3e2274441c92ebafcdbb6d0cf3cc934b6d87602d,lxmls/deep_learning/rnn.py,NumpyRNN,grads,#NumpyRNN#Any#Any#,125

Before Change

            
            nabla_W_h += dh_raw[:,None].dot(h[:, t][None,:])
            
            nabla_W_x += dh_raw[:,None].dot(z[:, t][None,:])
            
            nabla_W_e[:, x[t]] += W_x.T.dot(dh_raw)

After Change

            // Weight gradients
            nabla_W_y += np.outer(e[:, t], h[:, t+1])
            nabla_W_h += np.outer(e_raw, h[:, t])
            nabla_W_x += np.outer(e_raw, z[:, t])
            nabla_W_e[:, x[t]] += W_x.T.dot(e_raw)
           
        // Normalize to be in agreement with the loss
        nabla_params = [nabla_W_e/nr_steps, nabla_W_x/nr_steps, 

Instances