What does it mean if test and train accuracy is oscillating too much for CNN model in tensorflow?



I have implemented a CNN with images as input and 2 classes as output. I have applied mean subtraction to the input for normalisation before giving it as input to the network. But on compiling the code multiple times, there testing accuracy seems to be too unstable as I have in the pastebins


Testing Accuracy – > 0.40

Testing Accuracy – > 0.57 https://pastebin.com/e2LxssAB

Testing Accuracy – > 0.77 https://pastebin.com/WiiGW0nS

I am TRAINING on 1028 samples and testing on 300 samples