一.整體流程概述
1.輸入引數，其實輸入1個就行了（D:\tempFile\VOCdevkit），另外一個resnet權重只是為了加快訓練，如圖:

2.從VOC2007資料集中讀取資料，變成想要的資料格式
3.定義生成資料的迭代器
4.定義3個網路，一個是resnet共享卷積層，一個rpn層，一個分類器層
5.進入迭代，每次只訓練一張圖片
6.是否要進行圖片增強
7.根據特徵圖和定義框的比例，IOU等計算出y_train值，作為網路的label
8.訓練rpn層，輸出物體，和物體框的座標
9.然後再進行分類器層層的訓練

二.程式碼(關鍵部位已經給出註釋)

from __future__ import
 
 division
import random
import pprint
import sys
import time
import numpy as np
from optparse import OptionParser
import pickle

from keras import backend as K
from keras.optimizers import Adam, SGD, RMSprop
from keras.layers import Input
from keras.models import Model
from keras_frcnn import config, data_generators
from
 
 keras_frcnn import losses as losses
import keras_frcnn.roi_helpers as roi_helpers
from keras.utils import generic_utils

sys.setrecursionlimit(40000)

parser = OptionParser()

parser.add_option("-p", "--path", dest="train_path", help="Path to training data.")
parser.add_option("-o", "--parser", dest="parser"
 
, help="Parser to use. One of simple or pascal_voc",
                default="pascal_voc")
parser.add_option("-n", "--num_rois", type="int", dest="num_rois", help="Number of RoIs to process at once.", default=32)
parser.add_option("--network", dest="network", help="Base network to use. Supports vgg or resnet50.", default='resnet50')
parser.add_option("--hf", dest="horizontal_flips", help="Augment with horizontal flips in training. (Default=false).", action="store_true", default=False)
parser.add_option("--vf", dest="vertical_flips", help="Augment with vertical flips in training. (Default=false).", action="store_true", default=False)
parser.add_option("--rot", "--rot_90", dest="rot_90", help="Augment with 90 degree rotations in training. (Default=false).",
                  action="store_true", default=False)
parser.add_option("--num_epochs", type="int", dest="num_epochs", help="Number of epochs.", default=2000)
parser.add_option("--config_filename", dest="config_filename", help=
                "Location to store all the metadata related to the training (to be used when testing).",
                default="config.pickle")
parser.add_option("--output_weight_path", dest="output_weight_path", help="Output path for weights.", default='./model_frcnn.hdf5')
parser.add_option("--input_weight_path", dest="input_weight_path", help="Input path for weights. If not specified, will try to load default weights provided by keras.")

(options, args) = parser.parse_args()

if not options.train_path:   # if filename is not given
    parser.error('Error: path to training data must be specified. Pass --path to command line')

if options.parser == 'pascal_voc':
    from keras_frcnn.pascal_voc_parser import get_data
elif options.parser == 'simple':
    from keras_frcnn.simple_parser import get_data
else:
    raise ValueError("Command line option parser must be one of 'pascal_voc' or 'simple'")

# pass the settings from the command line, and persist them in the config object
C = config.Config()

C.use_horizontal_flips = bool(options.horizontal_flips)
C.use_vertical_flips = bool(options.vertical_flips)
C.rot_90 = bool(options.rot_90)

C.model_path = options.output_weight_path
C.num_rois = int(options.num_rois)

#有基於VGG和resnet兩種網路模型
if options.network == 'vgg':
    C.network = 'vgg'
    from keras_frcnn import vgg as nn
elif options.network == 'resnet50':
    from keras_frcnn import resnet as nn
    C.network = 'resnet50'
else:
    print('Not a valid model')
    raise ValueError


# check if weight path was passed via command line
if options.input_weight_path:
    C.base_net_weights = options.input_weight_path
else:
    # set the path to weights based on backend and model
    C.base_net_weights = nn.get_weight_path()

all_imgs, classes_count, class_mapping = get_data(options.train_path)
print(len(all_imgs)) #所有圖片的資訊，圖片名稱，位置等
print(len(classes_count)) #dict,類別:數量，例如'chair': 1432
print(len(class_mapping)) #dict,各個類別對應的標籤向量,0-19,例如chair:0,car:1

#再加入'背景'這個類別
if 'bg' not in classes_count:
    classes_count['bg'] = 0
    class_mapping['bg'] = len(class_mapping)
C.class_mapping = class_mapping

# 將class_mapping中的key和value對調
inv_map = {v: k for k, v in class_mapping.items()}

print('Training images per class:')
pprint.pprint(classes_count)
print('Num classes (including bg) = {}'.format(len(classes_count)))

config_output_filename = options.config_filename

with open(config_output_filename, 'wb') as config_f:
    pickle.dump(C,config_f)
    print('Config has been written to {}, and can be loaded when testing to ensure correct results'.format(config_output_filename))

# shuffle資料
random.shuffle(all_imgs)

num_imgs = len(all_imgs)
# 將all_imgs分為訓練集和測試集
train_imgs = [s for s in all_imgs if s['imageset'] == 'trainval']
val_imgs = [s for s in all_imgs if s['imageset'] == 'test']

print('Num train samples {}'.format(len(train_imgs)))
print('Num val samples {}'.format(len(val_imgs)))
# 生成anchor
data_gen_train = data_generators.get_anchor_gt(train_imgs, classes_count, C, nn.get_img_output_length, K.image_dim_ordering(), mode='train')
# data_gen_train = data_generators.get_anchor_gt(train_imgs, classes_count, C, nn.get_img_output_length, K.image_dim_ordering(), mode='train')
data_gen_val = data_generators.get_anchor_gt(val_imgs, classes_count, C, nn.get_img_output_length,K.image_dim_ordering(), mode='val')

#檢視後端是th還是tf，糾正輸入方式
if K.image_dim_ordering() == 'th':
    input_shape_img = (3, None, None)
else:
    input_shape_img = (None, None, 3)

img_input = Input(shape=input_shape_img)
roi_input = Input(shape=(None, 4))

# define the base network (resnet here, can be VGG, Inception, etc)
#定義nn的輸入層，還有faster rcnn共享卷積層
shared_layers = nn.nn_base(img_input, trainable=True)
print("shared_layers",shared_layers.shape)

# define the RPN, built on the base layers
#獲取anchor的個數，3重基準大小快，3種比例框，3*3=9
num_anchors = len(C.anchor_box_scales) * len(C.anchor_box_ratios)
#定義rpn層，return [x_class, x_regr, base_layers]
rpn = nn.rpn(shared_layers, num_anchors)

classifier = nn.classifier(shared_layers, roi_input, C.num_rois, nb_classes=len(classes_count), trainable=True)

#定義rpn模型的輸入和輸出一個框2分類（最後使用的sigmod而不是softmax）和框的迴歸
model_rpn = Model(img_input, rpn[:2])
#定義classifier的輸入和輸出
model_classifier = Model([img_input, roi_input], classifier)

# this is a model that holds both the RPN and the classifier, used to load/save weights for the models
model_all = Model([img_input, roi_input], rpn[:2] + classifier)

try:
    print('loading weights from {}'.format(C.base_net_weights))
    model_rpn.load_weights(C.base_net_weights, by_name=True)
    model_classifier.load_weights(C.base_net_weights, by_name=True)
except:
    print('Could not load pretrained model weights. Weights can be found in the keras application folder \
        https://github.com/fchollet/keras/tree/master/keras/applications')

optimizer = Adam(lr=1e-5)
optimizer_classifier = Adam(lr=1e-5)
model_rpn.compile(optimizer=optimizer, loss=[losses.rpn_loss_cls(num_anchors), losses.rpn_loss_regr(num_anchors)])
model_classifier.compile(optimizer=optimizer_classifier, loss=[losses.class_loss_cls, losses.class_loss_regr(len(classes_count)-1)], metrics={'dense_class_{}'.format(len(classes_count)): 'accuracy'})
model_all.compile(optimizer='sgd', loss='mae')

epoch_length = 1000
num_epochs = int(options.num_epochs)
iter_num = 0

losses = np.zeros((epoch_length, 5))
rpn_accuracy_rpn_monitor = []
rpn_accuracy_for_epoch = []
start_time = time.time()

best_loss = np.Inf

class_mapping_inv = {v: k for k, v in class_mapping.items()}
print('Starting training')

vis = True

for epoch_num in range(num_epochs):

    progbar = generic_utils.Progbar(epoch_length)
    print('Epoch {}/{}'.format(epoch_num + 1, num_epochs))

    while True:
        try:

            if len(rpn_accuracy_rpn_monitor) == epoch_length and C.verbose:
                mean_overlapping_bboxes = float(sum(rpn_accuracy_rpn_monitor))/len(rpn_accuracy_rpn_monitor)
                rpn_accuracy_rpn_monitor = []
                print('Average number of overlapping bounding boxes from RPN = {} for {} previous iterations'.format(mean_overlapping_bboxes, epoch_length))
                if mean_overlapping_bboxes == 0:
                    print('RPN is not producing bounding boxes that overlap the ground truth boxes. Check RPN settings or keep training.')
            print("生成data_gen_train")
            #X為經過最小邊600的比例變換的原始影象，Y為[所有框位置的和類別(正例還是反例)，所有框的前36層為位置和後36層（框和gt的比值）]，img_data增強影象後的影象資訊
            #那麼RPN的reg輸出也是比值
            X, Y, img_data = next(data_gen_train)
            print(X.shape,Y[0].shape,Y[1].shape)

            loss_rpn = model_rpn.train_on_batch(X, Y)
            print("loss_rpn",len(loss_rpn))
            print("loss_rpn0",loss_rpn[0])
            print("loss_rpn1",loss_rpn[1])
            print("loss_rpn2",loss_rpn[2])

            P_rpn = model_rpn.predict_on_batch(X)
#           print("P_rpn_cls",P_rpn[0].reshape((P_rpn[0].shape[1],P_rpn[0].shape[2],P_rpn[0].shape[3]))[:,:,0])
            print("P_rpn_cls",P_rpn[0].shape)
            print("P_rpn_reg",P_rpn[1].shape)

            #獲得最終選中的框
            R = roi_helpers.rpn_to_roi(P_rpn[0], P_rpn[1], C, K.image_dim_ordering(), use_regr=True, overlap_thresh=0.7, max_boxes=300)

            # note: calc_iou converts from (x1,y1,x2,y2) to (x,y,w,h) format
            #再對迴歸出來的框進行一次iou的計算，再一次過濾，只保留bg框和物體框
            #X2 from (x1,y1,x2,y2) to (x,y,w,h)
            #Y1為每個框對應類別標籤，one-host編碼
            #Y2為每個框和gt的比值,(x,x,160)，前80表示框是否正確，後80為20個類別可能的框
            X2, Y1, Y2, IouS = roi_helpers.calc_iou(R, img_data, C, class_mapping)
            print("X2",X2.shape)
#           print("X2_0",X2[0,0,:])
#           print("X2_1",X2[0,1,:])
            print("Y1",Y1.shape)
            print("Y2",Y2.shape)

            if X2 is None:
                rpn_accuracy_rpn_monitor.append(0)
                rpn_accuracy_for_epoch.append(0)
                continue
            #選出正例還是反例的index，背景的為反例，物體為正例
            neg_samples = np.where(Y1[0, :, -1] == 1)
            pos_samples = np.where(Y1[0, :, -1] == 0)
            print("neg_samples",len(neg_samples[0]))
            print("pos_samples",len(pos_samples[0]))

            if len(neg_samples) > 0:
                neg_samples = neg_samples[0]
            else:
                neg_samples = []

            if len(pos_samples) > 0:
                pos_samples = pos_samples[0]
            else:
                pos_samples = []

            rpn_accuracy_rpn_monitor.append(len(pos_samples))
            rpn_accuracy_for_epoch.append((len(pos_samples)))
            #num_rois=32，正例要求小於num_rois//2，其它全部由反例填充
            if C.num_rois > 1:
                if len(pos_samples) < C.num_rois//2:
                    selected_pos_samples = pos_samples.tolist()
                    print("selected_pos_samples",len(selected_pos_samples))
                else:
                    selected_pos_samples = np.random.choice(pos_samples, C.num_rois//2, replace=False).tolist()
                    print("selected_pos_samples",len(selected_pos_samples))
                try:
                    selected_neg_samples = np.random.choice(neg_samples, C.num_rois - len(selected_pos_samples), replace=False).tolist()
                    print("selected_neg_samples",len(selected_neg_samples))
                except:
                    selected_neg_samples = np.random.choice(neg_samples, C.num_rois - len(selected_pos_samples), replace=True).tolist()
                    print("selected_neg_samples",len(selected_neg_samples))
                sel_samples = selected_pos_samples + selected_neg_samples
            else:
                # in the extreme case where num_rois = 1, we pick a random pos or neg sample
                selected_pos_samples = pos_samples.tolist()
                selected_neg_samples = neg_samples.tolist()
                if np.random.randint(0, 2):
                    sel_samples = random.choice(neg_samples)
                else:
                    sel_samples = random.choice(pos_samples)

            print("sel_samples",len(sel_samples))
            print("sel_samples",sel_samples)
            loss_class = model_classifier.train_on_batch([X, X2[:, sel_samples, :]], [Y1[:, sel_samples, :], Y2[:, sel_samples, :]])
#           P_classifier = model_classifier.predict([X, X2[:, sel_samples, :]])
#           #[out_class, out_regr]
#           print("P_classifier_out_class",P_classifier[0].shape)
#           print("P_classifier_out_regr",P_classifier[1].shape)
#           import cv2
#           cv2.waitKey(0)
            losses[iter_num, 0] = loss_rpn[1]
            losses[iter_num, 1] = loss_rpn[2]

            losses[iter_num, 2] = loss_class[1]
            losses[iter_num, 3] = loss_class[2]
            losses[iter_num, 4] = loss_class[3]

            iter_num += 1

            progbar.update(iter_num, [('rpn_cls', np.mean(losses[:iter_num, 0])), ('rpn_regr', np.mean(losses[:iter_num, 1])),
                                      ('detector_cls', np.mean(losses[:iter_num, 2])), ('detector_regr', np.mean(losses[:iter_num, 3]))])

            if iter_num == epoch_length:
                loss_rpn_cls = np.mean(losses[:, 0])
                loss_rpn_regr = np.mean(losses[:, 1])
                loss_class_cls = np.mean(losses[:, 2])
                loss_class_regr = np.mean(losses[:, 3])
                class_acc = np.mean(losses[:, 4])

                mean_overlapping_bboxes = float(sum(rpn_accuracy_for_epoch)) / len(rpn_accuracy_for_epoch)
                rpn_accuracy_for_epoch = []

                if C.verbose:
                    print('Mean number of bounding boxes from RPN overlapping ground truth boxes: {}'.format(mean_overlapping_bboxes))
                    print('Classifier accuracy for bounding boxes from RPN: {}'.format(class_acc))
                    print('Loss RPN classifier: {}'.format(loss_rpn_cls))
                    print('Loss RPN regression: {}'.format(loss_rpn_regr))
                    print('Loss Detector classifier: {}'.format(loss_class_cls))
                    print('Loss Detector regression: {}'.format(loss_class_regr))
                    print('Elapsed time: {}'.format(time.time() - start_time))

                curr_loss = loss_rpn_cls + loss_rpn_regr + loss_class_cls + loss_class_regr
                iter_num = 0
                start_time = time.time()

                if curr_loss < best_loss:
                    if C.verbose:
                        print('Total loss decreased from {} to {}, saving weights'.format(best_loss,curr_loss))
                    best_loss = curr_loss
                    model_all.save_weights(C.model_path)

                break

        except Exception as e:
            print('Exception: {}'.format(e))
            continue

print('Training complete, exiting.')