e8caea8ea26a18f93c49ceb6e8d9a48403ca9e30,lxmls/deep_learning/numpy_models/mlp.py,NumpyMLP,log_forward,#NumpyMLP#,43

Before Change

        z = np.dot(tilde_z, weight.T) + bias

        // Softmax is computed in log-domain to prevent underflow/overflow
        log_tilde_z = z - logsumexp(z, axis=1)[:, None]

        return log_tilde_z, layer_inputs

    def cross_entropy_loss(self, input, output):

After Change

        z = np.dot(tilde_z, weight.T) + bias

        // Softmax is computed in log-domain to prevent underflow/overflow
        log_tilde_z = z - logsumexp(z, axis=1, keepdims=True)

        return log_tilde_z, layer_inputs

    def cross_entropy_loss(self, input, output):

Instances