训练定制YOLO模型时损失归于NAN [英] Loss goes to NAN when training the custom YOLO model

问题描述

我使用Keras为YOLO实现了自定义损失函数和模型.我正在使用Tensorflow作为后端.

I implemented a custom loss function and model for YOLO using Keras. I am using Tensorflow as backend.

import pickle
import tensorflow as tf
import numpy as np 
import matplotlib.pyplot as plt 
from keras.models import Sequential,load_model
from keras.layers import Dense,Conv2D,Activation,MaxPooling2D,Flatten
import keras as k
from keras import optimizers
import cv2

batch=12

sess= tf.Session()

#loss function
def yolo_loss(yTrue,yPred):
    coord=5
    noobj=0.5
    L_noobj=1
    L_obj=1
    if yTrue[6] == 1: 
       L_obj=0

    if yTrue[5] == 1:
       L_noobj=0

    w=coord*L_obj*(tf.square([tf.sqrt(yTrue[2])-tf.sqrt(yPred[2])])) 
    h=coord*L_obj*(tf.square([yTrue[3]-yPred[3]]))
    x=coord*L_obj*(tf.square([yTrue[0]-yPred[0]]))
    y=coord*L_obj*(tf.square([yTrue[1]-yPred[1]])) 
    no_obj=noobj*L_noobj*(tf.square([yTrue[6]-yPred[6]])) 
    obj=L_obj*(tf.square([yTrue[5]-yPred[5]])) 
    clss=L_obj*(tf.square([yTrue[4]-yPred[4]]))
    loss=w+h+x+y+no_obj+obj+clss
    return loss

def custom_loss(yTrue,yPred):
    loss=None
    for a in range(batch):
        loss_per_sample=0
        for b in range(4):
            for c in range(4):
                loss_per_sample += yolo_loss(yTrue[a,b,c,0:],yPred[a,b,c,0:])  
        if loss == None:
            loss=tf.stack(loss_per_sample)
        else:
            x=tf.stack(loss_per_sample)
            loss=tf.concat([loss,x],0)

    loss=tf.reshape(loss,[-1,1])      
    return loss 

#load data and labels 
x_train=pickle.load(open('data_image.pickle','rb'))
y_train=pickle.load(open('data_label.pickle','rb'))
test=pickle.load(open('test_image.pickle','rb'))


# model
model=Sequential()

model.add(Conv2D(16,(7,7),input_shape=x_train.shape[1:],padding="same"))
model.add(Activation("relu"))
model.add(MaxPooling2D((2,2)))

model.add(Conv2D(32,(3,3),padding="same"))
model.add(Activation("relu"))
model.add(MaxPooling2D((2,2)))

model.add(Conv2D(64,(3,3),padding="same"))
model.add(Activation("relu"))
model.add(MaxPooling2D((2,2)))

model.add(Conv2D(128,(3,3),padding="same"))
model.add(Activation("relu"))
model.add(MaxPooling2D((2,2)))

model.add(Conv2D(512,(3,3),padding="same"))
model.add(Activation("relu"))
model.add(MaxPooling2D((2,2)))

model.add(Conv2D(512,(3,3),padding="same"))
model.add(Activation("relu"))

model.add(Conv2D(1024,(3,3),padding="same"))
model.add(Activation("relu"))

model.add(Conv2D(7,(3,3),padding="same"))
model.add(Activation("relu"))

adam = optimizers.adam(lr=0.001)
model.compile(loss=custom_loss,optimizer=adam,metrics=["accuracy"]) 


model.fit(x_train,y_train,batch_size=batch,epochs=100)

model.save('yolo.model')

当我训练模型时，损失值将变为NAN.但是在自定义损失函数的"W"和"h"中删除了tf.sqrt()之后，损失几乎为零.但是问题是边界框的"W"和"h"值始终为零.我认为tf.sqrt()函数中有一些东西.请有人能告诉我这是怎么回事.

When I train the model Loss value goes to NAN.but after I remove the tf.sqrt() from the "W" and "h" in Custom loss function Loss is almost come to zero. But the problem is "W" and "h" value of the bounding box is always zero. I think there something in tf.sqrt() function. Please can someone tell me what is going on here.

推荐答案

您正在最后一层中使用relu，这是不期望的.这可能会导致渐变的消失.

You are using relu in last layer, which is not expected. This may be causing dying gradients.

此外，在使用sqrt函数之前，请做一些检查，例如负值.

Also, do some checks before using sqrt function such as negative values.

model.add(Conv2D(7,(3,3),padding="same"))
model.add(Activation("relu"))

adam = optimizers.adam(lr=0.001)
model.compile(loss=custom_loss,optimizer=adam,metrics=["accuracy"]) 

这篇关于训练定制YOLO模型时损失归于NAN的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！