r-cnn学习（四）：train_faster_rcnn_alt_opt.py源码学习

论文看的云里雾里，希望通过阅读其代码来进一步了解。

参考：http://blog.csdn.net/sloanqin/article/details/51525692

首先是./tools/train_faster_rcnn_alt_opt.py，通过其main函数了解整个训练流程。

if __name__ == '__main__': #建议读者调试这个函数，进去看看每个变量是怎么回事
    args = parse_args() #解析系统传入的argv参数，解析完放到args中返回  

    print('Called with args:')
    print(args)  

    if args.cfg_file is not None:
        cfg_from_file(args.cfg_file) #如果输入了这个参数，就调用该函数，应该是做某些配置操作
    if args.set_cfgs is not None:
        cfg_from_list(args.set_cfgs)
    cfg.GPU_ID = args.gpu_id # cfg是一个词典（edict）数据结构，从faster-rcnn.config引入的  

    # --------------------------------------------------------------------------
    # Pycaffe doesn't reliably free GPU memory when instantiated nets are
    # discarded (e.g. "del net" in Python code). To work around this issue, each
    # training stage is executed in a separate process using
    # multiprocessing.Process. #这里说的要使用多进程,因为在pycaffe中当某个网络被discard后，不能可靠保证释放内存资源；进程关闭后资源自然会释放
    # --------------------------------------------------------------------------  

    # queue for communicated results between processes
    mp_queue = mp.Queue() #mp指的是multiprocessing库，所以这里返回了一个用于多进程通信的队列对象
    # solves, iters, etc. for each training stage
    solvers, max_iters, rpn_test_prototxt = get_solvers(args.net_name) #这里返回了solvers的路径，maxiters的值，rpn_test_prototxt的路径  

    print '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'
    print 'Stage 1 RPN, init from ImageNet model'
    print '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'
    # 这一步是用imageNet的模型初始化，然后训练rpn网络（整个训练过程可以参考作者的论文）
    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) # 这里把该阶段需要的参数都放到这里来了，即函数train_rpn的输入参数
    p = mp.Process(target=train_rpn, kwargs=mp_kwargs) # 显然，这里准备启动一个新进程，调用函数train_rpn，传入参数kwargs，所以我们进入train_rpn函数看看是如何工作的
    p.start()
    rpn_stage1_out = mp_queue.get()
    p.join()  

    print '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'
    print 'Stage 1 RPN, generate proposals'
    print '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'
    # 这一步是利用上一步训练好的rpn网络，产生proposals供后面使用
    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()  

    print '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'
    print 'Stage 1 Fast R-CNN using RPN proposals, init from ImageNet model'
    print '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'
    #这一步是再次用imageNet的模型初始化前5层卷积层，然后用上一步得到的proposals训练检测网络
    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()  

    print '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'
    print 'Stage 2 RPN, init from stage 1 Fast R-CNN model'
    print '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'
    #这一步固定上一步训练好的前五层卷积层，再次训练RPN，这样就得到最终RPN网络的参数了
    cfg.TRAIN.SNAPSHOT_INFIX = 'stage2'
    mp_kwargs = dict(
            queue=mp_queue,
            imdb_name=args.imdb_name,
            init_model=str(fast_rcnn_stage1_out['model_path']),
            solver=solvers[2],
            max_iters=max_iters[2],
            cfg=cfg)
    p = mp.Process(target=train_rpn, kwargs=mp_kwargs)
    p.start()
    rpn_stage2_out = mp_queue.get()#保留训练的权重
    p.join()  

    print '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'
    print 'Stage 2 RPN, generate proposals'
    print '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'
    #利用最终确定的RPN网络产生proposals
    mp_kwargs = dict(
            queue=mp_queue,
            imdb_name=args.imdb_name,
            rpn_model_path=str(rpn_stage2_out['model_path']),
            cfg=cfg,
            rpn_test_prototxt=rpn_test_prototxt)
    p = mp.Process(target=rpn_generate, kwargs=mp_kwargs)
    p.start()
    rpn_stage2_out['proposal_path'] = mp_queue.get()['proposal_path']
    p.join()  

    print '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'
    print 'Stage 2 Fast R-CNN, init from stage 2 RPN R-CNN model'
    print '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'
    #利用上一步产生的proposals，训练出最终的检测网络
    cfg.TRAIN.SNAPSHOT_INFIX = 'stage2'
    mp_kwargs = dict(
            queue=mp_queue,
            imdb_name=args.imdb_name,
            init_model=str(rpn_stage2_out['model_path']),
            solver=solvers[3],
            max_iters=max_iters[3],
            cfg=cfg,
            rpn_file=rpn_stage2_out['proposal_path'])
    p = mp.Process(target=train_fast_rcnn, kwargs=mp_kwargs)
    p.start()
    fast_rcnn_stage2_out = mp_queue.get()
    p.join()  

    # Create final model (just a copy of the last stage)
    final_path = os.path.join(
            os.path.dirname(fast_rcnn_stage2_out['model_path']),
            args.net_name + '_faster_rcnn_final.caffemodel')
    print 'cp {} -> {}'.format(
            fast_rcnn_stage2_out['model_path'], final_path)
    shutil.copy(fast_rcnn_stage2_out['model_path'], final_path)
    print 'Final model: {}'.format(final_path)  

通过上面的代码可以看出，整个迭代过程分为四步（参考论文）。其中后面两步固定共享卷积

层，只对RPN和fc层进行微调。

接着看看每一步是怎样的。

首先是train_rpn。从代码看出，这个函数的主要任务是，配置参数，准备数据集，

传入第一阶段的solver，调用train_net训练模型并将结果返回。

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.
    """
    #首先进来后继续配置了一些cfg这个对象的一些参数
    # 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) #格式化输出字符串
    print('Using config:')
    pprint.pprint(cfg)  

    import caffe
    _init_caffe(cfg)  

    #这里是关键，准备数据集，我们在debug的时候可以发现，imdb是一个类，而roidb是该类的一个成员
    roidb, imdb = get_roidb(imdb_name)#我们进入这个数据准备的函数看看
    print 'roidb len: {}'.format(len(roidb))
    output_dir = get_output_dir(imdb)
    print 'Output will be saved to `{:s}`'.format(output_dir)
    #这个solver传入的是./models/pascal_voc/ZF/faster_rcnn_alt_opt/stage1_rpn_solver60k80k.pt
    model_paths = train_net(solver, roidb, output_dir,
                            pretrained_model=init_model,
                            max_iters=max_iters) #进入train_net函数，看训练如何实现的
    # Cleanup all but the final model
    for i in model_paths[:-1]: #把训练过程中保存的中间结果的模型删掉，只返回最终模型的结果
        os.remove(i)
    rpn_model_path = model_paths[-1]
    # Send final model path through the multiprocessing queue
    queue.put({'model_path': rpn_model_path}) #通过队列将该进程运行的模型结果的路径返回  

顺着train_rpn，查看train_net函数，该函数位于：./lib/fast_rcnn/train.py文件中

调用该文件中定义的类SolverWrapper的构造函数，返回该类的一个对象sw，然后调用了sw的train_model方法进行训练，

传入参数，搭建caffe的网络结构，用预训练模型完成初始化，整个过程在构造函数中完成。

"""Train a Fast R-CNN network."""  

import caffe
from fast_rcnn.config import cfg
import roi_data_layer.roidb as rdl_roidb
from utils.timer import Timer
import numpy as np
import os  

from caffe.proto import caffe_pb2
import google.protobuf as pb2  

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.
    """  

    #这就是SolverWrapper的构造函数
    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'  

        # 这句话调用了caffe的SGDSolver，这个是caffe在C++中实现的一个类，用来进行随机梯度下降优化，该类根据solver_prototxt中定义的网络和求解参数，完成网络
        # 初始化，然后返回类SGDSolver的一个实例，关于该类的设计可以参考caffe的网站：http://caffe.berkeleyvision.org/doxygen/classcaffe_1_1SGDSolver.html
        # 然后作者把该对象作为SolverWrapper的一个成员，命名为solver
        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)#这句话完成对网络的初始化  

        self.solver_param = caffe_pb2.SolverParameter()
        with open(solver_prototxt, 'rt') as f:
            pb2.text_format.Merge(f.read(), self.solver_param)#这句话应该是设置了self.solver_param这个成员的参数  

        self.solver.net.layers[0].set_roidb(roidb)#这句话传入训练的数据：roidb  

    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):
        """Network training loop."""
        last_snapshot_iter = -1
        timer = Timer()
        model_paths = []
        while self.solver.iter < max_iters:
            # Make one SGD update
            timer.tic()#作者测量一次迭代花的时间
            self.solver.step(1)# 做一次梯度下降优化
            timer.toc()
            if self.solver.iter % (10 * self.solver_param.display) == 0:
                print 'speed: {:.3f}s / iter'.format(timer.average_time)  

            if self.solver.iter % cfg.TRAIN.SNAPSHOT_ITERS == 0:
                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  

def get_training_roidb(imdb):
    """Returns a roidb (Region of Interest database) for use in training."""
    if cfg.TRAIN.USE_FLIPPED:
        print 'Appending horizontally-flipped training examples...'
        imdb.append_flipped_images()
        print 'done'  

    print 'Preparing training data...'
    rdl_roidb.prepare_roidb(imdb)
    print 'done'  

    return imdb.roidb  

def filter_roidb(roidb):
    """Remove roidb entries that have no usable RoIs."""
    #判断是否是有效roidb
    def is_valid(entry):
        # Valid images have:
        #   (1) At least one foreground RoI OR
        #   (2) At least one background RoI
        overlaps = entry['max_overlaps']
        # find boxes with sufficient overlap
        fg_inds = np.where(overlaps >= cfg.TRAIN.FG_THRESH)[0]#大于某个阈值为前景
        # Select background RoIs as those within [BG_THRESH_LO, BG_THRESH_HI)
        bg_inds = np.where((overlaps < cfg.TRAIN.BG_THRESH_HI) & #在某两个阈值之间为背景
                           (overlaps >= cfg.TRAIN.BG_THRESH_LO))[0]
        # image is only valid if such boxes exist
        valid = len(fg_inds) > 0 or len(bg_inds) > 0#要么为前景，要么为背景，则为有效roidb
        return valid  

    num = len(roidb)
    filtered_roidb = [entry for entry in roidb if is_valid(entry)]
    num_after = len(filtered_roidb)
    print 'Filtered {} roidb entries: {} -> {}'.format(num - num_after,
                                                       num, num_after)
    return filtered_roidb  

# 该函数先是调用了该文件中定义的类SolverWrapper的构造函数，返回了该类的一个对象sw，然后调用了sw的train_model方法进行训练
# 传入参数，搭建caffe的网络结构，用预训练模型完成初始化，这些过程就是在该构造函数中实现的，进入这个构造函数看看
def train_net(solver_prototxt, roidb, output_dir,
              pretrained_model=None, max_iters=40000):
    """Train a Fast R-CNN network."""  

    roidb = filter_roidb(roidb)#删除一些不满足要求的输入图片
    sw = SolverWrapper(solver_prototxt, roidb, output_dir,
                       pretrained_model=pretrained_model)#调用构造函数  

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