l1_proximal_step_snippet.py

    # Apply proximal gradient for the variables with l1 lasso loss
    # Non-negative weights constraint
    if L1_LOSS_WEIGHT > 0:
      for var in tf.get_collection(utils.LASSO_KEY):
        th_t = tf.fill(tf.shape(var), tf.convert_to_tensor(L1_LOSS_WEIGHT) * lr)
        zero_t = tf.zeros(tf.shape(var))
        var_temp = var - th_t * tf.sign(var)
        assign_op = var.assign(tf.select(tf.less(tf.abs(var), th_t), zero_t, var_temp))
        l1_op_list.append(assign_op)
        print('\tL1 loss added: %s(strength: %f)' % (var.name, L1_LOSS_WEIGHT))