Using TensorFlow training samples, what are the main factors that affect loss values?

problem description

recently, using TensorFlow training samples, the neural network chooses that the yolov3,loss value has been maintained at about 8, and there is no global convergence, and the final recognition rate is not high.

the environmental background of the problems and what methods you have tried

the loss of 8 is too large, so I want to optimize it, because I have little experience in machine learning, and I only summarize the following optimization points:

  • modify the activation function
  • modify the loss function
  • optimize training samples: increase the number of training sets; reduce interference items when marking

apart from these points, where else can we optimize? For the time being, I haven"t thought about setting the super-parameter, such as the step size when the gradient drops, and so on.


I don't know what dataset you use. I think the model is not expressive enough for your dataset, so you can try to change the network.

and why only look at loss? the main evaluation index of target detection should be mAP and so on.


how much is batch_size now? set it a little bit and try it, and the learning rate

.
MySQL Query : SELECT * FROM `codeshelper`.`v9_news` WHERE status=99 AND catid='6' ORDER BY rand() LIMIT 5
MySQL Error : Disk full (/tmp/#sql-temptable-64f5-7c0073-1f81e.MAI); waiting for someone to free some space... (errno: 28 "No space left on device")
MySQL Errno : 1021
Message : Disk full (/tmp/#sql-temptable-64f5-7c0073-1f81e.MAI); waiting for someone to free some space... (errno: 28 "No space left on device")
Need Help?