What is the mini-batch accuracy in CNN training?

hello~

when i train my CNN, log text appear in my command window.

for example,

my question is, what is the meaning of mini-batch accuracy value for each line?

is it average accuracy for every 50 iteration, or exactly for '50th' iteration?

thank you for reading~

Answer 1

[Sally Al Khamees]

The mini-batch accuracy reported during training corresponds to the accuracy of the particular mini-batch at the given iteration. It is not a running average over iterations. During training by stochastic gradient descent with momentum (SGDM), the algorithm groups the full dataset into disjoint mini-batches.

An iteration corresponds to the calculation of the network’s gradients for each mini-batch.

An epoch corresponds to moving through every available mini-batch.

Hope this helps

[jonghyun kim]

thank you very much!

[Ravish Raj]

What should be the overall accuracy for the training data as well as the testing data? I am asking for the optimum, for prediction of steering angle. Can you also suggest , how can I increase the accuracy ?

[shefali saxena]

Hello Ravish

Do you find any solution??

I am working on 1D(ECG Signal) with CNN model and the overall accuracy of my model is 75%

I have 40 records each record consists of 1x15000 data. My model consists of 15-22 layers.

how can I increase accuracy....

[Pendela Neelesh]

Hello @shefali saxena,

Train you model different kernel sizes. Validation accuracy can be low due to overfitting, try using dropouts in the model(if not included), add non linearity to the model by relu aactivation.

You can also use agumentation if the training dataset is small. Use feature extraction techniques(such as PCA, ICA etc) before training to decrease the computational complexity.