py-faster-rcnn使用experiments/scripts/faster_rcnn_alt_opt.sh脚本启动训练过程。在faster_rcnn_alt_opt.sh脚本中，调用tools/train_faster_rcnn_alt_opt.py接口
1、parse_args() #传递参数 2、 get_roidb(imdb_name, rpn_file= None) #获取roidb与lmdb 3、get_solvers(net_name) # 获取solvers 4、_init_caffe(cfg) #根据config加载caffe对象 5、train_rpn(queue, imdb_name, init_model, solver=None, max_iters, cfg) #训练RPN网络 6、 rpn_generate(queue, imdb_name, rpn_model_path, cfg, rpn_test_prototxt) #用RPN网络生产proposal 7、 train_fast_rcnn(queue, imdb_name, init_model, solver, max_iters, cfg, rpn_file) #用RPN网络产生的proposal来训练fast_rcnn网络 8、mian() #详细的描述了四个训练步骤
imdb根据imdb_name(默认是“voc_2007_trainval)来获取；
roidb是通过get_training_roidb来获取的，roidb是imdb的成员变量，包含了训练集图片中框出的每个区域。
def get_roidb(imdb_name, rpn_file=None): imdb = get_imdb(imdb_name) print 'Loaded dataset {:s} for training'.format(imdb.name) imdb.set_proposal_method(cfg.TRAIN.PROPOSAL_METHOD) print 'Set proposal method: {:s}'.format(cfg.TRAIN.PROPOSAL_METHOD) if rpn_file is not None: imdb.config['rpn_file'] = rpn_file roidb = get_training_roidb(imdb) return roidb, imdb 123456789 123456789在models/pascal_voc/netname/faster_rcnn_alt_opt文件夹下有stage1_rpn_solver60k80k.pt等不同阶段所对应的solver文件，并定义了各个阶段最大迭代次数，这里支持的net_name有 VGG16、VGG_CNN_M_1024、ZF三种
def get\_solvers(net\_name): # Faster R-CNN Alternating Optimization n = 'faster_rcnn_alt_opt' # Solver for each training stage solvers = [[net_name, n, 'stage1_rpn_solver60k80k.pt'], [net_name, n, 'stage1_fast_rcnn_solver30k40k.pt'], [net_name, n, 'stage2_rpn_solver60k80k.pt'], [net_name, n, 'stage2_fast_rcnn_solver30k40k.pt']] solvers = [os.path.join(cfg.MODELS_DIR, *s) for s in solvers] # Iterations for each training stage max_iters = [80000, 40000, 80000, 40000] # max_iters = [100, 100, 100, 100] # Test prototxt for the RPN rpn_test_prototxt = os.path.join( cfg.MODELS_DIR, net_name, n, 'rpn_test.pt') return solvers, max_iters, rpn_test_prototxt 12345678910111213141516 12345678910111213141516初始化caffe对象，一是初始化随机种子，二是设置GPU。
def _init_caffe(cfg): """Initialize pycaffe in a training process. """ import caffe # fix the random seeds (numpy and caffe) for reproducibility np.random.seed(cfg.RNG_SEED) caffe.set_random_seed(cfg.RNG_SEED) # set up caffe caffe.set_mode_gpu() caffe.set_device(cfg.GPU_ID) 123456789101112131415 123456789101112131415训练rpn网络。首先这里初始化了一些配置；随后做了一些训练的准备工作，初始化模型，确定路径等；然后开始正式训练。
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 box regression # 用ground_truth作为proposal cfg.TRAIN.PROPOSAL_METHOD = 'gt' # 每个minibatch训练用到的Image数 cfg.TRAIN.IMS_PER_BATCH = 1 print 'Init model: {}'.format(init_model) print('Using config:') pprint.pprint(cfg) #初始化caffe import caffe _init_caffe(cfg) roidb, imdb = get_roidb(imdb_name) print 'roidb len: {}'.format(len(roidb)) output_dir = get_output_dir(imdb) print 'Output will be saved to {:s}'.format(output_dir) 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) rpn_model_path = model_paths[-1] # Send final model path through the multiprocessing queue queue.put({'model_path': rpn_model_path}) 123456789101112131415161718192021222324252627282930313233 123456789101112131415161718192021222324252627282930313233根据输入的数据与模型与prototxt产生proposal，可作为下一步的训练所用。
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 # limit top boxes after NMS print 'RPN model: {}'.format(rpn_model_path) print('Using config:') pprint.pprint(cfg) 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) 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) output_dir = get_output_dir(imdb) 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}) 12345678910111213141516171819202122232425262728293031323334 12345678910111213141516171819202122232425262728293031323334训练fast-rcnn，首先它将产生roidb的方法设置成rpn_roidb；接下来就是准备一些参数、路径等等，用于送入网络训练，最后保存模型。
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 # not generating prosals on-the-fly cfg.TRAIN.PROPOSAL_METHOD = 'rpn' # use pre-computed RPN proposals instead #每个mini-batch包含两张图片，以及他们proposal的roi区域 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) 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}) 12345678910111213141516171819202122232425262728293031323334 12345678910111213141516171819202122232425262728293031323334 if __name__ == '__main__': args = parse_args() print('Called with args:') print(args) if args.cfg_file is not None: cfg_from_file(args.cfg_file) if args.set_cfgs is not None: cfg_from_list(args.set_cfgs) cfg.GPU_ID = args.gpu_id # -------------------------------------------------------------------------- # 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. # -------------------------------------------------------------------------- # 这里不同的训练stage将通过多线程来提高效率，mp_queue是进程间用于通讯的数据结构 # queue for communicated results between processes mp_queue = mp.Queue() # 获取 solves, iters, etc. for each training stage solvers, max_iters, rpn_test_prototxt = get_solvers(args.net_name) # 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） print print 'Stage 1 RPN, init from ImageNet model' print # 第一个stage开始啦~~ 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) p.start() rpn_stage1_out = mp_queue.get() p.join() # 这一步是利用上一步训练好的rpn网络，产生proposals供后面使用 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() #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 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() #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 print print 'Stage 2 RPN, init from stage 1 Fast R-CNN model' print 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() p.join() #利用最终确定的RPN网络产生proposals print print 'Stage 2 RPN, generate proposals' print 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() #利用上一步产生的proposals，训练出最终的检测网络 print print 'Stage 2 Fast R-CNN, init from stage 2 RPN R-CNN model' print 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() #输出最终模型 # 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) 123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134 123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134