TensorFlow just learning, in dealing with the problems found in the training images classification accuracy rate is very low, only 20%, the use of the code is http://www.cnblogs.com/denny402/p/6931338.html, is this blogger example code, identification data set is just my own data sets, are some signal image, image is 256 * 128, then use this example to pick the two kinds of signal accuracy is high, now identify 5 kinds of signals, accuracy is only 20%, training loss is decreased, but the accuracy basically remain unchanged, what reason be excuse me? Need from which aspects?

CodePudding user response:

2 choose 1 accuracy is about 50%, specification models complete failure, because a random guess there will be 50% of the time,
20% of probability to choose a random guess, you model the accuracy is only 20%, establish the model

CodePudding user response:

Ha, ha, ha, polite smile

CodePudding user response:

1 data set quantity too little? The number of the training set is the number of levels?

CodePudding user response:

2 test set of data and training set of data is the same, if the difference is very big, accuracy is very low!

CodePudding user response:

3 training epoch is too little?

CodePudding user response:

Ask you a question, I design the training accuracy of CNN model consistent with the ratio of positive and negative samples, such as positive and negative samples ratio of 1:3, the accuracy of 75, plus or minus a sample ratio is 1:1, the accuracy of % 50, and the loss in gradually decline, but accuracy is changeless, what reason is this excuse me, is dirty? Need to change the loss function?

CodePudding user response:

Initialize such as super parameter setting problem?

CodePudding user response:

You may have many reasons, such as:
1 network structure itself has a problem, you can't do the classification task,
2 your data sets and your network do not match, the clarity of the picture, the network layer will affect the classification effect of
3 according to different data sets, vector and optimizer need to try to appropriate value
4 is very important for the pretreatment of data sets, such as normalization, scaling, change the number range, balance, a lot of problems need to be considered, such as
If you need help, can communicate privately,

CodePudding user response:

references to the tenth floor ITMessager response:

your problem there could be many reasons, such as:
1 network structure itself has a problem, you can't do the classification task,
2 your data sets and your network do not match, the clarity of the picture, the network layer will affect the classification effect of
3 according to different data sets, vector and optimizer need to try to appropriate value
4 is very important for the pretreatment of data sets, such as normalization, scaling, change the number range, balance, a lot of problems need to be considered, such as
If you need help, can communicate privately,

binary classification problems, could you tell me if it is CNN direct reference others' binary classification code, prepares the data accuracy is always in the process of training at about 50%, said no training? Try to change parameters, etc., has no effect

CodePudding user response:

11 references Edward_graduate response:

binary classification problems, could you tell me if it is CNN direct reference others' binary classification code, prepares the data accuracy is always in the process of training at about 50%, said no training? Try to modify the above parameters, such as no results

can say so, you can analyze the curve of loss function, see if falling or difficult to convergence, or no convergence at all,

CodePudding user response:

refer to 12 floor ITMessager reply:

Quote: reference 11 floor Edward_graduate response:

Is if CNN binary classification problems, directly quote someone else's binary classification code, prepares the data accuracy is always in the process of training at about 50%, said no training? Try to modify the above parameters, such as no results

can say so, you can analyze the curve of loss function, see if falling or difficult to convergence, or no convergence at all,

loss function basically did not move, how want to consult with the error screen

CodePudding user response:

references to the tenth floor ITMessager response:

your problem there could be many reasons, such as:
1 network structure itself has a problem, you can't do the classification task,
2 your data sets and your network do not match, the clarity of the picture, the network layer will affect the classification effect of
3 according to different data sets, vector and optimizer need to try to appropriate value
4 is very important for the pretreatment of data sets, such as normalization, scaling, change the number range, balance, a lot of problems need to be considered, such as
If you need help, can communicate privately,

Hello, my blogger similar problems, your answer I feel very reasonable, can you private chat, I WeChat love_qinging thanks

CodePudding user response:

I now have a similar problem, and the second category is effective, can correctly to what, but with their own five classification is very poor, at the time of prediction, sometimes completely to a certain class, or directly to the output of various types of prediction probability is 0.2. Head all dizzy