时间戳

首页 > 解决方案 > 在 Keras 中正确构建 YOLOv2 的 Classifier 模型（Darknet19）

python - 在 Keras 中正确构建 YOLOv2 的 Classifier 模型（Darknet19）

问题描述

我正在尝试阅读一篇论文并从头开始构建分类器部分，但我似乎遇到了一个错误，并且不确定我是否正确构建它：

这是论文： https ://pjreddie.com/media/files/papers/YOLO9000.pdf

import keras
from keras.layers import Conv2D, Input, concatenate
from keras.layers import LeakyReLU, MaxPooling2D, BatchNormalization,GlobalAveragePooling2D
from keras.models import Model
from keras.activations import softmax
from functools import partial

#First train body for image classification
#Then train head

new_conv = partial(Conv2D ,padding = "same")

def _base_block(out,x):
    "(3,3), Leaky, Batch"
    x =new_conv(out, (3,3))(x)
    x =LeakyReLU(alpha=0.1)(x)
    x =BatchNormalization()(x) 
    return x

def _block_1(out, x):     
    """
    output follows:
    out//2, out

    """
    x =  new_conv(out//2, (1,1))(x)
    x =LeakyReLU(alpha=0.1)(x)
    x = BatchNormalization()(x)
    x = _base_block(out,x)
    return x

def _block_2(out, x):    
    """
    output follows:
    out, out//2, out

    """  
    x =_base_block(out,x)
    x = _block_1(out, x)
    return x


def Darknet19():
    input_layer = Input((img_size, img_size, 3))
    x = _base_block(32,input_layer)
    x =  MaxPooling2D((2,2),strides = 2)(x)
    x = _base_block(64,x)
    x = MaxPooling2D((2,2),strides = 2)(x)
    x = _block_2(128, x)
    x = MaxPooling2D((2,2),strides = 2)(x)
    x = _block_2(256, x)
    x = MaxPooling2D((2,2),strides = 2)(x)
    x = _block_2(512, x)
    x = _block_1(512, x)
    x = MaxPooling2D((2,2),strides = 2)(x)
    x =_block_2(1024, x)
    x = _block_1(512, x)
    x = new_conv(1, (1,1), activation = "linear")(x)
    model = Model(inputs = input_layer, outputs = x)  
    return model

def Darknet_classifier():
    base_model = Darknet19()
    x = base_model.output
    x = GlobalAveragePooling2D()(x)
    output = softmax(x)
    model = Model(inputs = base_model.inputs, outputs = output)

    return model

img_size = 426 #multiple of 30
model = Darknet19()
model =Darknet_classifier()
print(model.summary())

我收到的错误：

'Found: ' + str(x)) ValueError: 模型的输出张量必须是 Keras 的输出Layer（因此保存过去的层元数据）。找到：Tensor("Softmax:0", shape=(?, 1), dtype=float32)

看来我不能从 GAP 转到 softmax。我对模型的解释不正确吗？

标签： pythonkerasdeep-learningconv-neural-network

解决方案

您没有正确应用 softmax，您可以使用该Activation层： 

from keras.layers import Activation

output = Activation("softmax")(x)

推荐阅读