TensorFlow：BasicRNNCell（）从哪里获得权重和偏差？

Question

I don't really understand how TensorFlow's tf.contrib.rnn.BasicRNNCell works.
We defined a dict of weights, which should be used in different places in the graph, but I don't get how I can tell the command above where it needs which weight, or in general how I can tell it to use especially said dict as weights.
Can anyone elaborate this?

Answer 1

When you call tf.contrib.rnn.BasicRNNCell , say,

cell = tf.contrib.rnn.BasicRNNCell(num_units=100)

its weights (in tensorflow it's called kernel ) and bias will not be created. At this point cell only knows the number of neurons it's gonna create and how to activate afterward. (Logically, you need to tell cell the input size to create its weights variable.)

Then you probably will do:

rnn = tf.nn.dynamic_rnn(cell, rnn_inputs, dtype=tf.float32)

During tf.nn.dynamic_rnn , 2 important things will happen sequentially:

1. cell.build will be called. Since input is specified, cell will create its weights (kernel) and bias. After that you can access to them using cell._kernel and cell._bias . The corresponding nodes will be added to tensorflow's computation graph, but not linked to other nodes yet.

2. cell.call will be called. The weight and bias nodes created in 1 are linked to some newly nodes to produce the RNN result.

So literally you don't need to tell the command how to do things exactly. What you need to do is to use TensorFlow APIs in an appropriate way and most of the implementation details will be taken very good care of.