theano——shared,function(outputs,updates,givens)

• theano 下的 functions &＃xff08;from theano import function&＃xff09;的两个重要性质&＃xff1a;
  
  • 多个 outputs&＃xff1b;
  • 多个 updates&＃xff1b;

1. shared

• shared 与 theano.tensor.matrix()/vector() 的区别在于&＃xff0c;tensor.matrix/vector 符号变量不需要初始化&＃xff0c;而 shared 变量则需要初始化赋值&＃xff1b;

>>> import theano
>>> import theano.tensor as T >>> state &＃61; theano.shared(0)
>>> type(state)
theano.tensor.sharedvar.ScalarSharedVariable
>>> state.get_value(borrow&＃61;True)
array(0)# shared 型变量&＃xff0c;通过 get_value 取出其值&＃xff1b;>>> x &＃61; T.iscalar(&＃39;x&＃39;)
>>> type(x)
theano.tensor.var.TensorVariable# tensor value&＃xff08;thenao.tensor模块下的变量&＃xff1f;&＃xff09;是没有get_value成员函数的>>> type(x&＃43;state)
theano.tensor.var.TensorVariable

2. updates

>>> counter &＃61; function([x], state, updates&＃61;[(state, state&＃43;x)])
>>> counter(5)
>>> state.get_value(borrow&＃61;True)
5
>>> counter(6)
>>> state.get_value(borrow&＃61;True)
11

不使用可选参数updates&＃xff1a;

>>> state &＃43;&＃61; x# 此时state的类型已自动转换为tensor value&＃xff0c;不再具有get_value的成员
>>> counter &＃61; theano.function([x], state)
>>> counter(5)
array(5)
>>> state.get_value()
AttributeError: &＃39;TensorVariable&＃39; object has no attribute &＃39;get_value&＃39;

theano.function 的 updates, 二元 tuple 构成的 list&＃xff0c;

[(a, a&＃43;1), (b, b&＃43;1)]

其含义其实为&＃xff0c;a &＃61; a&＃43;1, b &＃61; b&＃43;1&＃xff1b;

所以常见与&＃xff0c;梯度下降算法中的权值的更新&＃xff0c;如下&＃xff1a;

updates &＃61; [(W, W-eta*delta_W),(b, b-eta*delta_b)]

3. givens

test_model &＃61; theano.function(inputs&＃61;[idx],outputs&＃61;layer3.error(y),givens&＃61;{x: test_set_x[idx*batch_size:(idx&＃43;1)*batch_size],y: test_set_y[idx*batch_size:(idx&＃43;1)*batch_size]}
)
valid_model &＃61; theano.function(inputs&＃61;[idx],outputs&＃61;layer3.error(y),givens&＃61;{x: valid_set_x[idx*batch_size:(idx&＃43;1)*batch_size],y: valid_set_y[idx*batch_size:(idx&＃43;1)*batch_size]}
)