faster rcnn默认有三种网络模型 ZF(小)、VGG_CNN_M_1024(中)、VGG16 (大)
训练图片大小为500*500,类别数1。
一. 修改VGG_CNN_M_1024模型配置文件
4) 测试模型时需要改的文件faster_rcnn_test.pt
cls_score层的num_output数值由21改为2;
bbox_pred层的num_output数值由84改为8;
二. 解读训练测试配置参数文件config.py
import os import os.path as osp import numpy as np # `pip install easydict` if you don't have it from easydict import EasyDict as edict __C = edict() # Consumers can get config by: # 在其他文件使用config要加的命令,例子见train_net.py # from fast_rcnn_config import cfg cfg = __C # # Training options # 训练的选项 # __C.TRAIN = edict() # Scales to use during training (can list multiple scales) # Each scale is the pixel size of an image's shortest side # 最短边Scale成600 __C.TRAIN.SCALES = (600,) # Max pixel size of the longest side of a scaled input image # 最长边最大为1000 __C.TRAIN.MAX_SIZE = 1000 # Images to use per minibatch # 一个minibatch包含两张图片 __C.TRAIN.IMS_PER_BATCH = 2 # Minibatch size (number of regions of interest [ROIs]) # Minibatch大小,即ROI的数量 __C.TRAIN.BATCH_SIZE = 128 # Fraction of minibatch that is labeled foreground (i.e. class > 0) # minibatch中前景样本所占的比例 __C.TRAIN.FG_FRACTION = 0.25 # Overlap threshold for a ROI to be considered foreground (if >= FG_THRESH) # 与前景的overlap大于等于0.5认为该ROI为前景样本 __C.TRAIN.FG_THRESH = 0.5 # Overlap threshold for a ROI to be considered background (class = 0 if # overlap in [LO, HI)) # 与前景的overlap在0.1-0.5认为该ROI为背景样本 __C.TRAIN.BG_THRESH_HI = 0.5 __C.TRAIN.BG_THRESH_LO = 0.1 # Use horizontally-flipped images during training? # 水平翻转图像,增加数据量 __C.TRAIN.USE_FLIPPED = True # Train bounding-box regressors # 训练bb回归器 __C.TRAIN.BBOX_REG = True # Overlap required between a ROI and ground-truth box in order for that ROI to # be used as a bounding-box regression training example # BBOX阈值,只有ROI与gt的重叠度大于阈值,这样的ROI才能用作bb回归的训练样本 __C.TRAIN.BBOX_THRESH = 0.5 # Iterations between snapshots # 每迭代1000次产生一次snapshot __C.TRAIN.SNAPSHOT_ITERS = 10000 # solver.prototxt specifies the snapshot path prefix, this adds an optional # infix to yield the path: <prefix>[_<infix>]_iters_XYZ.caffemodel # 为产生的snapshot文件名称添加一个可选的infix. solver.prototxt指定了snapshot名称的前缀 __C.TRAIN.SNAPSHOT_INFIX = '' # Use a prefetch thread in roi_data_layer.layer # So far I haven't found this useful; likely more engineering work is required # 在roi_data_layer.layer使用预取线程,作者认为不太有效,因此设为False __C.TRAIN.USE_PREFETCH = False # Normalize the targets (subtract empirical mean, divide by empirical stddev) # 归一化目标BBOX_NORMALIZE_TARGETS,减去经验均值,除以标准差 __C.TRAIN.BBOX_NORMALIZE_TARGETS = True # Deprecated (inside weights) # 弃用 __C.TRAIN.BBOX_INSIDE_WEIGHTS = (1.0, 1.0, 1.0, 1.0) # Normalize the targets using "precomputed" (or made up) means and stdevs # (BBOX_NORMALIZE_TARGETS must also be True) # 在BBOX_NORMALIZE_TARGETS为True时,归一化targets,使用经验均值和方差 __C.TRAIN.BBOX_NORMALIZE_TARGETS_PRECOMPUTED = False __C.TRAIN.BBOX_NORMALIZE_MEANS = (0.0, 0.0, 0.0, 0.0) __C.TRAIN.BBOX_NORMALIZE_STDS = (0.1, 0.1, 0.2, 0.2) # Train using these proposals # 使用'selective_search'的proposal训练!注意该文件来自fast rcnn,下文提到RPN __C.TRAIN.PROPOSAL_METHOD = 'selective_search' # Make minibatches from images that have similar aspect ratios (i.e. both # tall and thin or both short and wide) in order to avoid wasting computation # on zero-padding. # minibatch的两个图片应该有相似的宽高比,以避免冗余的zero-padding计算 __C.TRAIN.ASPECT_GROUPING = True # Use RPN to detect objects # 使用RPN检测目标 __C.TRAIN.HAS_RPN = False # IOU >= thresh: positive example # RPN的正样本阈值 __C.TRAIN.RPN_POSITIVE_OVERLAP = 0.7 # IOU < thresh: negative example # RPN的负样本阈值 __C.TRAIN.RPN_NEGATIVE_OVERLAP = 0.3 # If an anchor statisfied by positive and negative conditions set to negative # 如果一个anchor同时满足正负样本条件,设为负样本(应该用不到) __C.TRAIN.RPN_CLOBBER_POSITIVES = False # Max number of foreground examples # 前景样本的比例 __C.TRAIN.RPN_FG_FRACTION = 0.5 # Total number of examples # batch size大小 __C.TRAIN.RPN_BATCHSIZE = 256 # NMS threshold used on RPN proposals # 非极大值抑制的阈值 __C.TRAIN.RPN_NMS_THRESH = 0.7 # Number of top scoring boxes to keep before apply NMS to RPN proposals # 在对RPN proposal使用NMS前,要保留的top scores的box数量 __C.TRAIN.RPN_PRE_NMS_TOP_N = 12000 # Number of top scoring boxes to keep after applying NMS to RPN proposals # 在对RPN proposal使用NMS后,要保留的top scores的box数量 __C.TRAIN.RPN_POST_NMS_TOP_N = 2000 # Proposal height and width both need to be greater than RPN_MIN_SIZE (at orig image scale) # proposal的高和宽都应该大于RPN_MIN_SIZE,否则,映射到conv5上不足一个像素点 __C.TRAIN.RPN_MIN_SIZE = 16 # Deprecated (outside weights) # 弃用 __C.TRAIN.RPN_BBOX_INSIDE_WEIGHTS = (1.0, 1.0, 1.0, 1.0) # Give the positive RPN examples weight of p * 1 / {num positives} # 给定正RPN样本的权重 # and give negatives a weight of (1 - p) # 给定负RPN样本的权重 # Set to -1.0 to use uniform example weighting # 这里正负样本使用相同权重 __C.TRAIN.RPN_POSITIVE_WEIGHT = -1.0 # # Testing options # 测试选项 ,类同 # __C.TEST = edict() # Scales to use during testing (can list multiple scales) # Each scale is the pixel size of an image's shortest side __C.TEST.SCALES = (600,) # Max pixel size of the longest side of a scaled input image __C.TEST.MAX_SIZE = 1000 # Overlap threshold used for non-maximum suppression (suppress boxes with # IoU >= this threshold) # 测试时非极大值抑制的阈值 __C.TEST.NMS = 0.3 # Experimental: treat the (K+1) units in the cls_score layer as linear # predictors (trained, eg, with one-vs-rest SVMs). # 分类不再用SVM,设置为False __C.TEST.SVM = False # Test using bounding-box regressors # 使用bb回归 __C.TEST.BBOX_REG = True # Propose boxes # 不使用RPN生成proposal __C.TEST.HAS_RPN = False # Test using these proposals # 使用selective_search生成proposal __C.TEST.PROPOSAL_METHOD = 'selective_search' ## NMS threshold used on RPN proposals # RPN proposal的NMS阈值 __C.TEST.RPN_NMS_THRESH = 0.7 ## Number of top scoring boxes to keep before apply NMS to RPN proposals __C.TEST.RPN_PRE_NMS_TOP_N = 6000 ## Number of top scoring boxes to keep after applying NMS to RPN proposals __C.TEST.RPN_POST_NMS_TOP_N = 300 # Proposal height and width both need to be greater than RPN_MIN_SIZE (at orig image scale) __C.TEST.RPN_MIN_SIZE = 16 # # MISC # # The mapping from image coordinates to feature map coordinates might cause # 从原图到feature map的坐标映射,可能会造成在原图上不同的box到了feature map坐标系上变得相同了 # some boxes that are distinct in image space to become identical in feature # coordinates. If DEDUP_BOXES > 0, then DEDUP_BOXES is used as the scale factor # for identifying duplicate boxes. # 1/16 is correct for {Alex,Caffe}Net, VGG_CNN_M_1024, and VGG16 # 缩放因子 __C.DEDUP_BOXES = 1./16. # Pixel mean values (BGR order) as a (1, 1, 3) array # We use the same pixel mean for all networks even though it's not exactly what # they were trained with # 所有network所用的像素均值设为相同 __C.PIXEL_MEANS = np.array([[[102.9801, 115.9465, 122.7717]]]) # For reproducibility __C.RNG_SEED = 3 # A small number that's used many times # 极小的数 __C.EPS = 1e-14 # Root directory of project # 项目根路径 __C.ROOT_DIR = osp.abspath(osp.join(osp.dirname(__file__), '..', '..')) # Data directory # 数据路径 __C.DATA_DIR = osp.abspath(osp.join(__C.ROOT_DIR, 'data')) # Model directory # 模型路径 __C.MODELS_DIR = osp.abspath(osp.join(__C.ROOT_DIR, 'models', 'pascal_voc')) # Name (or path to) the matlab executable # matlab executable __C.MATLAB = 'matlab' # Place outputs under an experiments directory # 输出在experiments路径下 __C.EXP_DIR = 'default' # Use GPU implementation of non-maximum suppression # GPU实施非极大值抑制 __C.USE_GPU_NMS = True # Default GPU device id # 默认GPU id __C.GPU_ID = 0 def get_output_dir(imdb, net=None): #返回输出路径,在experiments路径下 """Return the directory where experimental artifacts are placed. If the directory does not exist, it is created. A canonical标准 path is built using the name from an imdb and a network (if not None). """ outdir = osp.abspath(osp.join(__C.ROOT_DIR, 'output', __C.EXP_DIR, imdb.name)) if net is not None: outdir = osp.join(outdir, net.name) if not os.path.exists(outdir): os.makedirs(outdir) return outdir def _merge_a_into_b(a, b): #两个配置文件融合 """Merge config dictionary a into config dictionary b, clobbering the options in b whenever they are also specified in a. """ if type(a) is not edict: return for k, v in a.iteritems(): # a must specify keys that are in b if not b.has_key(k): raise KeyError('{} is not a valid config key'.format(k)) # the types must match, too old_type = type(b[k]) if old_type is not type(v): if isinstance(b[k], np.ndarray): v = np.array(v, dtype=b[k].dtype) else: raise ValueError(('Type mismatch ({} vs. {}) ' 'for config key: {}').format(type(b[k]), type(v), k)) # recursively merge dicts if type(v) is edict: try: _merge_a_into_b(a[k], b[k]) except: print('Error under config key: {}'.format(k)) raise #用配置a更新配置b的对应项 else: b[k] = v def cfg_from_file(filename): """Load a config file and merge it into the default options.""" # 导入配置文件并与默认选项融合 import yaml with open(filename, 'r') as f: yaml_cfg = edict(yaml.load(f)) _merge_a_into_b(yaml_cfg, __C) def cfg_from_list(cfg_list): # 命令行设置config """Set config keys via list (e.g., from command line).""" from ast import literal_eval assert len(cfg_list) % 2 == 0 for k, v in zip(cfg_list[0::2], cfg_list[1::2]): key_list = k.split('.') d = __C for subkey in key_list[:-1]: assert d.has_key(subkey) d = d[subkey] subkey = key_list[-1] assert d.has_key(subkey) try: value = literal_eval(v) except: # handle the case when v is a string literal value = v assert type(value) == type(d[subkey]), 'type {} does not match original type {}'.format( type(value), type(d[subkey])) d[subkey] = value
三. cache问题
在重新训练新的数据之前将cache删除
1) py-faster-rcnn/output
2) py-faster-rcnn/data/cache
四. 超参数
py-faster-rcnn/models/pascal_voc/VGG16/faster_rcnn_alt_opt/stage_fast_rcnn_solver*.pt
base_lr:
0.001
lr_policy:
'step'
step_size:
30000
display:
20
....
总结solver文件个参数的意义
iteration: 数据进行一次前向-后向的训练
batchsize:每次迭代训练图片的数量
epoch:1个epoch就是将所有的训练图像全部通过网络训练一次
例如:假如有1280000张图片,batchsize=256,则1个epoch需要1280000/256=5000次iteration
它的max-iteration=450000,则共有450000/5000=90个epoch
而lr什么时候衰减与stepsize有关,减少多少与gamma有关,即:若stepsize=500, base_lr=0.01, gamma=0.1,则当迭代到第一个500次时,lr第一次衰减,衰减后的lr=lr*gamma=0.01*0.1=0.001,以后重复该过程,所以
stepsize是lr的衰减步长,gamma是lr的衰减系数。
在训练过程中,每到一定的迭代次数都会测试,迭代次数是由test-interval决定的,如test_interval=1000,则训练集每迭代1000次测试一遍网络,而
test_size, test_iter, 和test图片的数量决定了怎样test, test-size决定了test时每次迭代输入图片的数量,test_iter就是test所有的图片的迭代次数,如:500张test图片,test_iter=100,则test_size=5, 而solver文档里只需要根据test图片总数量来设置test_iter,以及根据需要设置test_interval即可。
迭代次数在文件py-faster-rcnn/tools/train_faster_rcnn_alt_opt.py
中进行修改
max_iters=[80000, 40000, 80000, 40000]
分别对应rpn第1阶段,fast rcnn第1阶段,rpn第2阶段,fast rcnn第2阶段的迭代次数。
-
预训练的ImageNet模型,放在下面的文件夹下,我的是VGG_CNN_M_1024.v2.caffemodel
- 两种训练数据的算法
(1) 使用交替优化(alternating optimization)算法来训练和测试Faster R-CNN1
2
3
4
5
6
7
8cd $FRCN_ROOT
./experiments/scripts/faster_rcnn_alt_opt.sh [GPU_ID] [NET] [--set ...]
# GPU_ID是你想要训练的GPUID
# 你可以选择如下的网络之一进行训练:ZF, VGG_CNN_M_1024, VGG16
# --set ... 运行你自定义fast_rcnn.config参数,例如.
# --set EXP_DIR seed_rng1701 RNG_SEED 1701
#例如命令
./experiments/scripts/faster_rcnn_alt_opt.sh 0 ZF pascal_voc
输出的结果在 $FRCN_ROOT/output下。训练过程截图:
(2) 使用近似联合训练( approximate joint training)
cd $FRCN_ROOT
./experiments/scripts/faster_rcnn_end2end.sh [GPU_ID] [NET] [--set ...]
这个方法是联合RPN模型和Fast R-CNN网络训练。而不是交替训练。用此种方法比交替优化快1.5倍,但是准确率相近。所以推荐使用这种方法
开始训练:
cd py-faster-rcnn
./experiments/scripts/faster_rcnn_end2end.sh 0 VGG_CNN_M_1024 pascal_voc
参数表明使用第一块GPU(0);模型是VGG_CNN_M_1024;训练数据是pascal_voc(voc2007)。
训练Fast R-CNN网络的结果保存在这个目录下:
output/<experiment directory>/<dataset name>/
测试保存在这个目录下:
output/<experiment directory>/<dataset name>/<network snapshot name>/