【說明】：歡迎加入：faster-rcnn 交流群 238138700，本文分析faster-rcnn 訓練的python指令碼；

【debug】：我是把__main__修改了下，放在一個自己定義的函式裡面，然後呼叫debug一步步看執行效果的，讀者不妨也這樣做，可以清晰看到程式是如何執行的；

#!/usr/bin/env python

# --------------------------------------------------------
# Faster R-CNN
# Copyright (c) 2015 Microsoft
# Licensed under The MIT License [see LICENSE for details]
# Written by Ross Girshick
# --------------------------------------------------------

"""Train a Faster R-CNN network using alternating optimization.
This tool implements the alternating optimization algorithm described in our
NIPS 2015 paper ("Faster R-CNN: Towards Real-time Object Detection with Region
Proposal Networks." Shaoqing Ren, Kaiming He, Ross Girshick, Jian Sun.)
"""

import _init_paths
from fast_rcnn.train import get_training_roidb, train_net
from fast_rcnn.config import cfg, cfg_from_file, cfg_from_list, get_output_dir
from datasets.factory import get_imdb
from rpn.generate import imdb_proposals
import argparse
import pprint
import numpy as np
import sys, os
import multiprocessing as mp
import cPickle
import shutil

def parse_args():
    """
    Parse input arguments
    """
    parser = argparse.ArgumentParser(description='Train a Faster R-CNN network')
    parser.add_argument('--gpu', dest='gpu_id',
                        help='GPU device id to use [0]',
                        default=0, type=int)
    parser.add_argument('--net_name', dest='net_name',
                        help='network name (e.g., "ZF")',
                        default='ZF', type=str) #qyy
    parser.add_argument('--weights', dest='pretrained_model',
                        help='initialize with pretrained model weights',
                        default='./data/imagenet_models/ZF.v2.caffemodel', type=str) #qyy
    parser.add_argument('--cfg', dest='cfg_file',
                        help='optional config file',
                        default='./experiments/cfgs/faster_rcnn_alt_opt.yml', type=str)# qyy
    parser.add_argument('--imdb', dest='imdb_name',
                        help='dataset to train on',
                        default='voc_2007_trainval', type=str)
    parser.add_argument('--set', dest='set_cfgs',
                        help='set config keys', default=None,
                        nargs=argparse.REMAINDER)

    if len(sys.argv) == 1:
        parser.print_help()
        sys.exit(1)

    args = parser.parse_args()
    return args

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

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

# ------------------------------------------------------------------------------
# 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.
# ------------------------------------------------------------------------------

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)

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.
    """
    #首先進來後繼續配置了一些cfg這個物件的一些引數
    # 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) #格式化輸出字串
    print('Using config:')
    pprint.pprint(cfg)

    import caffe
    _init_caffe(cfg)

    #這裡是關鍵，準備資料集，我們在debug的時候可以發現，imdb是一個類，而roidb是該類的一個成員
    roidb, imdb = get_roidb(imdb_name)#我們進入這個資料準備的函式看看
    print 'roidb len: {}'.format(len(roidb))
    output_dir = get_output_dir(imdb)
    print 'Output will be saved to `{:s}`'.format(output_dir)
    #這個solver傳入的是./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) #進入train_net函式，看訓練如何實現的
    # Cleanup all but the final model
    for i in model_paths[:-1]: #把訓練過程中儲存的中間結果的模型刪掉，只返回最終模型的結果
        os.remove(i)
    rpn_model_path = model_paths[-1]
    # Send final model path through the multiprocessing queue
    queue.put({'model_path': rpn_model_path}) #通過佇列將該程序執行的模型結果的路徑返回

#這個函式利用rpn網路來生成proposals的
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})
#這個函式是用來訓練檢測網路的
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
    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})

if __name__ == '__main__': #建議讀者除錯這個函式，進去看看每個變數是怎麼回事
    args = parse_args() #解析系統傳入的argv引數，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) #如果輸入了這個引數，就呼叫該函式，應該是做某些配置操作
    if args.set_cfgs is not None:
        cfg_from_list(args.set_cfgs)
    cfg.GPU_ID = args.gpu_id # cfg是一個詞典（edict）資料結構，從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. #這裡說的要使用多程序,因為在pycaffe中當某個網路被discard後，不能可靠保證釋放記憶體資源；程序關閉後資源自然會釋放
    # --------------------------------------------------------------------------

    # queue for communicated results between processes
    mp_queue = mp.Queue() #mp指的是multiprocessing庫，所以這裡返回了一個用於多程序通訊的佇列物件
    # solves, iters, etc. for each training stage
    solvers, max_iters, rpn_test_prototxt = get_solvers(args.net_name) #這裡返回了solvers的路徑，maxiters的值，rpn_test_prototxt的路徑

    print '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'
    print 'Stage 1 RPN, init from ImageNet model'
    print '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'
    # 這一步是用imageNet的模型初始化，然後訓練rpn網路（整個訓練過程可以參考作者的論文）
    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) # 這裡把該階段需要的引數都放到這裡來了，即函式train_rpn的輸入引數
    p = mp.Process(target=train_rpn, kwargs=mp_kwargs) # 顯然，這裡準備啟動一個新程序，呼叫函式train_rpn，傳入引數kwargs，所以我們進入train_rpn函式看看是如何工作的
    p.start()
    rpn_stage1_out = mp_queue.get()
    p.join()

    print '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'
    print 'Stage 1 RPN, generate proposals'
    print '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'
    # 這一步是利用上一步訓練好的rpn網路，產生proposals供後面使用
    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 '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'
    #這一步是再次用imageNet的模型初始化前5層卷積層，然後用上一步得到的proposals訓練檢測網路
    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 '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'
    #這一步固定上一步訓練好的前五層卷積層，再次訓練RPN，這樣就得到最終RPN網路的引數了
    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()
    p.join()

    print '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'
    print 'Stage 2 RPN, generate proposals'
    print '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'
    #利用最終確定的RPN網路產生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 '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'
    #利用上一步產生的proposals，訓練出最終的檢測網路
    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)
 

分析上面訓練呼叫的函式train_net，該函式位於：./lib/fast_rcnn/train.py檔案中

# --------------------------------------------------------
# Fast R-CNN
# Copyright (c) 2015 Microsoft
# Licensed under The MIT License [see LICENSE for details]
# Written by Ross Girshick
# --------------------------------------------------------

"""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.
    """

    #這就是SolverWrapper的建構函式
    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'

        # 這句話呼叫了caffe的SGDSolver，這個是caffe在C++中實現的一個類，用來進行隨機梯度下降優化，該類根據solver_prototxt中定義的網路和求解引數，完成網路
               # 初始化，然後返回類SGDSolver的一個例項，關於該類的設計可以參考caffe的網站：http://caffe.berkeleyvision.org/doxygen/classcaffe_1_1SGDSolver.html
        # 然後作者把該物件作為SolverWrapper的一個成員，命名為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)#這句話完成對網路的初始化

        self.solver_param = caffe_pb2.SolverParameter()
        with open(solver_prototxt, 'rt') as f:
            pb2.text_format.Merge(f.read(), self.solver_param)#這句話應該是設定了self.solver_param這個成員的引數

        self.solver.net.layers[0].set_roidb(roidb)#這句話傳入訓練的資料：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()#作者測量一次迭代花的時間
            self.solver.step(1)# 做一次梯度下降優化
            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."""

    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]
        # Select background RoIs as those within [BG_THRESH_LO, BG_THRESH_HI)
        bg_inds = np.where((overlaps < cfg.TRAIN.BG_THRESH_HI) &
                           (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
        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

# 該函式先是呼叫了該檔案中定義的類SolverWrapper的建構函式，返回了該類的一個物件sw，然後呼叫了sw的train_model方法進行訓練
# 傳入引數，搭建caffe的網路結構，用預訓練模型完成初始化，這些過程就是在該建構函式中實現的，進入這個建構函式看看
def train_net(solver_prototxt, roidb, output_dir,
              pretrained_model=None, max_iters=40000):
    """Train a Fast R-CNN network."""

    roidb = filter_roidb(roidb)#刪除一些不滿足要求的輸入圖片
    sw = SolverWrapper(solver_prototxt, roidb, output_dir,
                       pretrained_model=pretrained_model)#呼叫建構函式

    print 'Solving...'
    model_paths = sw.train_model(max_iters)#開始訓練模型
    print 'done solving'
    return model_paths
 

作者：香蕉麥樂迪--sloanqin--覃元元