py-faster-rcnn 訓練參數修改(轉)

faster rcnn默認有三種網絡模型 ZF（小）、VGG_CNN_M_1024（中）、VGG16 （大）

 

訓練圖片大小爲500*500，類別數1。

一. 修改VGG_CNN_M_1024模型配置文件

1）train.prototxt文件

      input-data層的num_class數值由21改成2；

      roi-data層的num_class數值由21改成2；

      cls_score層的num_output數值由21改成2（1+1）；

      bbox_pred層的num_output數值由84改成8（2*4）；

2）test.prototxt文件(c++dll調用的.prototxt也要改)

cls_score層的num_output數值由21改成2（1+1）；

bbox_pred層的num_output數值由84改成8（2*4）;

3）lib/datasets/pascal_voc.py文件

       修改self._classes = ('__background__',  '訓練的數據類別')

4) 測試模型時須要改的文件faster_rcnn_test.pt

cls_score層的num_output數值由21改成2；

bbox_pred層的num_output數值由84改成8；

 二. 解讀訓練測試配置參數文件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:
# 在其餘文件使用config要加的命令，例子見train_net.py
#   from fast_rcnn_config import cfg
cfg = __C
 
#
# Training options
# 訓練的選項
#
 
__C.TRAIN = edict()
 
# Scales to use during training (can list multiple scales)
# Each scale is the pixel size of an image's shortest side
# 最短邊Scale成600
__C.TRAIN.SCALES = (600,)
 
# Max pixel size of the longest side of a scaled input image
# 最長邊最大爲1000
__C.TRAIN.MAX_SIZE = 1000
 
# Images to use per minibatch
# 一個minibatch包含兩張圖片
__C.TRAIN.IMS_PER_BATCH = 2
 
# Minibatch size (number of regions of interest [ROIs])
#  Minibatch大小，即ROI的數量
__C.TRAIN.BATCH_SIZE = 128
 
# Fraction of minibatch that is labeled foreground (i.e. class > 0)
# minibatch中前景樣本所佔的比例
__C.TRAIN.FG_FRACTION = 0.25
 
# Overlap threshold for a ROI to be considered foreground (if >= FG_THRESH)
# 與前景的overlap大於等於0.5認爲該ROI爲前景樣本
__C.TRAIN.FG_THRESH = 0.5
 
# Overlap threshold for a ROI to be considered background (class = 0 if
# overlap in [LO, HI))
# 與前景的overlap在0.1-0.5認爲該ROI爲背景樣本
__C.TRAIN.BG_THRESH_HI = 0.5
__C.TRAIN.BG_THRESH_LO = 0.1
 
# Use horizontally-flipped images during training?
# 水平翻轉圖像，增長數據量
__C.TRAIN.USE_FLIPPED = True
 
# Train bounding-box regressors
# 訓練bb迴歸器
__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閾值，只有ROI與gt的重疊度大於閾值，這樣的ROI才能用做bb迴歸的訓練樣本
__C.TRAIN.BBOX_THRESH = 0.5
 
# Iterations between snapshots
# 每迭代1000次產生一次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
# 爲產生的snapshot文件名稱添加一個可選的infix. solver.prototxt指定了snapshot名稱的前綴
__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使用預取線程，做者認爲不太有效，所以設爲False
__C.TRAIN.USE_PREFETCH = False
 
# Normalize the targets (subtract empirical mean, divide by empirical stddev)
# 歸一化目標BBOX_NORMALIZE_TARGETS，減去經驗均值，除以標準差
__C.TRAIN.BBOX_NORMALIZE_TARGETS = True
# Deprecated (inside weights)
# 棄用
__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爲True時，歸一化targets,使用經驗均值和方差
__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訓練！注意該文件來自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的兩個圖片應該有類似的寬高比，以免冗餘的zero-padding計算
__C.TRAIN.ASPECT_GROUPING = True
 
# Use RPN to detect objects
# 使用RPN檢測目標
__C.TRAIN.HAS_RPN = False
# IOU >= thresh: positive example
# RPN的正樣本閾值
__C.TRAIN.RPN_POSITIVE_OVERLAP = 0.7
# IOU < thresh: negative example
# RPN的負樣本閾值
__C.TRAIN.RPN_NEGATIVE_OVERLAP = 0.3
# If an anchor statisfied by positive and negative conditions set to negative
# 若是一個anchor同時知足正負樣本條件，設爲負樣本（應該用不到）
__C.TRAIN.RPN_CLOBBER_POSITIVES = False
# Max number of foreground examples
# 前景樣本的比例
__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
# 非極大值抑制的閾值
__C.TRAIN.RPN_NMS_THRESH = 0.7
# Number of top scoring boxes to keep before apply NMS to RPN proposals
# 在對RPN proposal使用NMS前，要保留的top scores的box數量
__C.TRAIN.RPN_PRE_NMS_TOP_N = 12000
# Number of top scoring boxes to keep after applying NMS to RPN proposals
# 在對RPN proposal使用NMS後，要保留的top scores的box數量
__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的高和寬都應該大於RPN_MIN_SIZE，不然，映射到conv5上不足一個像素點
__C.TRAIN.RPN_MIN_SIZE = 16
# Deprecated (outside weights)
# 棄用
__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}
# 給定正RPN樣本的權重
# and give negatives a weight of (1 - p)
# 給定負RPN樣本的權重
# Set to -1.0 to use uniform example weighting
# 這裏正負樣本使用相同權重
__C.TRAIN.RPN_POSITIVE_WEIGHT = -1.0
 
#
# Testing options
# 測試選項 ,類同
#
 
__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)
# 測試時非極大值抑制的閾值
__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).
# 分類再也不用SVM，設置爲False
__C.TEST.SVM = False
 
# Test using bounding-box regressors
# 使用bb迴歸
__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閾值
__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
# 從原圖到feature map的座標映射，可能會形成在原圖上不一樣的box到了feature map座標系上變得相同了
# 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
# 縮放因子
__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
# 全部network所用的像素均值設爲相同
__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
# 極小的數
__C.EPS = 1e-14
 
# Root directory of project
# 項目根路徑
__C.ROOT_DIR = osp.abspath(osp.join(osp.dirname(__file__), '..', '..'))
 
# Data directory
# 數據路徑
__C.DATA_DIR = osp.abspath(osp.join(__C.ROOT_DIR, 'data'))
 
# Model directory
# 模型路徑
__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
# 輸出在experiments路徑下
__C.EXP_DIR = 'default'
 
# Use GPU implementation of non-maximum suppression
# GPU實施非極大值抑制
__C.USE_GPU_NMS = True
 
# Default GPU device id
# 默認GPU id
__C.GPU_ID = 0
 
def get_output_dir(imdb, net=None):
    #返回輸出路徑，在experiments路徑下
    """Return the directory where experimental artifacts are placed.
    If the directory does not exist, it is created.
 
    A canonical標準 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):
    #兩個配置文件融合
    """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的對應項
        else:
            b[k] = v
 
def cfg_from_file(filename):
    """Load a config file and merge it into the default options."""
    # 導入配置文件並與默認選項融合
    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):
    # 命令行設置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問題

在從新訓練新的數據以前將cache刪除

1) py-faster-rcnn/output 
2) py-faster-rcnn/data/cache




四. 超參數

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

....

總結solver文件個參數的意義

iteration： 數據進行一次前向-後向的訓練
batchsize：每次迭代訓練圖片的數量
epoch：1個epoch就是將全部的訓練圖像所有經過網絡訓練一次
例如：假若有1280000張圖片，batchsize=256，則1個epoch須要1280000/256=5000次iteration
它的max-iteration=450000，則共有450000/5000=90個epoch
而lr何時衰減與stepsize有關，減小多少與gamma有關，即:若stepsize=500, base_lr=0.01, gamma=0.1,則當迭代到第一個500次時，lr第一次衰減，衰減後的lr=lr*gamma=0.01*0.1=0.001,之後重複該過程，因此
stepsize是lr的衰減步長，gamma是lr的衰減係數。
在訓練過程當中，每到必定的迭代次數都會測試，迭代次數是由test-interval決定的，如test_interval=1000，則訓練集每迭代1000次測試一遍網絡，而
test_size, test_iter, 和test圖片的數量決定了怎樣test, test-size決定了test時每次迭代輸入圖片的數量，test_iter就是test全部的圖片的迭代次數，如：500張test圖片，test_iter=100，則test_size=5, 而solver文檔裏只須要根據test圖片總數量來設置test_iter，以及根據須要設置test_interval便可。

迭代次數在文件py-faster-rcnn/tools/train_faster_rcnn_alt_opt.py中進行修改

max_iters=[80000, 40000, 80000, 40000]

分別對應rpn第1階段，fast rcnn第1階段，rpn第2階段，fast rcnn第2階段的迭代次數。

 

1. 預訓練的ImageNet模型,放在下面的文件夾下,個人是VGG_CNN_M_1024.v2.caffemodel

    

2. 兩種訓練數據的算法
   (1) 使用交替優化（alternating optimization）算法來訓練和測試Faster R-CNN

   1 2 3 4 5 6 7 8

   cd $FRCN_ROOT ./experiments/scripts/faster_rcnn_alt_opt.sh [GPU_ID] [NET] [--set ...] # GPU_ID是你想要訓練的GPUID # 你能夠選擇以下的網絡之一進行訓練：ZF, VGG_CNN_M_1024, VGG16 # --set ... 運行你自定義fast_rcnn.config參數,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

輸出的結果在 $FRCN_ROOT/output下。訓練過程截圖：

(2) 使用近似聯合訓練（ approximate joint training）

cd $FRCN_ROOT
./experiments/scripts/faster_rcnn_end2end.sh [GPU_ID] [NET] [--set ...]

這個方法是聯合RPN模型和Fast R-CNN網絡訓練。而不是交替訓練。用此種方法比交替優化快1.5倍，可是準確率相近。因此推薦使用這種方法

開始訓練:

cd py-faster-rcnn

./experiments/scripts/faster_rcnn_end2end.sh 0 VGG_CNN_M_1024 pascal_voc

參數代表使用第一塊GPU(0)；模型是VGG_CNN_M_1024；訓練數據是pascal_voc(voc2007)。

訓練Fast R-CNN網絡的結果保存在這個目錄下：

output/<experiment directory>/<dataset name>/

測試保存在這個目錄下：

output/<experiment directory>/<dataset name>/<network snapshot name>/