二、訓練

接下來回到train.py第160行，通過呼叫sw.train_model方法進行訓練：

 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)即是網路進行一次前向傳播和反向傳播。前向傳播時，資料流會從第一層流動到最後一層，最後計算出loss，然後loss相對於各層輸入的梯度會從最後一層計算回第一層。下面逐層來介紹faster-rcnn演算法的執行過程。

2.1、input-data layer

第一層是由python程式碼構成的，其prototxt描述為：

從中可以看出，input-data層有三個輸出：data、im_info、gt_boxes。其實現為faster-rcnn/lib/roi_data_layer/layer.py中的RoIDataLayer類。網路在構造過程中（即self.solver = caffe.SGDSolver(solver_prototxt)）會呼叫該類的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)

主要是對輸出的shape進行定義（同時申請記憶體）。要說明的是，在前向傳播的過程中，仍然會對輸出的各top的shape進行重定義，並且二者定義的shape往往都是不同的。