了解darknet的yolo.cfg配置文件

Question

我在互联网上搜索但发现这方面的信息很少，我不明白yolo的.cfg文件中每个变量/值的含义。所以我希望你们中的一些人可以提供帮助，我不认为我是唯一一个有这个问题的人，所以如果有人知道2或3个变量，请发布它们，以便将来需要这些信息的人可以找到它们。

我想知道的主要是：

• 批次

• 细分

• 衰变

• 动量

• 信道

• 过滤器

• 激活

Answer 1

这是我目前对一些变量的理解。虽然不一定正确：

[净]

• batch：在前向传递中使用了许多图像+标签来计算渐变并通过反向传播更新权重。
• 细分：批次细分为这么多“块”。块的图像在gpu上并行运行。
• 衰变：可能是一个减少重量的术语，以避免产生大的值。出于稳定性原因，我猜。
• 频道：在此图片中有更好的解释：

在左侧，我们有一个4x4像素的单通道，重组图层将大小减小到一半，然后在不同通道中创建4个通道与相邻像素。

• 动量：我猜新梯度是由动量计算 * previous_gradient +（1- 动量）* gradient_of_current_batch 。使渐变更稳定。
• adam：使用adam优化工具？
虽然对我不起作用
• burn_in：对于前x个批次，慢慢增加学习率直到其最终值（您的 learning_rate 参数值）。使用此方法通过监控确定学习率，直到损失减少的值（在开始分歧之前）。
• policy = steps：使用以下步骤和比例参数调整培训期间的学习率
• steps = 500,1000：调整500和1000批后的学习率
• scale = 0.1,0.2：在500之后，将LR乘以0.1，然后再乘以1000再乘以0.2
• angle：通过旋转增大图像直到此角度（以度为单位）

层

• 过滤器：层中有多少个卷积内核。
• 激活：激活功能，relu，漏电relu等。参见src / activations.h
• stopbackward：只进行反向传播直到此层。将它放在第一个yolo层之前的倒数第二个卷积层中，以仅训练其后面的层，例如，当使用预训练的重量时。
• random：放入yolo图层。如果设置为1，则通过每隔几批将图像调整为不同大小来进行数据增强。用于概括对象大小。

许多事情或多或少都是不言自明的（大小，步幅，batch_normalize，max_batches，宽度，高度）。如果您有更多问题，请随时发表评论。

请再次注意，我并非百分之百确定其中的许多内容。

Answer 2

有关cfg参数的更完整说明，摘自YOLO v4 https://github.com/AlexeyAB/darknet/wiki/CFG-Parameters-in-the-%5Bnet%5D-section的作者 和https://github.com/AlexeyAB/darknet/wiki/CFG-Parameters-in-the-different-layers

下面只是文档的快照，请参考上面的链接以获取更好的格式

CFG-Parameters in the [net] section:

[net] section
batch=1 - number of samples (images, letters, ...) which will be precossed in one batch

subdivisions=1 - number of mini_batches in one batch, size mini_batch = batch/subdivisions, so GPU processes mini_batch samples at once, and the weights will be updated for batch samples (1 iteration processes batch images)

width=416 - network size (width), so every image will be resized to the network size during Training and Detection

height=416 - network size (height), so every image will be resized to the network size during Training and Detection

channels=3 - network size (channels), so every image will be converted to this number of channels during Training and Detection

inputs=256 - network size (inputs) is used for non-image data: letters, prices, any custom data

max_chart_loss=20 - max value of Loss in the image chart.png

For training only

Contrastive loss:

contrastive=1 - use Supervised contrastive loss for training Classifier (should be used with [contrastive] layer)

unsupervised=1 - use Unsupervised contrastive loss for training Classifier on images without labels (should be used with contrastive=1 parameter and with [contrastive] layer)

Data augmentation:

angle=0 - randomly rotates images during training (classification only)

saturation = 1.5 - randomly changes saturation of images during training

exposure = 1.5 - randomly changes exposure (brightness) during training

hue=.1 - randomly changes hue (color) during training https://en.wikipedia.org/wiki/HSL_and_HSV

blur=1 - blur will be applied randomly in 50% of the time: if 1 - will be blured background except objects with blur_kernel=31, if >1 - will be blured whole image with blur_kernel=blur (only for detection and if OpenCV is used)

min_crop=224 - minimum size of randomly cropped image (classification only)

max_crop=448 - maximum size of randomly cropped image (classification only)

aspect=.75 - aspect ration can be changed during croping from 0.75 - to 1/0.75 (classification only)

letter_box=1 - keeps aspect ratio of loaded images during training (detection training only, but to use it during detection-inference - use flag -letter_box at the end of detection command)

cutmix=1 - use CutMix data augmentation (for Classifier only, not for Detector)

mosaic=1 - use Mosaic data augmentation (4 images in one)

mosaic_bound=1 - limits the size of objects when mosaic=1 is used (does not allow bounding boxes to leave the borders of their images when Mosaic-data-augmentation is used)

data augmentation in the last [yolo]-layer

jitter=0.3 - randomly changes size of image and its aspect ratio from x(1 - 2*jitter) to x(1 + 2*jitter)
random=1 - randomly resizes network size after each 10 batches (iterations) from /1.4 to x1.4 with keeping initial aspect ratio of network size
adversarial_lr=1.0 - Changes all detected objects to make it unlike themselves from neural network point of view. The neural network do an adversarial attack on itself

attention=1 - shows points of attention during training

gaussian_noise=1 - add gaussian noise

Optimizator:

momentum=0.9 - accumulation of movement, how much the history affects the further change of weights (optimizer)

decay=0.0005 - a weaker updating of the weights for typical features, it eliminates dysbalance in dataset (optimizer) http://cs231n.github.io/neural-networks-3/

learning_rate=0.001 - initial learning rate for training

burn_in=1000 - initial burn_in will be processed for the first 1000 iterations, current_learning rate = learning_rate * pow(iterations / burn_in, power) = 0.001 * pow(iterations/1000, 4) where is power=4 by default

max_batches = 500200 - the training will be processed for this number of iterations (batches)

policy=steps - policy for changing learning rate: constant (by default), sgdr, steps, step, sig, exp, poly, random (f.e., if policy=random - then current learning rate will be changed in this way = learning_rate * pow(rand_uniform(0,1), power))

power=4 - if policy=poly - the learning rate will be = learning_rate * pow(1 - current_iteration / max_batches, power)

sgdr_cycle=1000 - if policy=sgdr - the initial number of iterations in cosine-cycle

sgdr_mult=2 - if policy=sgdr - multiplier for cosine-cycle https://towardsdatascience.com/https-medium-com-reina-wang-tw-stochastic-gradient-descent-with-restarts-5f511975163

steps=8000,9000,12000 - if policy=steps - at these numbers of iterations the learning rate will be multiplied by scales factor

scales=.1,.1,.1 - if policy=steps - f.e. if steps=8000,9000,12000, scales=.1,.1,.1 and the current iteration number is 10000 then current_learning_rate = learning_rate * scales[0] * scales[1] = 0.001 * 0.1 * 0.1 = 0.00001

label_smooth_eps=0.1 - use label smoothing for training Classifier

For training Recurrent networks:

Object Detection/Tracking on Video - if [conv-lstm] or [crnn] layers are used in additional to [connected] and [convolutional] layers

Text generation - if [lstm] or [rnn] layers are used in additional to [connected] layers

track=1 - if is set 1 then the training will be performed in Recurrents-tyle for image sequences

time_steps=16 - training will be performed for a random image sequence that contains 16 images from train.txt file

for [convolutional]-layers: mini_batch = time_steps*batch/subdivisions
for [conv_lstm]-recurrent-layers: mini_batch = batch/subdivisions and sequence=16
augment_speed=3 - if set 3 then can be used each 1st, 2nd or 3rd image randomly, i.e. can be used 16 images with indexes 0, 1, 2, ... 15 or 110, 113, 116, ... 155 from train.txt file

sequential_subdivisions=8 - lower value increases the sequence of images, so if time_steps=16 batch=16 sequential_subdivisions=8, then will be loaded time_steps*batch/sequential_subdivisions = 16*16/8 = 32 sequential images with the same data-augmentation, so the model will be trained for sequence of 32 video-frames

seq_scales=0.5, 0.5 - increasing sequence of images at some steps, i.e. the coefficients to which the original sequential_subdivisions value will be multiplied (and batch will be dividied, so the weights will be updated rarely) at correspond steps if is used policy=steps or policy=sgdr

CFG-Parameters in the different layers
Image processing [N x C x H x W]:

[convolutional] - convolutional layer

batch_normalize=1 - if 1 - will be used batch-normalization, if 0 will not (0 by default)

filters=64 - number of kernel-filters (1 by default)

size=3 - kernel_size of filter (1 by default)

groups = 32 - number of groups for grouped-convolutional (depth-wise) (1 by default)

stride=1 - stride (offset step) of kernel filter (1 by default)

padding=1 - size of padding (0 by default)

pad=1 - if 1 will be used padding = size/2, if 0 the will be used parameter padding= (0 by default)

dilation=1 - size of dilation (1 by default)

activation=leaky - activation function after convolution: logistic (by default), loggy, relu, elu, selu, relie, plse, hardtan, lhtan, linear, ramp, leaky, tanh, stair, relu6, swish, mish

[activation] - separate activation layer

activation=leaky - activation function: linear (by default), loggy, relu, elu, selu, relie, plse, hardtan, lhtan, linear, ramp, leaky, tanh, stair
[batchnorm] - separate Batch-normalization layer
[maxpool] - max-pooling layer (the maximum value)

size=2 - size of max-pooling kernel

stride=2 - stirde (offset step) of max-pooling kernel

[avgpool] - average pooling layer input W x H x C -> output 1 x 1 x C
[shortcut] - residual connection (ResNet)

from=-3,-5 - relative layer numbers, preforms element-wise adding of several layers: previous-layer and layers specified in from= parameter

weights_type=per_feature - will be used weights for shortcut y[i] = w1*layer1[i] + w2*layer2[i] ...

per_feature - 1 weights per layer/feature
per_channel - 1 weights per channel
none - weights will not be used (by default)
weights_normalization=softmax - will be used weights normalization

softmax - softmax normalization
relu - relu normalization
none - without weights normalization - unbound weights (by default)
activation=linear - activation function after shortcut/residual connection (linear by default)

[upsample] - upsample layer (increase W x H resolution of input by duplicating elements)

stride=2 - factor for increasing both Width and Height (new_w = w*stride, new_h = h*stride)
[scale_channels] - scales channels (SE: squeeze-and-excitation blocks) or (ASFF: adaptively spatial feature fusion) -it multiplies elements of one layer by elements of another layer

from=-3 - relative layer number, performs multiplication of all elements of channel N from layer -3, by one element of channel N from the previous layer -1 (i.e. for(int i=0; i < b*c*h*w; ++i) output[i] = from_layer[i] * previous_layer[i/(w*h)]; )

scale_wh=0 - SE-layer (previous layer 1x1xC), scale_wh=1 - ASFF-layer (previous layer WxHx1)

activation=linear - activation function after scale_channels-layer (linear by default)

[sam] - Spatial Attention Module (SAM) - it multiplies elements of one layer by elements of another layer

from=-3 - relative layer number (this and previous layers should be the same size WxHxC)
[reorg3d] - reorg layer (resize W x H x C)

stride=2 - if reverse=0 input will be resized to W/2 x H/2 x C4, if reverse=1thenW2 x H*2 x C/4`, (1 by default)

reverse=1 - if 0(by default) then decrease WxH, if1thenincrease WxH (0 by default)

[reorg] - OLD reorg layer from Yolo v2 - has incorrect logic (resize W x H x C) - depracated

stride=2 - if reverse=0 input will be resized to W/2 x H/2 x C4, if reverse=1thenW2 x H*2 x C/4`, (1 by default)

reverse=1 - if 0(by default) then decrease WxH, if1thenincrease WxH (0 by default)

[route] - concatenation layer, Concat for several input-layers, or Identity for one input-layer

layers = -1, 61 - layers that will be concatenated, output: W x H x C_layer_1 + C_layer_2
if index < 0, then it is relative layer number (-1 means previous layer)
if index >= 0, then it is absolute layer number
[yolo] - detection layer for Yolo v3 / v4

mask = 3,4,5 - indexes of anchors which are used in this [yolo]-layer

anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326 - initial sizes if bounded_boxes that will be adjusted

num=9 - total number of anchors

classes=80 - number of classes of objects which can be detected

ignore_thresh = .7 - keeps duplicated detections if IoU(detect, truth) > ignore_thresh, which will be fused during NMS (is used for training only)

truth_thresh = 1 - adjusts duplicated detections if IoU(detect, truth) > truth_thresh, which will be fused during NMS (is used for training only)

jitter=.3 - randomly crops and resizes images with changing aspect ratio from x(1 - 2*jitter) to x(1 + 2*jitter) (data augmentation parameter is used only from the last layer)

random=1 - randomly resizes network for each 10 iterations from 1/1.4 to 1.4(data augmentation parameter is used only from the last layer)

resize=1.5 - randomly resizes image in range: 1/1.5 - 1.5x

max=200 - maximum number of objects per image during training

counters_per_class=100,10,1000 - number of objects per class in Training dataset to eliminate the imbalance

label_smooth_eps=0.1 - label smoothing

scale_x_y=1.05 - eliminate grid sensitivity

iou_thresh=0.2 - use many anchors per object if IoU(Obj, Anchor) > 0.2

iou_loss=mse - IoU-loss: mse, giou, diou, ciou

iou_normalizer=0.07 - normalizer for delta-IoU

cls_normalizer=1.0 - normalizer for delta-Objectness

max_delta=5 - limits delta for each entry

[crnn] - convolutional RNN-layer (recurrent)

batch_normalize=1 - if 1 - will be used batch-normalization, if 0 will not (0 by default)

size=1 - convolutional kernel_size of filter (1 by default)

pad=0 - if 1 will be used padding = size/2, if 0 the will be used parameter padding= (0 by default)

output = 1024 - number of kernel-filters in one output convolutional layer (1 by default)

hidden=1024 - number of kernel-filters in two (input and hidden) convolutional layers (1 by default)

activation=leaky - activation function for each of 3 convolutional-layers in the [crnn]-layer (logistic by default)

[conv_lstm] - convolutional LSTM-layer (recurrent)

batch_normalize=1 - if 1 - will be used batch-normalization, if 0 will not (0 by default)

size=3 - convolutional kernel_size of filter (1 by default)

padding=1 - convolutional size of padding (0 by default)

pad=1 - if 1 will be used padding = size/2, if 0 the will be used parameter padding= (by default)

stride=1 - convolutional stride (offset step) of kernel filter (1 by default)

dilation=1 - convolutional size of dilation (1 by default)

output=256 - number of kernel-filters in each of 8 or 11 convolutional layers (1 by default)

groups=4 - number of groups for grouped-convolutional (depth-wise) (1 by default)

state_constrain=512 - constrains LSTM-state values [-512; +512] after each inference (time_steps*32 by default)

peephole=0 - if 1 then will be used Peephole (additional 3 conv-layers), if 0 will not (1 by default)

bottleneck=0 - if 1 then will be used reduced optimal versionn of conv-lstm layer

activation=leaky - activation function for each of 8 or 11 convolutional-layers in the [conv_lstm]-layer (linear by default)

lstm_activation=tanh - activation for G (gate: g = tanh(wg + ug)) and C (memory cell: h = o * tanh(c))

Detailed-architecture-of-the-peephole-LSTM

Free-form data processing [Inputs]:

[connected] - fully connected layer
output=256 - number of outputs (1 by default), so number of connections is equal to inputs*outputs
activation=leaky - activation after layer (logistic by default)
[dropout] - dropout layer
probability=0.5 - dropout probability - what part of inputs will be zeroed (0.5 = 50% by default)

dropblock=1 - use as DropBlock

dropblock_size_abs=7 - size of DropBlock in pixels 7x7

[softmax] - SoftMax CE (cross entropy) layer - Categorical cross-entropy for multi-class classification
[contrastive] - Contrastive loss layer for Supervised and Unsupervised learning (should be set [net] contrastive=1 and optionally [net] unsupervised=1)

classes=1000 - number of classes

temperature=1.0 - temperature

[cost] - cost layer calculates (linear)Delta and (squared)Loss
type=sse - cost type: sse (L2), masked, smooth (smooth-L1) (SSE by default)
[rnn] - fully connected RNN-layer (recurrent)
batch_normalize=1 - if 1 - will be used batch-normalization, if 0 will not (0 by default)
output = 1024 - number of outputs in one connected layer (1 by default)
hidden=1024 - number of outputs in two (input and hidden) connected layers (1 by default)
activation=leaky - activation after layer (logistic by default)
[lstm] - fully connected LSTM-layer (recurrent)
batch_normalize=1 - if 1 - will be used batch-normalization, if 0 will not (0 by default)
output = 1024 - number of outputs in all connected layers (1 by default)
[gru] - fully connected GRU-layer (recurrent)
batch_normalize=1 - if 1 - will be used batch-normalization, if 0 will not (0 by default)
output = 1024 - number of outputs in all 

connected layers (1 by default)

Answer 3

批处理，为减少丢失而在每批处理中选择的图像数量

细分将批量大小划分为no。批处理的子批处理

衰变是一个学习参数，并且在日记中指定使用的动量为0.9，衰减为0.0005

动量是一个学习参数，并且在期刊中指定使用0.9的动量和0.0005的衰减

通道通道是指BGR图像输入图像（3）的通道大小

过滤器用于CNN算法的过滤器数量

激活：CNN的激活功能：大多数使用泄漏的RELU功能（我在配置文件中看到的大部分内容）

Answer 4

尽管这是一个很老的帮助请求，但对于希望寻找答案的未来用户， 您可以在Wiki页面上的原始Yolo项目最著名的分支中找到所有说明。 https://github.com/AlexeyAB/darknet/wiki

尤其是，仅从here复制并粘贴[net]部分，如下所示：

  

[net]

     

  
• batch=1-将在一批样品中进行胶印的样品数量（图像，字母，...）
  
• subdivisions=1-一批中的mini_batches数量，大小为mini_batch = batch/subdivisions，因此GPU一次处理mini_batch个样本，权重将更新batch个样本（1迭代处理batch张图片）
  
• width=416-网络大小（宽度），因此在训练和检测过程中，每个图像的大小都会调整为网络大小
  
• height=416-网络大小（高度），因此在训练和检测期间，每个图像的大小都会调整为网络大小
  
• channels=3-网络大小（通道），因此在训练和检测过程中每张图像都将转换为该数量的通道
  
• inputs=256-网络大小（输入）用于非图像数据：字母，价格，任何自定义数据
  

无论如何，您甚至应该尝试在亲戚Github/issues part中寻找您想知道的内容，甚至是天真，因为通常已经有人问过并回答了。

祝你好运。