I don't know if this will help, but probably you are overfitting the dataset as you have increased the complexity of your model by adding neurons to the layers.

Yes, this seems like your network is overfitting. It would be advantageous to share your training and validation loss curves to see whether this assumption is correct.

Thanks for responding,

I tried the same dataset on Matlab over bigger networks and it gave good results. So I don't think the problem here is overfitting. I tried changing the learning rate and it improved the results. But I had to lower the learning rate and increase the iterations significantly. However, I am not sure if that is the right/only issue with the code or not. Eduardo Perez-Valero Simon Graham

The loss was constant

If you reduce the learning rate, your cost optimizer will take longer to converge to a semi-optimal result, so that is perfectly normal, as the steps it is taken are shorter in every parameter update.

As Simon Graham suggested, you should probably plot the learning curves of your model so you can check if you are having a high variance problem and then proceed accordingly to overcome the problem.