."""
        self._roidb = roidb
        self._shuffle_roidb_inds()  ##对水平图+垂直图进行随机打乱
        if cfg.TRAIN.USE_PREFETCH:  ##跳过
            self._blob_queue = Queue(10)
            self._prefetch_process = BlobFetcher(self._blob_queue,
                                                 self._roidb,
                                                 self._num_classes)
            self._prefetch_process.start()
            # Terminate the child process when the parent exists
            def cleanup():
                print 'Terminating BlobFetcher'
                self._prefetch_process.terminate()
                self._prefetch_process.join()
            import atexit
            atexit.register(cleanup)

 此时看看sloverwrapper包裹函数的变量：

然后我们进入函数model_paths = sw.train_model(max_iters)，从而进入训练。下面的函数大概是每20次显示一次，然后每10000次保存一次快照。

    def train_model(self, max_iters):  ##80000
        """Network training loop."""
        last_snapshot_iter = -1
        timer = Timer()
        model_paths = []
        while self.solver.iter < max_iters:
            # Make one SGD update
            timer.tic()
            self.solver.step(1)
            timer.toc()
            if self.solver.iter % (10 * self.solver_param.display) == 0:   ##10000 % 200，具体是每20次显示一次
                print 'speed: {:.3f}s / iter'.format(timer.average_time)

            if self.solver.iter % cfg.TRAIN.SNAPSHOT_ITERS == 0:  ##10000 % 10000 ，具体是每10000次保存一次快照
                last_snapshot_iter = self.solver.iter
                model_paths.append(self.snapshot())

        if last_snapshot_iter != self.solver.iter:
            model_paths.append(self.snapshot())
        return model_paths

至此，我们已经对第一阶段的训练完成。总的网络图为下图。总览一下，其实就是输入三个[data(pascal_voc格式)、im_info（图片维度）、gt_boxes]; 输出就是分类和回归得分。这里主要目的是获取训练完的模型 zf_rpn_stage1_iter_80000.caffemodel，以便后面的使用。

Stage 1 RPN, init from ImageNet model：

 接着我们进入第二阶段的训练。先回到train_faster_rcnn_alt_opt.py函数:

    print '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'
    print 'Stage 1 RPN, generate proposals'
    print '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'

    mp_kwargs = dict(
            queue=mp_queue,
            imdb_name=args.imdb_name,
            rpn_model_path=str(rpn_stage1_out['model_path']),
            cfg=cfg,
            rpn_test_prototxt=rpn_test_prototxt)
    p = mp.Process(target=rpn_generate, kwargs=mp_kwargs)
    p.start()
    rpn_stage1_out['proposal_path'] = mp_queue.get()['proposal_path']
    p.join()

进入子进程rpn_generate()函数：

def rpn_generate(queue=None, imdb_name=None, rpn_model_path=None, cfg=None,
                 rpn_test_prototxt=None):
    """Use a trained RPN to generate proposals.
    """

    cfg.TEST.RPN_PRE_NMS_TOP_N = -1     # no pre NMS filtering
    cfg.TEST.RPN_POST_NMS_TOP_N = 2000  # NMS后保留2000个框
    print 'RPN model: {}'.format(rpn_model_path)
    print('Using config:')
    pprint.pprint(cfg)  ##cfg除了上面改过的两项，其他都不变

    import caffe
    _init_caffe(cfg)

    # NOTE: the matlab implementation computes proposals on flipped images, too.
    # We compute them on the image once and then flip the already computed
    # proposals. This might cause a minor loss in mAP (less proposal jittering).
    imdb = get_imdb(imdb_name)  ##重新获取imdb数据
    print 'Loaded dataset `{:s}` for proposal generation'.format(imdb.name)

    # Load RPN and configure output directory
    rpn_net = caffe.Net(rpn_test_prototxt, rpn_model_path, caffe.TEST)  ##加载rpn_test_prototxt=
                        ##'py-faster-rcnn/models/pascal_voc/ZF/faster_rcnn_alt_opt/rpn_test.pt'
                        ##第一阶段的训练完毕的rpn_model_path=
                           ##py-faster-rcnn/output/faster_rcnn_alt_opt/voc_2007_trainval/zf_rpn_stage1_iter_80000.caffemodel

    output_dir = get_output_dir(imdb)  ##'py-faster-rcnn/output/faster_rcnn_alt_opt/voc_2007_trainval'
    print 'Output will be saved to `{:s}`'.format(output_dir)
    # Generate proposals on the imdb
    rpn_proposals = imdb_proposals(rpn_net, imdb)
    # Write proposals to disk and send the proposal file path through the
    # multiprocessing queue
    rpn_net_name = os.path.splitext(os.path.basename(rpn_model_path))[0]
    rpn_proposals_path = os.path.join(
        output_dir, rpn_net_name + '_proposals.pkl')
    with open(rpn_proposals_path, 'wb') as f:
        cPickle.dump(rpn_proposals, f, cPickle.HIGHEST_PROTOCOL)
    print 'Wrote RPN proposals to {}'.format(rpn_proposals_path)
    queue.put({'proposal_path': rpn_proposals_path})

进入rpn_net = caffe.Net()函数，该函数主要是创建caffe.Net，里面首先根据pt文件创建网络，网络最后一层是创建层ProposalLayer，进入该类的创建：

class ProposalLayer(caffe.Layer):
    """
    Outputs object detection proposals by applying estimated bounding-box
    transformations to a set of regular boxes (called "anchors").
    """

    def setup(self, bottom, top):
        # parse the layer parameter string, which must be valid YAML
        layer_params = yaml.load(self.param_str_)  ##加载该层参数layer_params={'feat_stride': 16}，其实也即16的scale

        self._feat_stride = layer_params['feat_stride']  ##16
        anchor_scales = layer_params.get('scales', (8, 16, 32))  ##（8，16，32）
        self._anchors = generate_anchors(scales=np.array(anchor_scales))
        self._num_anchors = self._anchors.shape[0]  ## 9

        if DEBUG:
            print 'feat_stride: {}'.format(self._feat_stride)
            print 'anchors:'
            print self._anchors

        # rois blob: holds R regions of interest, each is a 5-tuple
        # (n, x1, y1, x2, y2) specifying an image batch index n and a
        # rectangle (x1, y1, x2, y2)
        top[0].reshape(1, 5)  ##将rois blob输出层reshape成（1，5），列为（图片index，4个坐标值）

        # scores blob: holds scores for R regions of interest
        if len(top) > 1:  ##如果存在两个输出层，则将 scores blob 输出层reshape 成（1，1，1，1）？？？
            top[1].reshape(1, 1, 1, 1)

    def forward(self, bottom, top):...def backward(self, top, propagate_down, bottom):...
    def reshape(self, bottom, top):...
       

接着我们进入rpn_proposals = imdb_proposals(rpn_net, imdb)函数，该函数作用是读取所有图片并返回每张图片的imdb_boxes[i], scores：

def imdb_proposals(net, imdb):
    """Generate RPN proposals on all images in an imdb."""

    _t = Timer()
    imdb_boxes = [[] for _ in xrange(imdb.num_images)]
    for i in xrange(imdb.num_images):
        im = cv2.imread(imdb.image_path_at(i))
        _t.tic()
        imdb_boxes[i], scores = im_proposals(net, im)
        _t.toc()
        print 'im_proposals: {:d}/{:d} {:.3f}s' 
              .format(i + 1, imdb.num_images, _t.average_time)
        if 0:
            dets = np.hstack((imdb_boxes[i], scores))
            # from IPython import embed; embed()
            _vis_proposals(im, dets[:3, :], thresh=0.9)
            plt.show()

    return imdb_boxes

接着我们进入imdb_boxes[i], scores = im_proposals(net, im)函数：

def im_proposals(net, im):
    """Generate RPN proposals on a single image."""
    blobs = {}
    blobs['data'], blobs['im_info'] = _get_image_blob(im)
    net.blobs['data'].reshape(*(blobs['data'].shape))
    net.blobs['im_info'].reshape(*(blobs['im_info'].shape))
    blobs_out = net.forward(
            data=blobs['data'].astype(np.float32, copy=False),
            im_info=blobs['im_info'].astype(np.float32, copy=False))

    scale = blobs['im_info'][0, 2]
    boxes = blobs_out['rois'][:, 1:].copy() / scale
    scores = blobs_out['scores'].copy()
    return boxes, scores

继续进入blobs['data'], blobs['im_info'] = _get_image_blob(im)函数：

def _get_image_blob(im):
    """Converts an image into a network input.

    Arguments:
        im (ndarray): a color image in BGR order

    Returns:
        blob (ndarray): a data blob holding an image pyramid
        im_scale_factors (list): list of image scales (relative to im) used
            in the image pyramid
    """
    im_orig = im.astype(np.float32, copy=True)
    im_orig -= cfg.PIXEL_MEANS

    im_shape = im_orig.shape   ##（375，500，3）
    im_size_min = np.min(im_shape[0:2])
    im_size_max = np.max(im_shape[0:2])

    processed_ims = []

    assert len(cfg.TEST.SCALES) == 1   
    target_size = cfg.TEST.SCALES[0]

    im_scale = float(target_size) / float(im_size_min)   ##下面主要意思是图的最大边不能超过1000，目标是宽或高为600，另一个高或宽低于1000就行
    # Prevent the biggest axis from being more than MAX_SIZE
    if np.round(im_scale * im_size_max) > cfg.TEST.MAX_SIZE:
        im_scale = float(cfg.TEST.MAX_SIZE) / float(im_size_max)
    im = cv2.resize(im_orig, None, None, fx=im_scale, fy=im_scale,
                    interpolation=cv2.INTER_LINEAR)    ##对原始图片进行扩充，这里采用线性插值的方法  
    im_info = np.hstack((im.shape[:2], im_scale))[np.newaxis, :]  ##[[ 600.   800.     1.6]]
    processed_ims.append(im)

    # Create a blob to hold the input images
    blob = im_list_to_blob(processed_ims)  ##转化图片成blob格式

    return blob, im_info  ##返回该图片的blob格式数据和im_info图片信息，比如（600，800，3）

继续进入blob = im_list_to_blob(processed_ims)函数：

def im_list_to_blob(ims):
    """Convert a list of images into a network input.

    Assumes images are already prepared (means subtracted, BGR order, ...).
    """
    max_shape = np.array([im.shape for im in ims]).max(axis=0)
    num_images = len(ims)   ## 1张图
    blob = np.zeros((num_images, max_shape[0], max_shape[1], 3),
                    dtype=np.float32)
    for i in xrange(num_images):  ##将所有图片转换成caffe中blob的格式
        im = ims[i]
        blob[i, 0:im.shape[0], 0:im.shape[1], :] = im   
    # Move channels (axis 3) to axis 1
    # Axis order will become: (batch elem, channel, height, width)
    channel_swap = (0, 3, 1, 2)     ##交换axis=3和axis=1 列
    blob = blob.transpose(channel_swap)  ##最终变为(batch elem, channel, height, width)，注意这里的高和宽放到了后面，这是caffe中blob的格式
    return blob

回到im_proposals(net, im)函数，现在我们获得一张图片的blobs['data']和blobs['im_info']，继续往下运行。

def im_proposals(net, im):
    """Generate RPN proposals on a single image."""
    blobs = {}
    blobs['data'], blobs['im_info'] = _get_image_blob(im)
    net.blobs['data'].reshape(*(blobs['data'].shape))  ##将第一阶段的网络模型更改下输入结构，改成该图片的维度结构，比如（1，3，600，800）
    net.blobs['im_info'].reshape(*(blobs['im_info'].shape))  ##（1，3）
    blobs_out = net.forward(
            data=blobs['data'].astype(np.float32, copy=False),
            im_info=blobs['im_info'].astype(np.float32, copy=False))

    scale = blobs['im_info'][0, 2]  ##1.6
    boxes = blobs_out['rois'][:, 1:].copy() / scale
    scores = blobs_out['scores'].copy()
    return boxes, scores  ##返回2000个boxes的坐标和得分

 接着进入net.forward()函数，执行该网络的前向传播函数。在终端的显示是im_proposals: 1/5011 2476.064s，即有5011张图片，每张返回2000个propals和boxes。

这些proposals保存在文件'output/faster_rcnn_alt_opt/voc_2007_trainval/zf_rpn_stage1_iter_80000_proposals.pkl' 中，留着等下面的训练。最后把该路径通过进程传给下一个进程。总的网络图：

Stage 1 RPN, generate proposal

终于来到最后一个阶段了。先看看主代码：

    print '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'
    print 'Stage 1 Fast R-CNN using RPN proposals, init from ImageNet model'
    print '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'

    cfg.TRAIN.SNAPSHOT_INFIX = 'stage1'
    mp_kwargs = dict(
            queue=mp_queue,
            imdb_name=args.imdb_name,
            init_model=args.pretrained_model,
            solver=solvers[1],
            max_iters=max_iters[1],
            cfg=cfg,
            rpn_file=rpn_stage1_out['proposal_path'])
    p = mp.Process(target=train_fast_rcnn, kwargs=mp_kwargs)
    p.start()
    fast_rcnn_stage1_out = mp_queue.get()
    p.join()

我们进入子进程train_fast_rcnn()函数，其实这里的大部分内容和前面的相似，就不重复累述。

def train_fast_rcnn(queue=None, imdb_name=None, init_model=None, solver=None,
                    max_iters=None, cfg=None, rpn_file=None):
    """Train a Fast R-CNN using proposals generated by an RPN.
    """

    cfg.TRAIN.HAS_RPN = False           # 这次训练不需要RPN层了
    cfg.TRAIN.PROPOSAL_METHOD = 'rpn'   # 使用之前的propasals来训练
    cfg.TRAIN.IMS_PER_BATCH = 2
    print 'Init model: {}'.format(init_model)
    print 'RPN proposals: {}'.format(rpn_file)
    print('Using config:')
    pprint.pprint(cfg)

    import caffe
    _init_caffe(cfg)

    roidb, imdb = get_roidb(imdb_name, rpn_file=rpn_file)  ##这里rpn_file设置成rpn_roidb，因此不是调用gt_roidb，而是调用rpn_roidb函数
    output_dir = get_output_dir(imdb)
    print 'Output will be saved to `{:s}`'.format(output_dir)
    # Train Fast R-CNN
    model_paths = train_net(solver, roidb, output_dir,
                            pretrained_model=init_model,
                            max_iters=max_iters)
    # Cleanup all but the final model
    for i in model_paths[:-1]:
        os.remove(i)
    fast_rcnn_model_path = model_paths[-1]
    # Send Fast R-CNN model path over the multiprocessing queue
    queue.put({'model_path': fast_rcnn_model_path})

这里首先设置了训练使用的rpn_roidb方法（RPN用的是gt_roidb方法）,由于这时候的cfg.TRAIN.PROPOSAL_METHOD变成了rpn_roidb，所以相应的使用的get_roidb()也相应地改变，此时使用rpn_roidb（）方法，进入该函数： 

    def rpn_roidb(self):
        if int(self._year) == 2007 or self._image_set != 'test':
            gt_roidb = self.gt_roidb()  ##获得gt_roidb
            rpn_roidb = self._load_rpn_roidb(gt_roidb)  ##获得rpn_roidb
            roidb = imdb.merge_roidbs(gt_roidb, rpn_roidb)   ##融合gt_roidb和rpn_roidb变为最终输入使用到的roidb
        else:  
            roidb = self._load_rpn_roidb(None)

        return roidb

进入函数rpn_roidb = self._load_rpn_roidb(gt_roidb)：

    def _load_rpn_roidb(self, gt_roidb):
        filename = self.config['rpn_file']  ##第二阶段训练保存下来的pkl文件，也即propasals
        print 'loading {}'.format(filename)
        assert os.path.exists(filename), 
               'rpn data not found at: {}'.format(filename)
        with open(filename, 'rb') as f:
            box_list = cPickle.load(f)
        return self.create_roidb_from_box_list(box_list, gt_roidb)

接着进入self.create_roidb_from_box_list(box_list, gt_roidb)：

    def create_roidb_from_box_list(self, box_list, gt_roidb):  ##boxeslist就是第二阶段的保存下来的pkl文件读取出来的东西
        assert len(box_list) == self.num_images, 
                'Number of boxes must match number of ground-truth images'
        roidb = []
        for i in xrange(self.num_images):
            boxes = box_list[i]
            num_boxes = boxes.shape[0]   ##一副图有多少个boxes
            overlaps = np.zeros((num_boxes, self.num_classes), dtype=np.float32)

            if gt_roidb is not None and gt_roidb[i]['boxes'].size > 0:
                gt_boxes = gt_roidb[i]['boxes']
                gt_classes = gt_roidb[i]['gt_classes']
                gt_overlaps = bbox_overlaps(boxes.astype(np.float),  ##计算gt_boxes和RPN产生的IOU
                                            gt_boxes.astype(np.float))
                argmaxes = gt_overlaps.argmax(axis=1)
                maxes = gt_overlaps.max(axis=1)
                I = np.where(maxes > 0)[0]
                overlaps[I, gt_classes[argmaxes[I]]] = maxes[I]  ##存储每个propasal和gt_boxe的最大IOU，且IOU>0

            overlaps = scipy.sparse.csr_matrix(overlaps)  ##稀疏存储
            roidb.append({
                'boxes' : boxes,
                'gt_classes' : np.zeros((num_boxes,), dtype=np.int32),
                'gt_overlaps' : overlaps,
                'flipped' : False,
                'seg_areas' : np.zeros((num_boxes,), dtype=np.float32),
            })
        return roidb

到这里为止，我们就得到了有gt和rpn产生propasal融合而成的roidb。

接着回到训练train_net函数里面，这里的运作于前面的大部分相同，但这里设置了boxes回归项：

class SolverWrapper(object):
    """A simple wrapper around Caffe's solver.
    This wrapper gives us control over he snapshotting process, which we
    use to unnormalize the learned bounding-box regression weights.
    """

    def __init__(self, solver_prototxt, roidb, output_dir,
                 pretrained_model=None):
        """Initialize the SolverWrapper."""
        self.output_dir = output_dir

        if (cfg.TRAIN.HAS_RPN and cfg.TRAIN.BBOX_REG and
            cfg.TRAIN.BBOX_NORMALIZE_TARGETS):
            # RPN can only use precomputed normalization because there are no
            # fixed statistics to compute a priori
            assert cfg.TRAIN.BBOX_NORMALIZE_TARGETS_PRECOMPUTED

        if cfg.TRAIN.BBOX_REG:   ##注意这里训练RPN的时候设置为false，训练faster-rcnn的时候为true
            print 'Computing bounding-box regression targets...'
            self.bbox_means, self.bbox_stds = 
                    rdl_roidb.add_bbox_regression_targets(roidb)
            print 'done'

        self.solver = caffe.SGDSolver(solver_prototxt)
        if pretrained_model is not None:
            print ('Loading pretrained model '
                   'weights from {:s}').format(pretrained_model)
            self.solver.net.copy_from(pretrained_model)

        self.solver_param = caffe_pb2.SolverParameter()
        with open(solver_prototxt, 'rt') as f:
            pb2.text_format.Merge(f.read(), self.solver_param)

        self.solver.net.layers[0].set_roidb(roidb)

上图是它的类定义中的一部分，我们先来看看它的初始化函数，这里需要注意的是add_bbox_regression_targets()这个函数，它的作用是为RPN产生的proposal提供回归属性，该函数向roidb中再添加一个key:'bbox_targets'，它的格式如：targets[][5]:第一个元素是label，后面四个元素就是论文中谈及的tx,ty,tw,th；好的，我们进入这个函数：

def add_bbox_regression_targets(roidb):  ##添加回归用到的信息
    """Add information needed to train bounding-box regressors."""
    assert len(roidb) > 0
    assert 'max_classes' in roidb[0], 'Did you call prepare_roidb first?'

    num_images = len(roidb)
    # Infer number of classes from the number of columns in gt_overlaps
    num_classes = roidb[0]['gt_overlaps'].shape[1]  ##gt个数
    for im_i in xrange(num_images):
        rois = roidb[im_i]['boxes']
        max_overlaps = roidb[im_i]['max_overlaps']
        max_classes = roidb[im_i]['max_classes']
        roidb[im_i]['bbox_targets'] = 
                _compute_targets(rois, max_overlaps, max_classes)

    if cfg.TRAIN.BBOX_NORMALIZE_TARGETS_PRECOMPUTED:
        # Use fixed / precomputed "means" and "stds" instead of empirical values
        means = np.tile(
                np.array(cfg.TRAIN.BBOX_NORMALIZE_MEANS), (num_classes, 1))
        stds = np.tile(
                np.array(cfg.TRAIN.BBOX_NORMALIZE_STDS), (num_classes, 1))
    else:   ##对bbox的坐标进行归一化
        # Compute values needed for means and stds
        # var(x) = E(x^2) - E(x)^2
        class_counts = np.zeros((num_classes, 1)) + cfg.EPS
        sums = np.zeros((num_classes, 4))
        squared_sums = np.zeros((num_classes, 4))
        for im_i in xrange(num_images):
            targets = roidb[im_i]['bbox_targets']
            for cls in xrange(1, num_classes):
                cls_inds = np.where(targets[:, 0] == cls)[0]
                if cls_inds.size > 0:
                    class_counts[cls] += cls_inds.size
                    sums[cls, :] += targets[cls_inds, 1:].sum(axis=0)
                    squared_sums[cls, :] += 
                            (targets[cls_inds, 1:] ** 2).sum(axis=0)

        means = sums / class_counts
        stds = np.sqrt(squared_sums / class_counts - means ** 2)

    print 'bbox target means:'
    print means
    print means[1:, :].mean(axis=0) # ignore bg class
    print 'bbox target stdevs:'
    print stds
    print stds[1:, :].mean(axis=0) # ignore bg class

    # Normalize targets
    if cfg.TRAIN.BBOX_NORMALIZE_TARGETS:  ##减去均值，再除以方差
        print "Normalizing targets"
        for im_i in xrange(num_images):
            targets = roidb[im_i]['bbox_targets']
            for cls in xrange(1, num_classes):
                cls_inds = np.where(targets[:, 0] == cls)[0]
                roidb[im_i]['bbox_targets'][cls_inds, 1:] -= means[cls, :]
                roidb[im_i]['bbox_targets'][cls_inds, 1:] /= stds[cls, :]
    else:
        print "NOT normalizing targets"

    # These values will be needed for making predictions
    # (the predicts will need to be unnormalized and uncentered)
    return means.ravel(), stds.ravel()

主要看_compute_targets()函数，它产生了回归属性，进入该函数：

def _compute_targets(rois, overlaps, labels):  ##判断rpn产生的proposals回归哪一个gt_box
    """Compute bounding-box regression targets for an image."""
    # Indices of ground-truth ROIs
    gt_inds = np.where(overlaps == 1)[0]
    if len(gt_inds) == 0:
        # Bail if the image has no ground-truth ROIs
        return np.zeros((rois.shape[0], 5), dtype=np.float32)
    # Indices of examples for which we try to make predictions
    ex_inds = np.where(overlaps >= cfg.TRAIN.BBOX_THRESH)[0]  ##默认阈值为0.5,进行筛选

    # Get IoU overlap between each ex ROI and gt ROI
    ex_gt_overlaps = bbox_overlaps(   ##ex_gt_overlaps存储经筛选后的框和gt框的IOU值
        np.ascontiguousarray(rois[ex_inds, :], dtype=np.float),
        np.ascontiguousarray(rois[gt_inds, :], dtype=np.float))

    # Find which gt ROI each ex ROI has max overlap with:
    # this will be the ex ROI's gt target
    gt_assignment = ex_gt_overlaps.argmax(axis=1)##找满足条件的最大IOU对应的gt框
    gt_rois = rois[gt_inds[gt_assignment], :]  ##得到每一个ex_roid对应的gt_rois
    ex_rois = rois[ex_inds, :]

    targets = np.zeros((rois.shape[0], 5), dtype=np.float32)
    targets[ex_inds, 0] = labels[ex_inds]
    targets[ex_inds, 1:] = bbox_transform(ex_rois, gt_rois)  ##生成tx，ty，tw，th
    return targets

这部分主要得到rpn_roidb的坐标的均值和方差，可以用来进行坐标归一化；OK,再回到SolverWrapper类中，剩下的则是snapshot快照方法，和train_model方法，回到train_net()函数中，接着再调用train_model()方法，这些和上面的过程十分类似，不再累述。

最后阶段的图如下所示：

由图知我们设置不使用rpn层，将提取的proposals作为rois（rpn_roidb + gt_roidb）和前面VGG16（或者ZF）网络提取的最后特征（conv5_3）传入网络，计算bbox_inside_weights+bbox_outside_weights，作用与RPN一样，传入soomth_L1_loss layer，如图绿框。

这样就可以训练最后的识别softmax与最终的bounding box regression了。