Use Interior Point OPTimizer (IPOPT) with TensorFlow

tf-ipopt does constrained optimization in TensorFlow using IPOPT. Models are defined using TensorFlow variables and operations. Gradients for optimization are computed by TensorFlow and passed on to the IPOPT solver. This allows us to exploit the power of GPUs to speed up numerical optimization. See examples for details on how to model and solve constrained optimization.

Follow the steps to install. It is assumed that

• tensorflow
• ipopt
• pyipopt

are already installed. Details on how to install ipopt and pyipopt will be added later.

Step 1

We define a new class IpoptOptimizerInterface that can be used very much like ScipyOptimizerInterface. Make the following two changes in the external_optimizer.py in tensorflow package located at tensorflow/contrib/opt/python/training/.

Allow import of the new IpoptOptimizerInterface class.

__all__ = ['ExternalOptimizerInterface', 'ScipyOptimizerInterface', 'IpoptOptimizerInterface']

Step 2

Second, copy the class definition from the file ipopt_optimizer.py to external_optimizer.py.

Step 3

Modify _allowed_symbols to add IpoptOptimizerInterface in __init__.py located at tensorflow/contrib/opt/

_allowed_symbols = [
    'PowerSignOptimizer',
    'AddSignOptimizer'
    'DelayCompensatedGradientDescentOptimizer',
    'DropStaleGradientOptimizer',
    'ExternalOptimizerInterface',
    'LazyAdamOptimizer',
    'NadamOptimizer',
    'MovingAverageOptimizer',
    'ScipyOptimizerInterface',
    'IpoptOptimizerInterface',
    'VariableClippingOptimizer',
    'MultitaskOptimizerWrapper',
    'clip_gradients_by_global_norm',
    'ElasticAverageOptimizer',
    'ElasticAverageCustomGetter'
]

That's it! Check examples for details on how to use IpoptOptimizerInterface.