Low accuracy of CNN

I reshaped my data to 4D (9x2000x10x20000) i.e height x width x channel x instances. and label 20000x1. As my matlab is not able to processed these 20000 at a time so i took first 1000 instances and evaluated my CNN model with 70:30 holdout. but i am only getting 20% caccuracy. and model is overfitting. Model is predicting on one class for all available classes. Data is unbalenced also.

Answer 1

[Aditya Patil]

You can use all the data by passing some data at a time, also called as a MiniBatch. See the MiniBatchSize option in traningoptions.

If the data is imbalanced due to taking small sample, then using all data should solve the issue. On the other hand, if the entire data is imbalanced, consider using an appropriate loss function, such as focalLossLayer. Alternately, consider undersampling the data with more samples, or augment the data to increase samples in class with less samples.