Why is the accuracy on the training set dropping in certain iterations of the training?

I'm training using an ANN where i'm using relu activation on all layers except the last one which has a sigmoid activation.

I only have a small amount of data 891 examples. but i'm training using an ANN i've put together using numpy and pandas just as a training exercise.

I've implemented a fuction that would test combinations of layer formations and learning rate that would iterate x amount of times and give me the best average so that i can take a informed guess about the best dimentions and lreaning rate.

However i've been notecing something weird. Some of the iteration have a very low accuracy as shown in the below images.

I'm using 'he' weight initalization

for l in range(1, L):
    parameters['W' + str(l)] = np.random.randn(layer_dims[l], layer_dims[l-1])*np.sqrt(2/layer_dims[l-1])  # Using he initialization
    parameters['b' + str(l)] = np.zeros((layer_dims[l], 1))