I’m working in university on project that should do classification of leukemia cancer based on Protein sequence SNP in the form of Wavelet Coherence Transform scalograms. The datasets I have is about 312 images (unbalanced 212 for positive and 128 for negative).
The professor demand to work a Deep neural network (CNN) model and to have an accuracy above 90%
First I do the step of trial-observation and I build model with Convnet give me 0.624 after I use Transfer learning VGG16 and Resnet50 also gives me accuracy on both train and val 0.624.
I can’t find where is the fault ?
Because you have unbalanced classes you need to use precision, recall and F1 score also you might need to include more images on the training, be mindful on the ratio you split train/validation/testing!
After reviewing your Leukemia prototype notebook,
I would try exploring your model algorithm like reducing one convolution layer, removing dropout, using only one or two dense layer. Using different optimizer, reducing the epoch, also batch size for training and validation generator varies too much (try using same batch_size for both generator and see if there is any improvement), so try reducing the batch size for your training dataset.
I am sharing a link which also elaborate on how input shape in Keras. the input layer itself is not a layer, but a tensor. It’s the starting tensor you send to the first hidden layer. This tensor must have the same shape as your training data.
neural network - Keras input explanation: input_shape, units, batch_size, dim, etc - Stack Overflow)%20.
Regards
DP