How to understand the CRFModelWrapper weights after the training?

Hi everyone,

I trained a neural network with the CRFModelWrapper for a sequence labelling task and am trying to figure out what the trained weights of CRF layers look like. I applied the command “get_weight()” to the crf_layer in the CRFModelWrapper to extract the weights of the embedded crf_layer in CRFModelWrapper and found it contains 5 categories of weights, which I list below with their dimensions(shapes).

  1. Chain kernels: (11,11),
  2. The left boundary (11,1),
  3. The right boundary (11,1),
  4. crf_model_wrapper/crf/dense/kernel:0, (11,11),
  5. crf_model_wrapper/crf/dense/bias:0,(11,1),

The number of tags I encoded is 11, and I thought it should contain only one (11,11) array representing the probability of the transition between each tag. Is the Chain kernels the same as what I mean? Thanks.

1 Like