Generate most exciting inputs (MEIs).
This section describes the general usage of the MEI framework. Due to the fact that this framework uses DataJoint and NNFabrik general familiarity with these two packages is assumed.
This section describes how the tables used in the MEI generation process have to be set up.
This table contains ensembles of previously trained models. During the MEI generation process, all models in an ensemble will be given the same input and their output will be averaged. The framework provides a template class that you can use to create a trained ensemble model table by declaring a new class that inherits from the template. Afterwards you have to link up your class with your NNFabrik-style dataset and trained model tables via class attributes.
from mei.main import TrainedEnsembleModelTemplate
@schema
class TrainedEnsembleModel(TrainedEnsembleModelTemplate):
dataset_table = Dataset
trained_model_table = TrainedModel
This table has two jobs:
- Provide a method that can be used to get the to-be-optimized objective.
- Contain the information that is needed to come up with the aforementioned objective.
Note that you will have to implement your own objective table because the exact implementation is heavily dependent on the structure of the data you want to use and the architecture of your models.
The objective table implemented below contains information that can be used to map the ID of a real neuron (neuron_id
)
to the index of its corresponding output unit (output_unit
) in the model's output. The get_objective
method uses this information to constrain a given model to a single output unit and therefore to a single real neuron.
from mei.modules import ConstrainedOutputModel
@schema
class Objective(dj.Computed):
definition = """
-> self.dataset_table
neuron_id: int
---
output_unit: int
"""
dataset_table = Dataset
def make(self, key):
"""Fills the table."""
def get_objective(self, model, key):
output_unit = (self & key).fetch1("output_unit")
return ConstrainedOutputModel(model, output_unit)
Your implementation must provide a method called get_objective
that has a PyTorch module (model
) and a
dictionary (key
) as its only parameters and that must return a PyTorch module representing the objective. The returned
module must itself return
a scalar value when called.
This table contains generated MEIs. You can create your own MEI table by inheriting from the provided template class. Afterwards you have to link up your table with your trained (ensemble) model and objective tables via class attributes.
from mei.main import MEITemplate
@schema
class MEI(MEITemplate):
trained_model_table = TrainedEnsembleModel
objective_table = Objective
This section lays out the general steps one would execute when generating MEIs.
Note that this step is only required if you are using the trained ensemble model table.
You can create a new ensemble model by calling the create_ensemble
method of the trained ensemble model table with a
DataJoint restriction (key
). The passed restriction is used to restrict the trained model table and all models still
present in the restricted table will be made part of the new ensemble. Note that the provided restriction must be able
to restrict the dataset table down to a single entry because creating an ensemble consisting of models trained on
different datasets is currently not supported. For your own reference you can also pass a comment when creating a new
ensemble.
TrainedEnsembleModel().create_ensemble(key, comment="My ensemble")
Before generating MEIs you have to populate the objective table by either calling its populate
method if your
implementation provides it or by manually inserting entries.
Objective().populate()
Each MEI is generated according to a user-configurable method. You can specify a new method by adding it to the MEI
method table using its add_method
method (see example below). This method expects the name of a method function
(method_fn
) and method configuration object (method_config
).
The function name needs to be the absolute path to a callable object. A function that can be used to generate MEIs using
gradient ascent is provided with the framework and its path is mei.methods.gradient_ascent
. You can also implement
your own function and use it with the framework. Further information on how to do that can be found here.
The configuration object will be passed to the function by the framework and should contain information that will be used by the method function to alter its behavior.
In the case of the provided function the configuration object is a dictionary. It contains information about which
components to use when generating MEIs and how to configure those components. A component must be a callable object that
must return another callable object when called. The configuration dictionary contains the absolute path ("path"
) to
the corresponding component and can additionally contain a set of keyword arguments ("kwargs"
) that will be passed to
the corresponding component when it is initially called. Below is a list of supported components:
"device"
: Required, must be either"cpu"
or"cuda"
. The MEI will be generated on the CPU or the GPU depending on this value"initial"
: Required component used to generate an initial guess from which the MEI generation process will start"optimizer"
: Required component used to optimize the input to the model and in turn generate the MEI. Must be a PyTorch-style optimizer class"stopper"
: Required component used to determine whether or not to stop the MEI generation process in each iteration based on a user-defined condition"transform"
: Optional component used to transform the input before passing it through the model"regularization"
: Optional component used to compute a regularization term from the (transformed) input that is added to the model's output before taking the optimization step"precondition"
: Optional component used to precondition the gradient"postprocessing"
: Optional component that applies an operation to the input after each optimization step. The operation performed by this component does not influence the gradient"objectives"
: Optional component that consists of a list of sub-components. Each sub-component tracks an objective over the duration of the generation process
You can completely omit optional components from the configuration dictionary if you do not want to use them.
from mei.main import MEIMethod
method_fn = "mei.methods.gradient_ascent"
method_config = {
"initial": {"path": "mei.initial.RandomNormal"},
"optimizer": {"path": "torch.optim.SGD", "kwargs": {"lr": 0.1}},
"stopper": {"path": "mei.stoppers.NumIterations", "kwargs": {"num_iterations": 1000}},
"objectives": [
{"path": "mei.objectives.EvaluationObjective", "kwargs": {"interval": 10}}
],
"device": "cuda",
)
MEIMethod().add_method(method_fn, method_config, comment="My MEI method")
Next you have to specify a seed to make the MEI generation process random but reproducible by inserting a seed into the MEI seed table.
from mei.main import MEISeed
MEISeed().insert1({"mei_seed": 42})
After configuring everything you can generate MEIs by calling the populate
method of your MEI
table. The table will
insert one row for each generated MEI which itself can be found in the mei
attribute. Additionally each MEI is
associated with a score and an output object which can be found in the score
and output
attributes, respectively.
The score should express how well the generation process went but what it exactly represents is up to the used method function. In the case of the provided function it represents the final model evaluation.
The output object is an object that is returned by the method function at the end of the generation process. The included function will return a dictionary that contains the values of the objectives that were tracked during the generation process.
Note that the mei
and output
attributes are stored externally.
MEI().populate()
Instances of the State
class contain information describing a particular state encountered during the optimization
process. This information is used by various components in the framework. The attributes of a state instance are:
i_iter
: An integer representing the index of the optimization step this state corresponds toevaluation
: A float representing the evaluation of the model in response to the current inputreg_term
: A float representing the current regularization term added to the evaluation before the optimization step represented by this state was made. This value will be zero if no transformation is usedinput_
: A tensor representing the untransformed input to the model. This will be identical to the post-processed input from the last step for all steps except the first onetransformed_input
: A tensor representing the transformed input to the model. This will be identical to the untransformed input if no transformation is usedpost_processed_input
: A tensor representing the post-processed input. This will be identical to the untransformed input if no post-processing is donegrad
: A tensor representing the gradientpreconditioned_grad
: A tensor representing the preconditioned gradient. This will be identical to the gradient if no preconditioning is donestopper_output
: An object returned by the stopper component.
This section describes each component type in greater detail and how you can implement your own variants.
This component is the point of entry for the whole optimization process.
It will be called with your NNFabrik-style dataloaders dictionary, your model, your configuration object and an
integer representing the seed. It must return the MEI, a float representing the score the MEI achieved and an output
object. The MEI and the output object must be compatible with PyTorch's save
function.
def method(dataloaders, model, config, seed):
"""Generates a MEI."""
return mei, score, output
After you have implemented your method you can use it by adding it to the MEI method table as described here.
This component is used to check whether or not to stop the MEI generation process based on the current state of the generation process.
All custom stoppers must implement the __call__
method and they should inherit from a abstract base class (ABC) called
OptimizationStopper
. The stopper will be called after each optimization step with the current state of the
optimization process. It must return (True, output)
if the optimization process is to be stopped and
(False, output)
otherwise. output
can be any object or None
but it can not be omitted.
Below is the implementation of a custom stopper that stops the MEI generation process once the model's evaluation reaches a user-specified threshold:
"""Contents of mymodule.py"""
from mei.stoppers import OptimizationStopper
class EvaluationThreshold(OptimizationStopper):
def __init__(self, threshold):
self.threshold = threshold
def __call__(self, current_state):
if current_state.evaluation >= self.threshold:
return True, None
else:
return False, None
You can then use your custom stopper component by specifying it in the method_config
like so:
method_config = {
"stopper": {"path": "mymodule.EvaluationThreshold", "kwargs": {"threshold": 2.5}},
...
}