model_fn closure for TPUEstimatorReturns model_fn closure, which is an input to TPUEstimator.
This function is similar to model_fn_builder from run_classifier.R.
(RBERT issue #28.)
.model_fn_builder_EF(bert_config, init_checkpoint, layer_indexes, use_tpu)BertConfig instance.
Character; path to the checkpoint directory, plus checkpoint name stub (e.g. "bert_model.ckpt"). Path must be absolute and explicit, starting with "/".
Integer list; indexes (positive, or negative counting back from the end) indicating which layers to extract as "output features". (It needs to be specified here because we get them back as the model "predictions".)
Logical; whether to use TPU.
model_fn closure for TPUEstimator.
The model_fn function takes four parameters:
A list (or similar structure) that contains objects such as
input_ids, input_mask, tokens, and
input_type_ids. These objects will be inputs to the
create_model function.
Not used in this function, but presumably we need to keep this slot here.
Character; value such as "train", "infer", or "eval".
Not used in this function, but presumably we need to keep this slot here.
The output of model_fn is the result of a
tf$contrib$tpu$TPUEstimatorSpec call.