yolo3的学习

 yolo3与SSD、retinaNet等在MAP和时间上的对比：

(关于MAP:MAP的全称为mean average precision，也就是均值平均精度，好多时候都直接称为准确度，它是的AP取的均值，比如在一个识别任务里面会有好多类目标，而每一类里面有多个不同的具体目标，每个目标的预测结果与真实结果之间的差别用precision来表示，对这一类平均就是AP，对每一类取均值，那就是MAP，还有一个作为网络评判的标注是recall，即召回率，召回率描述的是把正确结果找出来的程度。如40个男生，60个女生，要找50个女生，结果找出来的50个里有40个女生，10个男生，那她的precision就是40/50 ，而recall是40/60，precision的分母是找出结果的总数量，recall的分母是原有结构里面正确 数目之和）

 yolo3论文上的内容比较少，更多的细节还是得分析代码。论文中主要讲了下面几点：

1.与yolo2相似的，yolo3仍旧使用dimension clusters来产生anchor box,预测位置上的4个值，也yolo2一致

 yolo3使用logistic回归对每个bounding box给出是否为目标的置信度，如果某个bounding box的IOU最大就取为1,对于

其他的非极大值的就忽略。这样对于每个groud truth只分配一个最好的bounding box，对于没有分配的bounding box,对于位置和类别没有损失，只有判别是否有目标的置信度有损失。

2.考虑标签的重叠性（如女人和人），类别预测上不用softmax，因为softmax有个前提是是每个box对应一个标签，这对于多标签的数据很难处理，因此使用的是独立的logistic classifiers，训练时损失函数采用的是二分的交叉熵损失。

3.借鉴FPN，在不同的尺度上预测，在实验中在不同尺度上预测3个box，80个类别得到的是N*N*[3*(4+1+80)],上采样2x，与之前的通过element-wide相加的到merge,然后再接一些conv，具体细节请看下面代码部分。

首先，是yolo3的网络结构,如下图所示，主体部分被称为darknet53，右下脚部分借鉴FPN的思想，采用上采样与merge的方法进行预测。

在代码中，作者先定义了DarknetConv2D函数，在此基础上又定义了DarknetConv2D_BN_Leaky，也就是在DarknetConv2D之后加上了bathnorm操作和LeakyReLU激活函数，并在此上定义resblock_body，让卷积成块。

compose（A，B，C）（x)=C（B（A（x）））

*args是不可变名参数，如一个固定值20;**kwargs是可变名参数，如k=20

def DarknetConv2D(*args, **kwargs):
    """Wrapper to set Darknet parameters for Convolution2D."""
    darknet_conv_kwargs = {'kernel_regularizer': l2(5e-4)}
    darknet_conv_kwargs['padding'] = 'valid' if kwargs.get('strides')==(2,2) else 'same'
    darknet_conv_kwargs.update(kwargs)
    return Conv2D(*args, **darknet_conv_kwargs)

def DarknetConv2D_BN_Leaky(*args, **kwargs):
    """Darknet Convolution2D followed by BatchNormalization and LeakyReLU."""
    no_bias_kwargs = {'use_bias': False}
    no_bias_kwargs.update(kwargs)
    return compose(
        DarknetConv2D(*args, **no_bias_kwargs),
        BatchNormalization(),
        LeakyReLU(alpha=0.1))

def resblock_body(x, num_filters, num_blocks):
    '''A series of resblocks starting with a downsampling Convolution2D'''
    # Darknet uses left and top padding instead of 'same' mode
    x = ZeroPadding2D(((1,0),(1,0)))(x)
    x = DarknetConv2D_BN_Leaky(num_filters, (3,3), strides=(2,2))(x)
    for i in range(num_blocks):
        y = compose(
                DarknetConv2D_BN_Leaky(num_filters//2, (1,1)),
                DarknetConv2D_BN_Leaky(num_filters, (3,3)))(x)
        x = Add()([x,y])
    return x

 在此基础上darknet很简洁

def darknet_body(x):
    '''Darknent body having 52 Convolution2D layers'''
    x = DarknetConv2D_BN_Leaky(32, (3,3))(x)
    x = resblock_body(x, 64, 1)
    x = resblock_body(x, 128, 2)
    x = resblock_body(x, 256, 8)#92层
    x = resblock_body(x, 512, 8)#152层
    x = resblock_body(x, 1024, 4)
    return x

右下角部分，先是定义make_last_layers，四层和六层卷积

def make_last_layers(x, num_filters, out_filters):
    '''6 Conv2D_BN_Leaky layers followed by a Conv2D_linear layer'''
    x = compose(
            DarknetConv2D_BN_Leaky(num_filters, (1,1)),
            DarknetConv2D_BN_Leaky(num_filters*2, (3,3)),
            DarknetConv2D_BN_Leaky(num_filters, (1,1)),
            DarknetConv2D_BN_Leaky(num_filters*2, (3,3)),
            DarknetConv2D_BN_Leaky(num_filters, (1,1)))(x)
    y = compose(
            DarknetConv2D_BN_Leaky(num_filters*2, (3,3)),
            DarknetConv2D(out_filters, (1,1)))(x)
    return x, y

利用make_last_layers定义yolo_body，scale1处四层conv得到x，x再通过两层conv得到y1,另一方面x通过一层conv和x2的上采样与scale2处merge，经过4层conv得到x,x再通过两层conv得到y2,同理得到y3，最后输出y1,y2,y3用作预测。

yolo_head按照如下公式，(feats[..., :])中0到3是tx,ty,tw,th，4是confidence，之后是class_probs，grid是cx,cy,K.title是在维度上进行复制操作。

def yolo_head(feats, anchors, num_classes, input_shape, calc_loss=False):
    """Convert final layer features to bounding box parameters."""
    num_anchors = len(anchors)
    # Reshape to batch, height, width, num_anchors, box_params.
    anchors_tensor = K.reshape(K.constant(anchors), [1, 1, 1, num_anchors, 2])

    grid_shape = K.shape(feats)[1:3] # height, width
    grid_y = K.tile(K.reshape(K.arange(0, stop=grid_shape[0]), [-1, 1, 1, 1]),
        [1, grid_shape[1], 1, 1])
    grid_x = K.tile(K.reshape(K.arange(0, stop=grid_shape[1]), [1, -1, 1, 1]),
        [grid_shape[0], 1, 1, 1])
    grid = K.concatenate([grid_x, grid_y])
    grid = K.cast(grid, K.dtype(feats))

    feats = K.reshape(
        feats, [-1, grid_shape[0], grid_shape[1], num_anchors, num_classes + 5])

    # Adjust preditions to each spatial grid point and anchor size.
    box_xy = (K.sigmoid(feats[..., :2]) + grid) / K.cast(grid_shape[::-1], K.dtype(feats))
    box_wh = K.exp(feats[..., 2:4]) * anchors_tensor / K.cast(input_shape[::-1], K.dtype(feats))
    box_confidence = K.sigmoid(feats[..., 4:5])
    box_class_probs = K.sigmoid(feats[..., 5:])

    if calc_loss == True:
        return grid, feats, box_xy, box_wh
    return box_xy, box_wh, box_confidence, box_class_probs

yolo_correct_boxes是将训练时用的(416,416)得出的bbox转化成原图size的bbox

def yolo_correct_boxes(box_xy, box_wh, input_shape, image_shape):
    '''Get corrected boxes'''
    box_yx = box_xy[..., ::-1]
    box_hw = box_wh[..., ::-1]
    input_shape = K.cast(input_shape, K.dtype(box_yx))
    image_shape = K.cast(image_shape, K.dtype(box_yx))
    new_shape = K.round(image_shape * K.min(input_shape/image_shape))
    offset = (input_shape-new_shape)/2./input_shape
    scale = input_shape/new_shape
    box_yx = (box_yx - offset) * scale
    box_hw *= scale

    box_mins = box_yx - (box_hw / 2.)
    box_maxes = box_yx + (box_hw / 2.)
    boxes =  K.concatenate([
        box_mins[..., 0:1],  # y_min
        box_mins[..., 1:2],  # x_min
        box_maxes[..., 0:1],  # y_max
        box_maxes[..., 1:2]  # x_max
    ])

    # Scale boxes back to original image shape.
    boxes *= K.concatenate([image_shape, image_shape])
    return boxes

在preprocess_true_boxes中，true_boxes批次数m，最大框数T，每个框5个值，y_true的第0和1位是中心点xy，范围是(0~13/26/52)，第2和3位是宽高wh，范围是0~1，第4位是置信度1或0，第5~n位是类别为1其余为0。

def preprocess_true_boxes(true_boxes, input_shape, anchors, num_classes):
    '''Preprocess true boxes to training input format

    Parameters
    ----------
    true_boxes: array, shape=(m, T, 5)
        Absolute x_min, y_min, x_max, y_max, class_id relative to input_shape.
    input_shape: array-like, hw, multiples of 32
    anchors: array, shape=(N, 2), wh
    num_classes: integer

    Returns
    -------
    y_true: list of array, shape like yolo_outputs, xywh are reletive value

    '''
    assert (true_boxes[..., 4]<num_classes).all(), 'class id must be less than num_classes'
    num_layers = len(anchors)//3 # default setting
    anchor_mask = [[6,7,8], [3,4,5], [0,1,2]] if num_layers==3 else [[3,4,5], [1,2,3]]

    true_boxes = np.array(true_boxes, dtype='float32')
    input_shape = np.array(input_shape, dtype='int32')
    boxes_xy = (true_boxes[..., 0:2] + true_boxes[..., 2:4]) // 2
    boxes_wh = true_boxes[..., 2:4] - true_boxes[..., 0:2]
    true_boxes[..., 0:2] = boxes_xy/input_shape[::-1]
    true_boxes[..., 2:4] = boxes_wh/input_shape[::-1]#除以416归一化

    m = true_boxes.shape[0]
    grid_shapes = [input_shape//{0:32, 1:16, 2:8}[l] for l in range(num_layers)]
                             #{0:32, 1:16, 2:8}是个字典，grid的3层大小13,26,52
    y_true = [np.zeros((m,grid_shapes[l][0],grid_shapes[l][1],len(anchor_mask[l]),5+num_classes),
        dtype='float32') for l in range(num_layers)]

    # Expand dim to apply broadcasting.
    anchors = np.expand_dims(anchors, 0)#扩充第一维,由（9,2)到（1,9,2)
    anchor_maxes = anchors / 2.
    anchor_mins = -anchor_maxes
    valid_mask = boxes_wh[..., 0]>0#存在box的为true

    for b in range(m):
        # Discard zero rows.
        wh = boxes_wh[b, valid_mask[b]]#对于每个batch中，任何一个框存在box，则得到其坐标
        if len(wh)==0: continue
        # Expand dim to apply broadcasting.
        wh = np.expand_dims(wh, -2)#扩充倒数第二维，由（num，2)到（num，1,2)
        box_maxes = wh / 2.
        box_mins = -box_maxes

        intersect_mins = np.maximum(box_mins, anchor_mins)#得到的是（num，9,2)
        intersect_maxes = np.minimum(box_maxes, anchor_maxes)
        intersect_wh = np.maximum(intersect_maxes - intersect_mins, 0.)
        intersect_area = intersect_wh[..., 0] * intersect_wh[..., 1]
        box_area = wh[..., 0] * wh[..., 1]
        anchor_area = anchors[..., 0] * anchors[..., 1]
        iou = intersect_area / (box_area + anchor_area - intersect_area)#（num，9)

        # Find best anchor for each true box
        best_anchor = np.argmax(iou, axis=-1)

        for t, n in enumerate(best_anchor):#t是box的序号；n是最优anchor的序号；l是层号；
            for l in range(num_layers):
                if n in anchor_mask[l]:
                    i = np.floor(true_boxes[b,t,0]*grid_shapes[l][1]).astype('int32')
                    j = np.floor(true_boxes[b,t,1]*grid_shapes[l][0]).astype('int32')
                    k = anchor_mask[l].index(n)
                    c = true_boxes[b,t, 4].astype('int32')
                    y_true[l][b, j, i, k, 0:4] = true_boxes[b,t, 0:4]
                    y_true[l][b, j, i, k, 4] = 1
                    y_true[l][b, j, i, k, 5+c] = 1

    return y_true

对于IOU的计算与preprocess_true_boxes中IOU的计算类似

def box_iou(b1, b2):
    '''Return iou tensor

    Parameters
    ----------
    b1: tensor, shape=(i1,...,iN, 4), xywh
    b2: tensor, shape=(j, 4), xywh

    Returns
    -------
    iou: tensor, shape=(i1,...,iN, j)

    '''

    # Expand dim to apply broadcasting.
    b1 = K.expand_dims(b1, -2)
    b1_xy = b1[..., :2]
    b1_wh = b1[..., 2:4]
    b1_wh_half = b1_wh/2.
    b1_mins = b1_xy - b1_wh_half
    b1_maxes = b1_xy + b1_wh_half

    # Expand dim to apply broadcasting.
    b2 = K.expand_dims(b2, 0)
    b2_xy = b2[..., :2]
    b2_wh = b2[..., 2:4]
    b2_wh_half = b2_wh/2.
    b2_mins = b2_xy - b2_wh_half
    b2_maxes = b2_xy + b2_wh_half

    intersect_mins = K.maximum(b1_mins, b2_mins)
    intersect_maxes = K.minimum(b1_maxes, b2_maxes)
    intersect_wh = K.maximum(intersect_maxes - intersect_mins, 0.)
    intersect_area = intersect_wh[..., 0] * intersect_wh[..., 1]
    b1_area = b1_wh[..., 0] * b1_wh[..., 1]
    b2_area = b2_wh[..., 0] * b2_wh[..., 1]
    iou = intersect_area / (b1_area + b2_area - intersect_area)

    return iou

args中是yolo_outputs和y_true的组合，两者格式一致，都是[(?, 13, 13, 3*85), (?, 26, 26, 3*85), (?, 52, 52, 3*85)]，前者是网络的预测值，后者是真值。input_shape是[13,13]*32也就是416*416，grid_shape就是[(13,13),(26,26),(52,52)]。mf就是batchsize。整体的损失函数可以参照yolo1中的损失函数。

 但在yolo3中，作者将平方损失改成了交叉熵的损失。权值取得是2-w*h

def yolo_loss(args, anchors, num_classes, ignore_thresh=.5, print_loss=False):
    '''Return yolo_loss tensor

    Parameters
    ----------
    yolo_outputs: list of tensor, the output of yolo_body or tiny_yolo_body
    y_true: list of array, the output of preprocess_true_boxes
    anchors: array, shape=(N, 2), wh
    num_classes: integer
    ignore_thresh: float, the iou threshold whether to ignore object confidence loss

    Returns
    -------
    loss: tensor, shape=(1,)

    '''
    num_layers = len(anchors)//3 # default setting
    yolo_outputs = args[:num_layers]
    y_true = args[num_layers:]
    anchor_mask = [[6,7,8], [3,4,5], [0,1,2]] if num_layers==3 else [[3,4,5], [1,2,3]]
    input_shape = K.cast(K.shape(yolo_outputs[0])[1:3] * 32, K.dtype(y_true[0]))
    grid_shapes = [K.cast(K.shape(yolo_outputs[l])[1:3], K.dtype(y_true[0])) for l in range(num_layers)]
    loss = 0
    m = K.shape(yolo_outputs[0])[0] # batch size, tensor
    mf = K.cast(m, K.dtype(yolo_outputs[0]))

    for l in range(num_layers):
        object_mask = y_true[l][..., 4:5]#是否有目标的置信度
        true_class_probs = y_true[l][..., 5:]#在有目标情况下，各类别出现的概率

        grid, raw_pred, pred_xy, pred_wh = yolo_head(yolo_outputs[l],
             anchors[anchor_mask[l]], num_classes, input_shape, calc_loss=True)
        pred_box = K.concatenate([pred_xy, pred_wh])

        # Darknet raw box to calculate loss.
        raw_true_xy = y_true[l][..., :2]*grid_shapes[l][::-1] - grid#与yolo_head对比来看
        raw_true_wh = K.log(y_true[l][..., 2:4] / anchors[anchor_mask[l]] * input_shape[::-1])
        raw_true_wh = K.switch(object_mask, raw_true_wh, K.zeros_like(raw_true_wh)) # avoid log(0)=-inf
        box_loss_scale = 2 - y_true[l][...,2:3]*y_true[l][...,3:4]#2-w*h

        # Find ignore mask, iterate over each of batch.
        ignore_mask = tf.TensorArray(K.dtype(y_true[0]), size=1, dynamic_size=True)
        object_mask_bool = K.cast(object_mask, 'bool')
        def loop_body(b, ignore_mask):
            true_box = tf.boolean_mask(y_true[l][b,...,0:4], object_mask_bool[b,...,0])
            iou = box_iou(pred_box[b], true_box)
            best_iou = K.max(iou, axis=-1)
            ignore_mask = ignore_mask.write(b, K.cast(best_iou<ignore_thresh, K.dtype(true_box)))
            return b+1, ignore_mask
        _, ignore_mask = K.control_flow_ops.while_loop(lambda b,*args: b<m, loop_body, [0, ignore_mask])
        ignore_mask = ignore_mask.stack()
        ignore_mask = K.expand_dims(ignore_mask, -1)#略去不是极大值IOU对应的anchor，维度是（bachsize，grid，grid，3,1)

        # K.binary_crossentropy is helpful to avoid exp overflow.
        xy_loss = object_mask * box_loss_scale * K.binary_crossentropy(raw_true_xy, raw_pred[...,0:2], from_logits=True)
        wh_loss = object_mask * box_loss_scale * 0.5 * K.square(raw_true_wh-raw_pred[...,2:4])
        confidence_loss = object_mask * K.binary_crossentropy(object_mask, raw_pred[...,4:5], from_logits=True)+ \
            (1-object_mask) * K.binary_crossentropy(object_mask, raw_pred[...,4:5], from_logits=True) * ignore_mask
        class_loss = object_mask * K.binary_crossentropy(true_class_probs, raw_pred[...,5:], from_logits=True)

        xy_loss = K.sum(xy_loss) / mf
        wh_loss = K.sum(wh_loss) / mf
        confidence_loss = K.sum(confidence_loss) / mf
        class_loss = K.sum(class_loss) / mf
        loss += xy_loss + wh_loss + confidence_loss + class_loss
        if print_loss:
            loss = tf.Print(loss, [loss, xy_loss, wh_loss, confidence_loss, class_loss, K.sum(ignore_mask)], message='loss: ')
    return loss