faster中的兩次的負樣本都參與位置回歸，不參與分類

# 最短邊Scale成600
__C.TRAIN.SCALES = (600,)
 
# 最長邊最大為1000
__C.TRAIN.MAX_SIZE = 1000
 
# 一個minibatch包含兩張圖片,但是呼叫是yaml裡改成1，而且只能為1
__C.TRAIN.IMS_PER_BATCH = 2
 -------------------------------------------------------------------------
# 這一部分是PropsalTargetCreator.py中的引數


# 1. 從rpn經過NMS抑制後得到的2000個roi中選出128個sampel_rois，進行標準化處理
 
，最後的訓練
# 2. RoIs 和 gt_bboxes 的 IoU 大於 0.5 的，選擇一些（128*0.25=32）
# **注意這裡如果不夠32個，比如滿足IOU大於0.5的只有20個則剩餘108都是負樣本是108個**
# **但是還需要主要的是這裡的位置回歸只對roi與gt的IOU>0.5[可設定，這裡感覺和上面的選擇條件有些重復，不明所以]正樣本計算損失. 而且是隻對正樣本中的這個類別 4 個引數計算損失，負樣本不進行loss計算***
# 3. 選擇 RoIs 和 gt_bboxes 的 IoU 小於等於 0（或者 0.1）的選擇一些（比如 128-正樣本 個）作為負樣本


#  Minibatch大小，即ROI的數量，如果你的目標比較多的話，可以適當設定的大一些
__C.TRAIN.BATCH_SIZE = 128
 
# minibatch中前景樣本所佔的比例，負樣本只在分類時參與
__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認為該ROI為前景樣本，如果你的anchor經過修改，變得很多，
# 比如小汽車長都在0-128個畫素，然後你的anchor的長寬從4 14 24 34 ...128 這裡的IOU可以比較大，這樣有利於快速收斂和精度的提高

__C.TRAIN.FG_THRESH = 0.5 
# 與前景的overlap在0.1-0.5認為該ROI為背景樣本
__C.TRAIN.BG_THRESH_HI = 0.5
__C.TRAIN.BG_THRESH_LO = 0.1

# Use horizontally-flipped images during training?
# 水平翻轉影象，增加資料量
__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才能用作bb迴歸的訓練樣本
# 如果你上面的.FG_THRESH做過調整，這裡最好做相應的調整 
__C.TRAIN.BBOX_THRESH = 0.5

# Iterations between snapshots
# 每迭代1000次產生一次snapshot
__C.TRAIN.SNAPSHOT_ITERS = 10000
 
# 為產生的snapshot檔名稱新增一個可選的infix. solver.prototxt指定了snapshot名稱的字首
__C.TRAIN.SNAPSHOT_INFIX = ''
 
# 在roi_data_layer.layer使用預取執行緒，作者認為不太有效，因此設為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為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！注意該檔案來自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的兩個圖片應該有相似的寬高比，以避免冗餘的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
# 如果一個anchor同時滿足正負樣本條件，設為負樣本（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，否則，對映到conv5上不足一個畫素點
__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
# 這裡正負樣本使用相同權重
__C.TRAIN.RPN_POSITIVE_WEIGHT = -1.0
 
#
# Testing options
# 測試選項
#
 
__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).
# 分類不再用SVM，設定為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的座標對映，可能會造成在原圖上不同的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
# 所有network所用的畫素均值設為相同
__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
# 專案根路徑
__C.ROOT_DIR = osp.abspath(osp.join(osp.dirname(__file__), '..', '..'))
 
# Data directory
# 資料路徑
__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
# 預設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):
    #兩個配置檔案融合
    """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."""
    # 匯入配置檔案並與預設選項融合
    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):
    # 命令列設定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

 

• AnchorTargetCreator裡的引數

• 將 20000 多個候選的 anchor 選出 256 個 anchor 進行分類和迴歸位置。選擇過程如下：

• 對於每一個 ground truth bounding box選擇和它重疊度（IoU）最高的一個 anchor 作為正樣本

• 對於剩下的 anchor，從中選擇和任意一個重疊度超過 0.7 的 anchor，作為正樣本，正樣本的數目不超過 128 個。

• 隨機選擇和重疊度小於 0.3 的 anchor 作為負樣本。負樣本和正樣本的總數為 256。

• 對於每個 anchor, gt_label 要麼為 1（前景），要麼為 0（背景），而 gt_loc 則是由 4 個位置引數 (tx,ty,tw,th) 組成，這樣比直接回歸座標更好。
• 計算分類損失用的是交叉熵損失，而計算迴歸損失用的是 Smooth_l1_loss. 在計算迴歸損失的時候，只計算正樣本（128個前景）的損失，不計算負樣本的位置損失。