How to load a python module from a string?

Question

Given the string

optimizer = "tensorflow.train.ProximalAdagradOptimizer"

How can I achieve the following:

import tensorflow
optimizer = tensorflow.train.ProximalAdagradOptimizer()

Context

To add context to my specific use case (to address one of the comments): I'm trying to create a text-based config that describes how my model (specifically: estimator) was configured. That way I can easily re-instantiate them after training if I want to train it more or do other stuff with them. I haven't found a simple way to do this; I'm not looking for a saved_model for this. My use case is to easily reload models prior to committing them to a saved_model state. The config would look something like this:

model_config = {
  "type": "DNNClassifier",
  "n_classes": 10,
  "feature_columns": [
    {
      "numeric_column": [
        {
          "key": "x"
        },
        {
          "key": "y"
        }
      ]
    }
  ],
  "optimizer": {
    "AdamOptimizer": {
      "learning_rate": 1.0
    }
  }
}

Given that "config" I can instantiate my estimator with:

estimator = load_estimator(model_config, model_dir=model_dir)

The value of type would resolve to tensorflow.estimator.DNNClassifier . The value of feature_column[0].<key> would resolve to tensorflow.feature_column.numeric_column . Finally, the value of optimizer.<key> would resolve to tensorflow.train.AdamOptimizer .

Answer 1

you could using the function eval its not the best answer but does the job

if "tensorflow" in optimizer:
    import tensorflow
    optimizer = eval(optimizer + '()')

you can"t do that : eval("import tensorflow")

Answer 2

You can do something like this:

import importlib

def get_object_by_name(qualname):
    module, _, object = qualname.rpartition(".")
    if module:
        # package parameter is only necessary for relative imports
        # (here relative to this package)
        vs = vars(importlib.import_module(module, package=__package__))
    else:
        # If no module name we assume it is from the current module
        vs = globals()
    return vs[object]

optimizer_qualname = "tensorflow.train.AdamOptimizer"
optimizer_class = get_object_by_name(optimizer_qualname)
optimizer = optimizer_class()

I changed the optimizer to avoid the error due to the missing learning rate parameter in the example.