二、训练
接下来回到train.py第160行，通过调用sw.train_model方法进行训练：
1 def train_model(self, max_iters): 2 """Network training loop.""" 3 last_snapshot_iter = -1 4 timer = Timer() 5 model_paths = [] 6 while self.solver.iter < max_iters: 7 # Make one SGD update 8 timer.tic() 9 self.solver.step(1) 10 timer.toc() 11 if self.solver.iter % (10 * self.solver_param.display) == 0: 12 print 'speed: {:.3f}s / iter'.format(timer.average_time) 13 14 if self.solver.iter % cfg.TRAIN.SNAPSHOT_ITERS == 0: 15 last_snapshot_iter = self.solver.iter 16 model_paths.append(self.snapshot()) 17 18 if last_snapshot_iter != self.solver.iter: 19 model_paths.append(self.snapshot()) 20 return model_paths
方法中的self.solver.step(1)即是网络进行一次前向传播和反向传播。前向传播时，数据流会从第一层流动到最后一层，最后计算出loss，然后loss相对于各层输入的梯度会从最后一层计算回第一层。下面逐层来介绍faster-rcnn算法的运行过程。
2.1、input-data layer
第一层是由python代码构成的，其prototxt描述为：
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': 2" } }
从中可以看出，input-data层有三个输出：data、im_info、gt_boxes，其实现为RoIDataLayer类。这一层对数据的预处理操作为：对图片进行长宽等比例缩放，使短边缩放至600；如果缩放后，长边的长度大于1000，则以长边为基准，将长边缩放至1000，短边作相应的等比例缩放。这一层的3个输出分别为：
1、data：1, 3, h, w(一个batch只支持输入一张图)
2、im_info: im_info[0], im_info[1], im_info[2]分别为h, w, target_size/im_origin_size(缩放比例)
3、gt_boxes: (x1, y1, x2, y2, cls)
预处理部分涉及到的函数有_get_next_minibatch，get_minibatch，_get_image_blob，prep_im_for_blob，im_list_to_blob。
网络在构造过程中（即self.solver = caffe.SGDSolver(solver_prototxt)）会调用该类的setup方法：
1 __C.TRAIN.IMS_PER_BATCH = 1 2 __C.TRAIN.SCALES = [600] 3 __C.TRAIN.MAX_SIZE = 1000 4 __C.TRAIN.HAS_RPN = True 5 __C.TRAIN.BBOX_REG = True 6 7 def setup(self, bottom, top): 8 """Setup the RoIDataLayer.""" 9 10 # parse the layer parameter string, which must be valid YAML 11 layer_params = yaml.load(self.param_str_) 12 13 self._num_classes = layer_params['num_classes'] 14 15 self._name_to_top_map = {} 16 17 # data blob: holds a batch of N images, each with 3 channels 18 idx = 0 19 top[idx].reshape(cfg.TRAIN.IMS_PER_BATCH, 3, 20 max(cfg.TRAIN.SCALES), cfg.TRAIN.MAX_SIZE) 21 self._name_to_top_map['data'] = idx 22 idx += 1 23 24 if cfg.TRAIN.HAS_RPN: 25 top[idx].reshape(1, 3) 26 self._name_to_top_map['im_info'] = idx 27 idx += 1 28 29 top[idx].reshape(1, 4) 30 self._name_to_top_map['gt_boxes'] = idx 31 idx += 1 32 else: # not using RPN 33 # rois blob: holds R regions of interest, each is a 5-tuple 34 # (n, x1, y1, x2, y2) specifying an image batch index n and a 35 # rectangle (x1, y1, x2, y2) 36 top[idx].reshape(1, 5) 37 self._name_to_top_map['rois'] = idx 38 idx += 1 39 40 # labels blob: R categorical labels in [0, ..., K] for K foreground 41 # classes plus background 42 top[idx].reshape(1) 43 self._name_to_top_map['labels'] = idx 44 idx += 1 45 46 if cfg.TRAIN.BBOX_REG: 47 # bbox_targets blob: R bounding-box regression targets with 4 48 # targets per class 49 top[idx].reshape(1, self._num_classes * 4) 50 self._name_to_top_map['bbox_targets'] = idx 51 idx += 1 52 53 # bbox_inside_weights blob: At most 4 targets per roi are active; 54 # thisbinary vector sepcifies the subset of active targets 55 top[idx].reshape(1, self._num_classes * 4) 56 self._name_to_top_map['bbox_inside_weights'] = idx 57 idx += 1 58 59 top[idx].reshape(1, self._num_classes * 4) 60 self._name_to_top_map['bbox_outside_weights'] = idx 61 idx += 1 62 63 print 'RoiDataLayer: name_to_top:', self._name_to_top_map 64 assert len(top) == len(self._name_to_top_map)
主要是对输出的shape进行定义。要说明的是，在前向传播的过程中，仍然会对输出的各top的shape进行重定义，并且二者定义的shape往往都是不同的。