Open xiaoshingshing opened 5 years ago
Hi, sorry for the very late response, I don't receive notifications on the repo anymore but I want to make sure this has an answer even if it is late.
Yes there is code to turn a spec to an architecuture:
If you only care about evaluating the spec and getting the same metrics out (e.g. with a spec which wasn't evaluated in the dataset), there is a convenience function written in the API: https://github.com/google-research/nasbench/blob/master/nasbench/api.py#L293 (NASBench.evaluate(...)
)
If you just want the TensorFlow graph associated with a spec, there's a library function you can call here: https://github.com/google-research/nasbench/blob/master/nasbench/lib/model_builder.py#L36 (model_builder.build_model_fn(...)
). Note that this function will create a function which returns an TPUEstimatorSpec, which might not be exactly what you want. You will need to modify the code a little bit to just include the part that builds the model itself (line 46 - line 99).
Given a model_spec, are there any api to turning it into a model architecture?