When running the minist tensorflow training demo, the more than 35200, the number of iterations will appear training accuracy and accuracy have been near 0.1 after fluctuations,
Do you have encountered a similar situation?

Step 34800, the training accuracy 1
Step 34900, the training accuracy 1
Step 35000, the training accuracy 1
Step 35100, the training accuracy 1
Step 35200, the training accuracy 0.06
Step 35300, the training accuracy 0.1
Step 35400, the training accuracy 0.06
Step 35500, the training accuracy 0.06
Step 35600, the training accuracy 0.12

CodePudding user response:

Is a fitting, typically a threshold and more than the threshold will appear the accuracy in test data sets is reduced,

CodePudding user response:

Vector is too big, too small, model assumption is in line with the data distribution, will affect the accuracy of training to further enhance,

CodePudding user response:

reference 1st floor ksx_120999 response:

is a fitting, typically a threshold and more than the threshold will appear the accuracy in test data sets is reduced,

More than the threshold, appeared fitting, accuracy will mutate? Accuracy in 0.1 or so, feeling and the probability of random guesses about the same,

CodePudding user response:

refer to the second floor of joy pig response:

vector is too big, too small, model assumption is in line with the data distribution, will affect the accuracy of training to further enhance,

This is to download the demo, the source code is to set up 20000 times, accuracy should also can also, should not be more pot

CodePudding user response:

reference dzyjjpy reply: 3/f

Quote: refer to 1st floor ksx_120999 response:

Is a fitting, typically a threshold and more than the threshold will appear the accuracy in test data sets is reduced,

More than the threshold, appeared fitting, accuracy will mutate? Accuracy at about 0.1, feeling is similar to the probability of random guesses,

Mutation is possible, mnist training, there are many online instructions,
Possibility:
1, may be the algorithm itself is wrong,
2, the training set and test set of the group has a problem, try not to take out the training test set
3, the number of training is not the more the better,

CodePudding user response:

reference 5 floor ksx_120999 reply:

Quote: refer to the third floor dzyjjpy response:

Quote: refer to 1st floor ksx_120999 response:

Is a fitting, typically a threshold and more than the threshold will appear the accuracy in test data sets is reduced,

More than the threshold, appeared fitting, accuracy will mutate? Accuracy at about 0.1, feeling is similar to the probability of random guesses,

Mutation is possible, mnist training, there are many online instructions,
Possibility:
1, may be the algorithm itself is wrong,
2, the training set and test set of the group has a problem, try not to take out the training test set
3, the number of training is not the more the better,

Ok, thank you, training set and test set, is completely separated,

CodePudding user response:

Hello, I want to ask, my training process, training accuracy is 1, but the testing accuracy has been the same, is 0.125, what is this

CodePudding user response:

Bloggers, want to consult the training accuracy when suddenly drop what reason be