Caffe2 - (二十三) Detectron 之 utils 函数(1)

Caffe2 - (二十三) Detectron 之 utils 函数(1)

1. blob.py

# Based on: Fast R-CNN
# Written by Ross Girshick
# --------------------------------------------------------

"""Caffe2 blob helper functions."""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals

import cPickle as pickle
import cv2
import numpy as np

from caffe2.proto import caffe2_pb2

from core.config import cfg


def im_list_to_blob(ims):
    """
    将 images 列表转化为网络输入 blob.
    假设 images 已经进行 prep_im_for_blob 处理，或类似的如下处理：
    - BGR channel 顺序
    - 减像素均值
    - 输入尺寸调整
    - 数据为 float32 numpy ndarray 格式

    函数输入： images 列表
    函数输出： 4D HCHW 的 imagess 张量，沿着 axis 0 轴连接而成.
    """

    max_shape = np.array([im.shape for im in ims]).max(axis=0)
    # 图片补零，以便于步长 stride 能够整除.
    if cfg.FPN.FPN_ON:
        stride = float(cfg.FPN.COARSEST_STRIDE)
        max_shape[0] = int(np.ceil(max_shape[0] / stride) * stride)
        max_shape[1] = int(np.ceil(max_shape[1] / stride) * stride)

    num_images = len(ims)
    blob = np.zeros((num_images, max_shape[0], max_shape[1], 3), dtype=np.float32)
    for i in range(num_images):
        im = ims[i]
        blob[i, 0:im.shape[0], 0:im.shape[1], :] = im

    # 将 channels (axis 3) 移动到 axis 1.
    # 处理后，axis 顺序变为：(batch elem, channel, height, width)
    channel_swap = (0, 3, 1, 2)
    blob = blob.transpose(channel_swap)
    return blob


def prep_im_for_blob(im, pixel_means, target_sizes, max_size):
    """
    预处理作为网络输入 blob 的 images，包括：
    - per-channel 减像素均值
    - 数据转化为 float32
    - 将 images 转换到特定尺寸(最大为 max_size)

    函数输出： 
        - 预处理 images 列表 ims；
        - 图片反变换的缩放因子(scale factors) im_scales.
    """
    im = im.astype(np.float32, copy=False)
    im -= pixel_means
    im_shape = im.shape
    im_size_min = np.min(im_shape[0:2])
    im_size_max = np.max(im_shape[0:2])

    ims = []
    im_scales = []
    for target_size in target_sizes:
        im_scale = float(target_size) / float(im_size_min)
        # 防止最大边的值超过 max_size
        if np.round(im_scale * im_size_max) > max_size:
            im_scale = float(max_size) / float(im_size_max)
        im = cv2.resize(im, None, None, fx=im_scale, fy=im_scale, interpolation=cv2.INTER_LINEAR)
        ims.append(im)
        im_scales.append(im_scale)
    return ims, im_scales


def zeros(shape, int32=False):
    """
    返回给定大小的值全为 0 的 blob，数据类型是 int 或 float.
    """
    return np.zeros(shape, dtype=np.int32 if int32 else np.float32)


def ones(shape, int32=False):
    """
    返回给定大小的值全为 1 的 blob，数据类型是 int 或 float.
    """
    return np.ones(shape, dtype=np.int32 if int32 else np.float32)


def py_op_copy_blob(blob_in, blob_out):
    """
    将 numpy ndarray 格式的 blob_in 复制到 Caffe2 CPUTensor blob 格式的 blob_out.

    函数用于将 numpy 数据复制到 PythonOps 中的 Caffe2 blob.
    """
    # 某些情况下可能要求 Caffe2 支持 int32 blobs
    needs_int32_init = False
    try:
        _ = blob.data.dtype  # noqa
    except Exception:
        needs_int32_init = blob_in.dtype == np.int32
    if needs_int32_init:
        # init can only take a list (failed on tuple)
        blob_out.init(list(blob_in.shape), caffe2_pb2.TensorProto.INT32)
    else:
        blob_out.reshape(blob_in.shape)
    blob_out.data[...] = blob_in


def get_loss_gradients(model, loss_blobs):
    """
    对于 loss_blobs 中指定的每个 loss，生成 1 的梯度gradient.
    """
    loss_gradients = {}
    for b in loss_blobs:
        loss_grad = model.net.ConstantFill(b, [b + '_grad'], value=1.0)
        loss_gradients[str(b)] = str(loss_grad)
    return loss_gradients


def serialize(obj):
    """
    采用 pickle 序列化 Python 对象object，并编码为数据类型为 float32 的数组，
    以便于送入 Caffe workspace.

    反序列化 - deserialize().
    """
    return np.fromstring(pickle.dumps(obj), dtype=np.uint8).astype(np.float32)


def deserialize(arr):
    """
    将 Caffe2 workspace 中拉取的数据类型为 float32 的数据反序列化为 Python对象.

    序列化 - serialize().
    """
    return pickle.loads(arr.astype(np.uint8).tobytes())

2. boxes.py

# Based on: Fast/er R-CNN
# Written by Ross Girshick
# --------------------------------------------------------

"""
Box 处理的相关函数.
Detectron 默认的 box 格式为：[x1, y1, x2, y2]
    - (x1, y1): box 的左上角(top-left) 坐标
    - (x2, y2): box 的右下角(bottom-right) 坐标
如果是其它格式 box ，如 [x, y, w, h]，需要进行转换.

boxes.py 提供了一些转换函数(刚开始看可能觉得奇怪的函数)：
    - box 的 width 计算为： x2 - x1 + 1
    - box 的 height 计算为： y2 - y1 + 1
这里 +1 可以追溯到早期目标检测，即坐标是整数像素值，而不是亚像素坐标(subpixel coordinate) 中的浮点数点坐标. x2=x1 和 y2=y1 时的 box 用于包括单个像素，width=1,因此需要 +1.
现在，大部分数据集提供的 boxes 格式一般都是浮点数坐标，因此，width 的计算为 x2 -x1 更合理.

实际上，只要模型训练和测试采用的变换函数是一致的，则结果都是 OK的(至少在 COCO 上已经被验证.)

由于在训练模型时，已经长期都是采用 +1 转换，因此，即使不喜欢使用这种方式，这里也不愿进行改变.
"""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals

import numpy as np

from core.config import cfg 
import utils.cython_bbox as cython_bbox
import utils.cython_nms as cython_nms

bbox_overlaps = cython_bbox.bbox_overlaps
"""
bbox_overlaps:
输入：
    - boxes: (N, 4) 的 float ndarray
    - query_boxes: (K, 4) 的 float ndarray
输出：
    - overlaps: (N, K) ndarray, boxes 和 query_boxes 间的重叠overlap.
"""


def boxes_area(boxes):
    """
    计算 boxes 数组中各 boxes 的面积.
    """
    w = (boxes[:, 2] - boxes[:, 0] + 1)
    h = (boxes[:, 3] - boxes[:, 1] + 1)
    areas = w * h
    assert np.all(areas >= 0), 'Negative areas founds'
    return areas


def unique_boxes(boxes, scale=1.0):
    """
    返回唯一 boxes(unique boxes）的索引.
    """
    v = np.array([1, 1e3, 1e6, 1e9])
    hashes = np.round(boxes * scale).dot(v)
    _, index = np.unique(hashes, return_index=True)
    return np.sort(index)


def xywh_to_xyxy(xywh):
    """
    将 [x1 y1 w h] 格式 box 转换为 [x1 y1 x2 y2] 格式.
    """
    if isinstance(xywh, (list, tuple)):
        # 单个 box 是以坐标值列表的形式给定，顺序是 xywh
        assert len(xywh) == 4
        x1, y1 = xywh[0], xywh[1]
        x2 = x1 + np.maximum(0., xywh[2] - 1.)
        y2 = y1 + np.maximum(0., xywh[3] - 1.)
        return (x1, y1, x2, y2)
    elif isinstance(xywh, np.ndarray):
        # 多个 boxes 是以 2D ndarray 的形式给定.
        return np.hstack((xywh[:, 0:2], xywh[:, 0:2] + np.maximum(0, xywh[:, 2:4] - 1)))
    else:
        raise TypeError('Argument xywh must be a list, tuple, or numpy array.')


def xyxy_to_xywh(xyxy):
    """
    将 [x1 y1 x2 y2] 格式 box 转换为 [x1 y1 w h] 格式.
    """
    if isinstance(xyxy, (list, tuple)):
        # 单个 box 是以坐标值列表的形式给定，顺序是 xyxy
        assert len(xyxy) == 4
        x1, y1 = xyxy[0], xyxy[1]
        w = xyxy[2] - x1 + 1
        h = xyxy[3] - y1 + 1
        return (x1, y1, w, h)
    elif isinstance(xyxy, np.ndarray):
        # 多个 boxes 是以 2D ndarray 的形式给定.
        return np.hstack((xyxy[:, 0:2], xyxy[:, 2:4] - xyxy[:, 0:2] + 1))
    else:
        raise TypeError('Argument xyxy must be a list, tuple, or numpy array.')


def filter_small_boxes(boxes, min_size):
    """
    过滤较小的 boxes;
    只保留 width 和 height 都大于 min_size 的 boxes.
    """
    w = boxes[:, 2] - boxes[:, 0] + 1
    h = boxes[:, 3] - boxes[:, 1] + 1
    keep = np.where((w > min_size) & (h > min_size))[0]
    return keep


def clip_boxes_to_image(boxes, height, width):
    """
    根据给定 height 和 width 的图片来裁剪 boxes array.
    """
    boxes[:, [0, 2]] = np.minimum(width - 1., np.maximum(0., boxes[:, [0, 2]]))
    boxes[:, [1, 3]] = np.minimum(height - 1., np.maximum(0., boxes[:, [1, 3]]))
    return boxes


def clip_xyxy_to_image(x1, y1, x2, y2, height, width):
    """
    根据给定 height 和 width 的图片来裁剪 boxes 的坐标.
    """
    x1 = np.minimum(width - 1., np.maximum(0., x1))
    y1 = np.minimum(height - 1., np.maximum(0., y1))
    x2 = np.minimum(width - 1., np.maximum(0., x2))
    y2 = np.minimum(height - 1., np.maximum(0., y2))
    return x1, y1, x2, y2


def clip_tiled_boxes(boxes, im_shape):
    """
    将 boxes 裁剪到图片边界boundaries.
        - img_shape - [height, width] 
        - boxes - (N, 4 * num_tiled_boxes).
    """
    assert boxes.shape[1] % 4 == 0, 'boxes.shape[1] is {:d}, but must be divisible by 4.'.format(boxes.shape[1])
    # x1 >= 0
    boxes[:, 0::4] = np.maximum(np.minimum(boxes[:, 0::4], im_shape[1] - 1), 0)
    # y1 >= 0
    boxes[:, 1::4] = np.maximum(np.minimum(boxes[:, 1::4], im_shape[0] - 1), 0)
    # x2 < im_shape[1]
    boxes[:, 2::4] = np.maximum(np.minimum(boxes[:, 2::4], im_shape[1] - 1), 0)
    # y2 < im_shape[0]
    boxes[:, 3::4] = np.maximum(np.minimum(boxes[:, 3::4], im_shape[0] - 1), 0)
    return boxes


def bbox_transform(boxes, deltas, weights=(1.0, 1.0, 1.0, 1.0)):
    """
    Forward 变换，采用边界框回归 deltas (bounding-box regression deltas) 将 proposals box 映射到 predicted boxes.
    详细的权重参数描述可见 bbox_transform_inv 函数.
    """
    if boxes.shape[0] == 0:
        return np.zeros((0, deltas.shape[1]), dtype=deltas.dtype)

    boxes = boxes.astype(deltas.dtype, copy=False)

    widths = boxes[:, 2] - boxes[:, 0] + 1.0
    heights = boxes[:, 3] - boxes[:, 1] + 1.0
    ctr_x = boxes[:, 0] + 0.5 * widths
    ctr_y = boxes[:, 1] + 0.5 * heights

    wx, wy, ww, wh = weights
    dx = deltas[:, 0::4] / wx
    dy = deltas[:, 1::4] / wy
    dw = deltas[:, 2::4] / ww
    dh = deltas[:, 3::4] / wh

    # 避免送入 np.exp() 的值太大.
    dw = np.minimum(dw, cfg.BBOX_XFORM_CLIP)
    dh = np.minimum(dh, cfg.BBOX_XFORM_CLIP)

    pred_ctr_x = dx * widths[:, np.newaxis] + ctr_x[:, np.newaxis]
    pred_ctr_y = dy * heights[:, np.newaxis] + ctr_y[:, np.newaxis]
    pred_w = np.exp(dw) * widths[:, np.newaxis]
    pred_h = np.exp(dh) * heights[:, np.newaxis]

    pred_boxes = np.zeros(deltas.shape, dtype=deltas.dtype)
    # x1
    pred_boxes[:, 0::4] = pred_ctr_x - 0.5 * pred_w
    # y1
    pred_boxes[:, 1::4] = pred_ctr_y - 0.5 * pred_h
    # x2 (note: "- 1" is correct; don't be fooled by the asymmetry)
    pred_boxes[:, 2::4] = pred_ctr_x + 0.5 * pred_w - 1
    # y2 (note: "- 1" is correct; don't be fooled by the asymmetry)
    pred_boxes[:, 3::4] = pred_ctr_y + 0.5 * pred_h - 1

    return pred_boxes


def bbox_transform_inv(boxes, gt_boxes, weights=(1.0, 1.0, 1.0, 1.0)):
    """
    Inverse 变换，给定 proposal boxes 和 groundtruth boxes 时，计算目标边界框回归deltas (target bounding-box regression deltas).
    weights 参数应该是 4-tuple 的乘法权重，用于要回归的目标(regression target).

    在旧版本的代码中(包括 py-faster-rcnn)，weights 参数的设置，是为了使训练数据集上，regression deltas 能够的标准偏差为1.
    现在，weights 的设置默认是固定集 (10., 10., 5., 5.)的，而不是精确统计计算得到的.
    这是一种近似权重，采用先前的单位标准偏差启发式从 COCO 上得到的.
    """
    ex_widths = boxes[:, 2] - boxes[:, 0] + 1.0
    ex_heights = boxes[:, 3] - boxes[:, 1] + 1.0
    ex_ctr_x = boxes[:, 0] + 0.5 * ex_widths
    ex_ctr_y = boxes[:, 1] + 0.5 * ex_heights

    gt_widths = gt_boxes[:, 2] - gt_boxes[:, 0] + 1.0
    gt_heights = gt_boxes[:, 3] - gt_boxes[:, 1] + 1.0
    gt_ctr_x = gt_boxes[:, 0] + 0.5 * gt_widths
    gt_ctr_y = gt_boxes[:, 1] + 0.5 * gt_heights

    wx, wy, ww, wh = weights
    targets_dx = wx * (gt_ctr_x - ex_ctr_x) / ex_widths
    targets_dy = wy * (gt_ctr_y - ex_ctr_y) / ex_heights
    targets_dw = ww * np.log(gt_widths / ex_widths)
    targets_dh = wh * np.log(gt_heights / ex_heights)

    targets = np.vstack((targets_dx, targets_dy, targets_dw, targets_dh)).transpose()
    return targets


def expand_boxes(boxes, scale):
    """
    跟根据给定的 scale 来扩展 boxes array.
    """
    w_half = (boxes[:, 2] - boxes[:, 0]) * .5
    h_half = (boxes[:, 3] - boxes[:, 1]) * .5
    x_c = (boxes[:, 2] + boxes[:, 0]) * .5
    y_c = (boxes[:, 3] + boxes[:, 1]) * .5

    w_half *= scale
    h_half *= scale

    boxes_exp = np.zeros(boxes.shape)
    boxes_exp[:, 0] = x_c - w_half
    boxes_exp[:, 2] = x_c + w_half
    boxes_exp[:, 1] = y_c - h_half
    boxes_exp[:, 3] = y_c + h_half

    return boxes_exp


def flip_boxes(boxes, im_width):
    """
    水平反转 boxes.
    """
    boxes_flipped = boxes.copy()
    boxes_flipped[:, 0::4] = im_width - boxes[:, 2::4] - 1
    boxes_flipped[:, 2::4] = im_width - boxes[:, 0::4] - 1
    return boxes_flipped


def aspect_ratio(boxes, aspect_ratio):
    """
    相对于 width (width-relative) 的长宽比(aspect ratio) 变换.
    """
    boxes_ar = boxes.copy()
    boxes_ar[:, 0::4] = aspect_ratio * boxes[:, 0::4]
    boxes_ar[:, 2::4] = aspect_ratio * boxes[:, 2::4]
    return boxes_ar


def box_voting(top_dets, all_dets, thresh, scoring_method='ID', beta=1.0):
    """
    对 all_dets 进行边界框投票(bounding-box voting) 来改善 top_dets.
    参见: https://arxiv.org/abs/1505.01749.
    可选的 score 平均方法(不在参考论文里) 可以根据 scoring_method 进行设置.
    """
    # top_dets is [N, 5] each row is [x1 y1 x2 y2, sore]
    # all_dets is [N, 5] each row is [x1 y1 x2 y2, sore]
    top_dets_out = top_dets.copy()
    top_boxes = top_dets[:, :4]
    all_boxes = all_dets[:, :4]
    all_scores = all_dets[:, 4]
    top_to_all_overlaps = bbox_overlaps(top_boxes, all_boxes)
    for k in range(top_dets_out.shape[0]):
        inds_to_vote = np.where(top_to_all_overlaps[k] >= thresh)[0]
        boxes_to_vote = all_boxes[inds_to_vote, :]
        ws = all_scores[inds_to_vote]
        top_dets_out[k, :4] = np.average(boxes_to_vote, axis=0, weights=ws)
        if scoring_method == 'ID':
            # Identity, nothing to do
            pass
        elif scoring_method == 'TEMP_AVG':
            # Average probabilities (considered as P(detected class) vs.
            # P(not the detected class)) after smoothing with a temperature
            # hyperparameter.
            P = np.vstack((ws, 1.0 - ws))
            P_max = np.max(P, axis=0)
            X = np.log(P / P_max)
            X_exp = np.exp(X / beta)
            P_temp = X_exp / np.sum(X_exp, axis=0)
            P_avg = P_temp[0].mean()
            top_dets_out[k, 4] = P_avg
        elif scoring_method == 'AVG':
            # Combine new probs from overlapping boxes
            top_dets_out[k, 4] = ws.mean()
        elif scoring_method == 'IOU_AVG':
            P = ws
            ws = top_to_all_overlaps[k, inds_to_vote]
            P_avg = np.average(P, weights=ws)
            top_dets_out[k, 4] = P_avg
        elif scoring_method == 'GENERALIZED_AVG':
            P_avg = np.mean(ws**beta)**(1.0 / beta)
            top_dets_out[k, 4] = P_avg
        elif scoring_method == 'QUASI_SUM':
            top_dets_out[k, 4] = ws.sum() / float(len(ws))**beta
        else:
            raise NotImplementedError('Unknown scoring method {}'.format(scoring_method) )

    return top_dets_out


def nms(dets, thresh):
    """
    采用经典的 DPM-style 的贪婪 NMS.
    """
    if dets.shape[0] == 0:
        return []
    return cython_nms.nms(dets, thresh)


def soft_nms(dets, sigma=0.5, overlap_thresh=0.3, score_thresh=0.001, method='linear'):
    """
    采用论文 https://arxiv.org/abs/1704.04503 的 soft NMS 算法.
    """
    if dets.shape[0] == 0:
        return dets, []

    methods = {'hard': 0, 'linear': 1, 'gaussian': 2}
    assert method in methods, 'Unknown soft_nms method: {}'.format(method)

    dets, keep = cython_nms.soft_nms(np.ascontiguousarray(dets, dtype=np.float32),
                                     np.float32(sigma),
                                     np.float32(overlap_thresh),
                                     np.float32(score_thresh),
                                     np.uint8(methods[method]) )
    return dets, keep

3. image.py

"""Image helper functions."""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals

import cv2
import numpy as np


def aspect_ratio_rel(im, aspect_ratio):
    """
    相对于 width(width-relative) 的长宽比(aspect ratio) 变换.
    """
    im_h, im_w = im.shape[:2]
    im_ar_w = int(round(aspect_ratio * im_w))
    im_ar = cv2.resize(im, dsize=(im_ar_w, im_h))
    return im_ar


def aspect_ratio_abs(im, aspect_ratio):
    """
    绝对长宽比(absolute aspect ratio) 变换.
    """
    im_h, im_w = im.shape[:2]
    im_area = im_h * im_w

    im_ar_w = np.sqrt(im_area * aspect_ratio)
    im_ar_h = np.sqrt(im_area / aspect_ratio)
    assert np.isclose(im_ar_w / im_ar_h, aspect_ratio)

    im_ar = cv2.resize(im, dsize=(int(im_ar_w), int(im_ar_h)))
    return im_ar

c