adastep.py 文件源码

def adastep(
        inputs, loss, params, outputs=(),
        max_iter=8, rho = 0.9, momentum=None,
        initial_learning_rate = 1.0e-3, max_learning_rate=1.0, max_delta = 1.0e-1, eps=1.0e-6):
  cache_inputs, cache_grads, get_loss, set_params = grad_base(
    inputs, loss, params, outputs, norm_gradients=False, momentum=momentum
  )

  one = T.constant(1.0, dtype='float32')

  v = theano.shared(
    np.float32(initial_learning_rate), name = 'v'
  )
  new_v = T.fscalar()

  upd_v = OrderedDict()
  upd_v[v] = v * rho + new_v * (one - rho)

  update_v = theano.function([new_v], v, updates=upd_v, no_default_updates=True)
  get_v = theano.function([], v, no_default_updates=True)

  return _adastep(
    cache_inputs, cache_grads, get_loss, set_params,
    get_v, update_v,
    max_iter=max_iter,
    max_learning_rate=max_learning_rate,
    max_delta=max_delta,
    eps=eps
  )