CNTK 层上的 Python 操作

Question

我不是 python 人，想了解如何将一些 Python 代码转换为 C# 代码：

# model parameters
with default_options(bias=False): # all the projections have no bias
    attn_proj_enc   = Stabilizer(enable_self_stabilization=enable_self_stabilization) >> Dense(attention_dim, init=init, input_rank=1) # projects input hidden state, keeping span axes intact
    attn_proj_dec   = Stabilizer(enable_self_stabilization=enable_self_stabilization) >> Dense(attention_dim, init=init, input_rank=1) # projects decoder hidden state, but keeping span and beam-search axes intact
    attn_proj_tanh  = Stabilizer(enable_self_stabilization=enable_self_stabilization) >> Dense(1            , init=init, input_rank=1) # projects tanh output, keeping span and beam-search axes intact
attn_final_stab = Stabilizer(enable_self_stabilization=enable_self_stabilization)

此代码是来自此处的 sn-p：CNTK/bindings/python/cntk/layers/models/attention.py

我的问题是，>> 运算符实际上在做什么？

Stabilizer 和Dense 一样是一个层，那么这些层发生了什么？这里有花哨的Bitwise Operation吗？

Answer 1

我以前从未见过这种情况，但模型实例化似乎是答案：

Example:
 >>> model = Dense(500) >> Activation(C.relu) >> Dense(10)
 >>> # is the same as
 >>> model = Dense(500) >> C.relu >> Dense(10)
 >>> # and also the same as
 >>> model = Dense(500, activation=C.relu) >> Dense(10)

发件人：python/cntk/layers/layers.py#L1361

的：

Another example is a GRU layer with projection, which could be realized as 
``Recurrence(GRU(500) >> Dense(200))``,
where the projection is applied to the hidden state as fed back to the next 
step.
``F>>G`` is a short-hand for ``Sequential([F, G])``.

发件人：python/cntk/layers/sequence.py#L331

特别是模型中的序列。