Optimizing a Model’s Hyper-parameters
The easiest way to optimize a model is with the pykeen.hpo.hpo_pipeline()
function.
All of the following examples are about getting the best model
when training pykeen.models.TransE
on the pykeen.datasets.Nations
dataset.
Each gives a bit of insight into usage of the hpo_pipeline()
function.
The minimal usage of the hyper-parameter optimization is to specify the
dataset, the model, and how much to run. The following example shows how to
optimize the TransE model on the Nations dataset a given number of times using
the n_trials
argument.
>>> from pykeen.hpo import hpo_pipeline
>>> hpo_pipeline_result = hpo_pipeline(
... n_trials=30,
... dataset='Nations',
... model='TransE',
... )
Alternatively, the timeout
can be set. In the following example,
as many trials as possible will be run in 60 seconds.
>>> from pykeen.hpo import hpo_pipeline
>>> hpo_pipeline_result = hpo_pipeline(
... timeout=60,
... dataset='Nations',
... model='TransE',
... )
The hyper-parameter optimization pipeline has the ability to optimize hyper-parameters for the corresponding
*_kwargs
arguments in the pykeen.pipeline.pipeline()
:
model
loss
regularizer
optimizer
negative_sampler
training
Defaults
Each component’s hyper-parameters have a reasonable default values. For example, every model in PyKEEN has default for its hyper-parameters chosen from the best-reported values in each model’s original paper unless otherwise stated on the model’s reference page. In case hyper-parameters for a model for a specific dataset were not available, we choose the hyper-parameters based on the findings in our large-scale benchmarking [ali2020a]. For most components (e.g., models, losses, regularizers, negative samples, training loops), these values are stored in the default valeues of the respective classes’ __init__() functions. They can be viewed in the corresponding reference section of the docs.
Some components contain strategies for doing hyper-parameter optimization. When you call the
pykeen.hpo.hpo_pipeline()
, the following steps are taken to determine what happens for each hyper-parameter
in each componenent:
If an explicit value was passed, use it.
If no explicit value was passed and an HPO strategy was passed, use the explicit strategy.
If no explicit value was passed and no HPO strategy was passed and there is a default HPO strategy, use the default strategy.
If no explicit value was passed, no HPO strategy was passed, and there is no default HPO strategy, use the default hyper-parameter value
If no explicit value was passed, no HPO strategy was passed, and there is no default HPO strategy, and there is no default hyper-parameter value, raise an
TypeError
For example, the TransE model’s default HPO strategy for its embedding_dim
argument is to search between
\([16, 256]\) with a step size of 16. The \(l_p\) norm is set to search as either 1 or 2. This will be overridden
with 50 in the following code:
>>> from pykeen.hpo import hpo_pipeline
>>> hpo_pipeline_result = hpo_pipeline(
... dataset='Nations',
... model='TransE',
... model_kwargs=dict(embedding_dim=50),
... )
The strategy can be explicitly overridden with:
>>> from pykeen.hpo import hpo_pipeline
>>> hpo_pipeline_result = hpo_pipeline(
... dataset='Nations',
... model='TransE',
... model_kwargs_ranges=dict(
... embedding_dim=dict(type=int, low=16, high=256, step=32),
... ),
... )
Each model, loss, regularizer, negative sampler, and training loop specify a class variable called hpo_defaults
in which there’s a dictionary with all of the default strategies. They keys match up to the arguments in their
respective __init__()
functions.
Since optimizers aren’t re-implemented in PyKEEN, there’s a specfic dictionary at
pykeen.optimizers.optimizers_hpo_defaults
containing their strategies. It’s debatable whether
you should optimize the optimizers (yo dawg), so you can always choose to set the learning rate lr
to a constant
value.
Strategies
An HPO strategy is a Python dict
with a type
key corresponding to a categorical variable, boolean
variable, integer variable, or floating point number variable. The value itself for type
should be one of the following:
"categorical"
bool
or"bool"
int
or"int"
float
or"float"
Several strategies can be grouped together in a dictionary where the key is the name of the hyper-parameter
for the component in the *_kwargs_ranges
arguments to the HPO pipeline.
Categorical
The only other key to use inside a categorical variable is choices
. For example, if you want to
choose between Kullback-Leibler divergence or expected likelihood as similarity used in the KG2E model,
you can write a strategy like:
>>> from pykeen.hpo import hpo_pipeline
>>> hpo_pipeline_result = hpo_pipeline(
... dataset='Nations',
... model='KG2E',
... model_kwargs_ranges=dict(
... dist_similarity=dict(type='categorical', choices=['KL', 'EL']),
... ),
... )
Boolean
The boolean variable actually doesn’t need any extra keys besides the type, so a strategy for a boolean
variable always looks like dict(type='bool')
. Under the hood, this is automatically translated to a categorical
variable with choices=[True, False]
.
>>> from pykeen.hpo import hpo_pipeline
>>> hpo_pipeline_result = hpo_pipeline(
... dataset='Nations',
... model='TransE',
... training_loop='sLCWA',
... negative_sampler_kwargs_ranges=dict(
... filtered=dict(type=boolean),
... ),
... )
Integers and Floating Point Numbers
The integer and floating point number strategies share several aspects. Both require a low
and high
entry
like in dict(type=float, low=0.0, high=1.0)
or dict(type=int, low=1, high=10)
.
Linear Scale
By default, you don’t need to specify a scale
, but you can be explicit by setting scale='linear'
.
This behavior should be self explanatory - there is no rescaling and you get back uniform distribution
within the bounds specified by the low
and high
arguments. This applies to both type=int
and
type=float
. The following example uniformly choose from [1,100]:
>>> from pykeen.hpo import hpo_pipeline
>>> hpo_pipeline_result = hpo_pipeline(
... dataset='Nations',
... model='TransE',
... training_loop='sLCWA',
... negative_sampler_kwargs_ranges=dict(
... num_negs_per_pos=dict(type=int, low=1, high=100),
... ),
... )
Power Scale (type=int
only)
The power scale was originally implemented as scale='power_two'
to support
pykeen.models.ConvE
’s output_channels
parameter. However, using two as a base is a bit limiting, so we
also implemented a more general scale='power'
where you can set set base
. Here’s an example to optimize over
the number of negatives per positive ratio using base=10:
>>> from pykeen.hpo import hpo_pipeline
>>> hpo_pipeline_result = hpo_pipeline(
... dataset='Nations',
... model='TransE',
... training_loop='sLCWA',
... negative_sampler_kwargs_ranges=dict(
... num_negs_per_pos=dict(type=int, scale='power', base=10, low=0, high=2),
... ),
... )
The power scale can only be used with type=int and not bool, categorical, or float. I like this scale because it can quickly discretize a large search space. In this example, you will get [10**0, 10**1, 10**2] as choices then uniformly choose from them.
Logarithmic Reweighting
The evil twin to the power scale is logarithmic reweighting on the linear scale. This is applicable type=int
and
type=float
. Rather than changing the choices themselves, the log scale uses Optuna’s built in log
functionality
to reassign the probabilities uniformly over the log’d distribution. The same example as above could be
accomplished with:
>>> from pykeen.hpo import hpo_pipeline
>>> hpo_pipeline_result = hpo_pipeline(
... dataset='Nations',
... model='TransE',
... training_loop='sLCWA',
... negative_sampler_kwargs_ranges=dict(
... num_negs_per_pos=dict(type=int, low=1, high=100, log=True),
... ),
... )
but this time, it’s not discretized. However, you’re just as likely to pick from \([1,10]\) as \([10, 100]\).
Stepping
With the linear scale, you can specify the step
size. This discretizes the distribution in linear space,
so if you want to pick from \(10, 20, ... 100\), you can do:
>>> from pykeen.hpo import hpo_pipeline
>>> hpo_pipeline_result = hpo_pipeline(
... dataset='Nations',
... model='TransE',
... training_loop='sLCWA',
... negative_sampler_kwargs_ranges=dict(
... num_negs_per_pos=dict(type=int, low=10, high=100, step=10),
... ),
... )
This actually also works with logarithmic reweighting, since it is still technically on a linear scale, but with probabilites reweighted logarithmically. So now you’d pick from one of \([10]\) or \([20, 30, 40, ..., 100]\) with the same probability
>>> from pykeen.hpo import hpo_pipeline
>>> hpo_pipeline_result = hpo_pipeline(
... dataset='Nations',
... model='TransE',
... training_loop='sLCWA',
... negative_sampler_kwargs_ranges=dict(
... num_negs_per_pos=dict(type=int, low=10, high=100, step=10, log=True),
... ),
... )
Custom Strategies
While the default values for hyper-parameters are encoded with the python syntax
for default values of the __init__()
function of each model, the ranges/scales can be
found in the class variable pykeen.models.Model.hpo_default
. For
example, the range for TransE’s embedding dimension is set to optimize
between 50 and 350 at increments of 25 in pykeen.models.TransE.hpo_default
.
TransE also has a scoring function norm that will be optimized by a categorical
selection of {1, 2} by default.
Note
These hyper-parameter ranges were chosen as reasonable defaults for the benchmark datasets FB15k-237 / WN18RR. When using different datasets, the ranges might be suboptimal.
All hyper-parameters defined in the hpo_default
of your chosen model will be
optimized by default. If you already have a value that you’re happy with for
one of them, you can specify it with the model_kwargs
attribute. In the
following example, the embedding_dim
for a TransE model is fixed at 200,
while the rest of the parameters will be optimized using the pre-defined HPO strategies in
the model. For TransE, that means that the scoring function norm will be optimized
as 1 or 2.
>>> from pykeen.hpo import hpo_pipeline
>>> hpo_pipeline_result = hpo_pipeline(
... model='TransE',
... model_kwargs=dict(
... embedding_dim=200,
... ),
... dataset='Nations',
... n_trials=30,
... )
If you would like to set your own HPO strategy for the model’s hyperparameters, you can do so with the
model_kwargs_ranges
argument. In the example below, the embeddings are
searched over a larger range (low
and high
), but with a higher step
size (q
), such that 100, 200, 300, 400, and 500 are searched.
>>> from pykeen.hpo import hpo_pipeline
>>> hpo_result = hpo_pipeline(
... n_trials=30,
... dataset='Nations',
... model='TransE',
... model_kwargs_ranges=dict(
... embedding_dim=dict(type=int, low=100, high=500, q=100),
... ),
... )
Warning
If the given range is not divisible by the step size, then the upper bound will be omitted.
If you want to optimize the entity initializer, you can use the type='categorical'
type,
which requres a choices=[...]
key with a list of choices. This works for strings, integers,
floats, etc.
>>> from pykeen.hpo import hpo_pipeline
>>> hpo_result = hpo_pipeline(
... n_trials=30,
... dataset='Nations',
... model='TransE',
... model_kwargs_ranges=dict(
... entity_initializer=dict(type='categorical', choices=[
... 'xavier_uniform',
... 'xavier_uniform_norm',
... 'uniform',
... ]),
... ),
... )
The same could be used for constrainers, normalizers, and regularizers over both entities and relations. However, different models might have different names for the initializer, normalizer, constrainer and regularizer since there could be multiple representations for either the entity, relation, or both. Check your desired model’s documentation page for the kwargs that you can optimize over.
Keys of pykeen.nn.emb.initializers
can be passed as initializers as strings and
keys of pykeen.nn.emb.constrainers
can be passed as constrainers as strings.
The HPO pipeline does not support optimizing over the hyper-parameters for each initializer. If you are interested in this, consider rolling your own ablation study pipeline.
Optimizing the Loss
While each model has its own default loss, you can explicitly specify a loss
the same way as in pykeen.pipeline.pipeline()
.
>>> from pykeen.hpo import hpo_pipeline
>>> hpo_pipeline_result = hpo_pipeline(
... n_trials=30,
... dataset='Nations',
... model='TransE',
... loss='MarginRankingLoss',
... )
As stated in the documentation for pykeen.pipeline.pipeline()
, each model
specifies its own default loss function in pykeen.models.Model.loss_default
.
For example, the TransE model defines the margin ranking loss as its default in
pykeen.models.TransE.loss_default
.
Each model also specifies default hyper-parameters for the loss function in
pykeen.models.Model.loss_default_kwargs
. For example, DistMultLiteral
explicitly sets the margin to 0.0 in pykeen.models.DistMultLiteral.loss_default_kwargs
.
Unlike the model’s hyper-parameters, the models don’t store the strategies for
optimizing the loss functions’ hyper-parameters. The pre-configured strategies
are stored in the loss function’s class variable pykeen.models.Loss.hpo_default
.
However, similarily to how you would specify model_kwargs_ranges
, you can
specify the loss_kwargs_ranges
explicitly, as in the following example.
>>> from pykeen.hpo import hpo_pipeline
>>> hpo_pipeline_result = hpo_pipeline(
... n_trials=30,
... dataset='Nations',
... model='TransE',
... loss='MarginRankingLoss',
... loss_kwargs_ranges=dict(
... margin=dict(type=float, low=1.0, high=2.0),
... ),
... )
Optimizing the Negative Sampler
When the stochastic local closed world assumption (sLCWA) training approach is used for training, a negative sampler
(subclass of pykeen.sampling.NegativeSampler
) is chosen.
Each has a strategy stored in pykeen.sampling.NegativeSampler.hpo_default
.
Like models and regularizers, the rules are the same for specifying negative_sampler
,
negative_sampler_kwargs
, and negative_sampler_kwargs_ranges
.
Optimizing the Optimizer
Yo dawg, I heard you liked optimization, so we put an optimizer around your
optimizer so you can optimize while you optimize. Since all optimizers used
in PyKEEN come from the PyTorch implementations, they obviously do not have
hpo_defaults
class variables. Instead, every optimizer has a default
optimization strategy stored in pykeen.optimizers.optimizers_hpo_defaults
the same way that the default strategies for losses are stored externally.
Optimizing the Optimized Optimizer - a.k.a. Learning Rate Schedulers
If optimizing your optimizer doesn’t cut it for you, you can turn it up a notch and use learning rate schedulers (lr_scheduler) that will vary the learning rate of the optimizer. This can e.g. be useful to have a more aggressive learning rate in the beginning to quickly make progress while lowering the learning rate over time to allow the model to smoothly converge to the optimum.
PyKEEN allows you to use the learning rate schedulers provided by PyTorch, which you can
simply specify as you would in the pykeen.pipeline.pipeline()
.
>>> from pykeen.hpo import hpo_pipeline
>>> hpo_pipeline_result = hpo_pipeline(
... dataset='Nations',
... model='TransE',
... lr_scheduler='ExponentialLR',
... )
>>> pipeline_result.save_to_directory('nations_transe')
The same way as the optimizers don’t come with hpo_defaults
class variables, lr_schedulers rely
on their own optimization strategies provided in pykeen.lr_schedulers.lr_schedulers_hpo_defaults
In case you are ready to explore even more you can of course also set your own ranges with the
lr_scheduler_kwargs_ranges
keyword argument as in:
>>> from pykeen.hpo import hpo_pipeline
>>> hpo_pipeline_result = hpo_pipeline(
... dataset='Nations',
... model='TransE',
... lr_scheduler='ExponentialLR',
... lr_scheduler_kwargs_ranges=dict(
... gamma=dict(type=float, low=0.8, high=1.0),
... ),
... )
>>> pipeline_result.save_to_directory('nations_transe')
Optimizing Everything Else
Without loss of generality, the following arguments to pykeen.pipeline.pipeline()
have corresponding *_kwargs and *_kwargs_ranges:
training_loop
(only kwargs, not kwargs_ranges)evaluator
evaluation
Early Stopping
Early stopping can be baked directly into the optuna
optimization.
The important keys are stopper='early'
and stopper_kwargs
.
When using early stopping, the hpo_pipeline()
automatically takes
care of adding appropriate callbacks to interface with optuna
.
>>> from pykeen.hpo import hpo_pipeline
>>> hpo_pipeline_result = hpo_pipeline(
... n_trials=30,
... dataset='Nations',
... model='TransE',
... stopper='early',
... stopper_kwargs=dict(frequency=5, patience=2, relative_delta=0.002),
... )
These stopper kwargs were chosen to make the example run faster. You will likely want to use different ones.
Configuring Optuna
Choosing a Search Algorithm
Because PyKEEN’s hyper-parameter optimization pipeline is powered by Optuna, it can directly use all of
Optuna’s built-in samplers listed on optuna.samplers
or any custom subclass of
optuna.samplers.BaseSampler
.
By default, PyKEEN uses the Tree-structured Parzen Estimator (TPE; optuna.samplers.TPESampler
),
a probabilistic search algorithm. You can explicitly set the sampler using the sampler
argument
(not to be confused with the negative sampler used when training under the sLCWA):
>>> from pykeen.hpo import hpo_pipeline
>>> from optuna.samplers import TPESampler
>>> hpo_pipeline_result = hpo_pipeline(
... n_trials=30,
... sampler=TPESampler,
... dataset='Nations',
... model='TransE',
... )
You can alternatively pass a string so you don’t have to worry about importing Optuna. PyKEEN knows that sampler classes always end in “Sampler” so you can pass either “TPE” or “TPESampler” as a string. This is case-insensitive.
>>> from pykeen.hpo import hpo_pipeline
>>> hpo_pipeline_result = hpo_pipeline(
... n_trials=30,
... sampler="tpe",
... dataset='Nations',
... model='TransE',
... )
It’s also possible to pass a sampler instance directly:
>>> from pykeen.hpo import hpo_pipeline
>>> from optuna.samplers import TPESampler
>>> sampler = TPESampler(prior_weight=1.1)
>>> hpo_pipeline_result = hpo_pipeline(
... n_trials=30,
... sampler=sampler,
... dataset='Nations',
... model='TransE',
... )
If you’re working in a JSON-based configuration setting, you won’t be able to instantiate the sampler
with your desired settings like this. As a solution, you can pass the keyword arguments via the
sampler_kwargs
argument in combination with specifying the sampler as a string/class to the HPO pipeline like in:
>>> from pykeen.hpo import hpo_pipeline
>>> hpo_pipeline_result = hpo_pipeline(
... n_trials=30,
... sampler="tpe",
... sampler_kwargs=dict(prior_weight=1.1),
... dataset='Nations',
... model='TransE',
... )
To emulate most hyper-parameter optimizations that have used random
sampling, use optuna.samplers.RandomSampler
like in:
>>> from pykeen.hpo import hpo_pipeline
>>> from optuna.samplers import RandomSampler
>>> hpo_pipeline_result = hpo_pipeline(
... n_trials=30,
... sampler=RandomSampler,
... dataset='Nations',
... model='TransE',
... )
Grid search can be performed using optuna.samplers.GridSampler
like in:
>>> from pykeen.hpo import hpo_pipeline
>>> from optuna.samplers import GridSampler
>>> hpo_pipeline_result = hpo_pipeline(
... n_trials=30,
... sampler=GridSampler,
... dataset='Nations',
... model='TransE',
... )
Full Examples
The examples above have shown the permutation of one setting at a time. This section has some more complete examples.
The following example sets the optimizer, loss, training, negative sampling, evaluation, and early stopping settings.
>>> from pykeen.hpo import hpo_pipeline
>>> hpo_pipeline_result = hpo_pipeline(
... n_trials=30,
... dataset='Nations',
... model='TransE',
... model_kwargs=dict(embedding_dim=20, scoring_fct_norm=1),
... optimizer='SGD',
... optimizer_kwargs=dict(lr=0.01),
... loss='marginranking',
... loss_kwargs=dict(margin=1),
... training_loop='slcwa',
... training_kwargs=dict(num_epochs=100, batch_size=128),
... negative_sampler='basic',
... negative_sampler_kwargs=dict(num_negs_per_pos=1),
... evaluator_kwargs=dict(filtered=True),
... evaluation_kwargs=dict(batch_size=128),
... stopper='early',
... stopper_kwargs=dict(frequency=5, patience=2, relative_delta=0.002),
... )
If you have the configuration as a dictionary:
>>> from pykeen.hpo import hpo_pipeline_from_config
>>> config = {
... 'optuna': dict(
... n_trials=30,
... ),
... 'pipeline': dict(
... dataset='Nations',
... model='TransE',
... model_kwargs=dict(embedding_dim=20, scoring_fct_norm=1),
... optimizer='SGD',
... optimizer_kwargs=dict(lr=0.01),
... loss='marginranking',
... loss_kwargs=dict(margin=1),
... training_loop='slcwa',
... training_kwargs=dict(num_epochs=100, batch_size=128),
... negative_sampler='basic',
... negative_sampler_kwargs=dict(num_negs_per_pos=1),
... evaluator_kwargs=dict(filtered=True),
... evaluation_kwargs=dict(batch_size=128),
... stopper='early',
... stopper_kwargs=dict(frequency=5, patience=2, relative_delta=0.002),
... )
... }
... hpo_pipeline_result = hpo_pipeline_from_config(config)
If you have a configuration (in the same format) in a JSON file:
>>> import json
>>> config = {
... 'optuna': dict(
... n_trials=30,
... ),
... 'pipeline': dict(
... dataset='Nations',
... model='TransE',
... model_kwargs=dict(embedding_dim=20, scoring_fct_norm=1),
... optimizer='SGD',
... optimizer_kwargs=dict(lr=0.01),
... loss='marginranking',
... loss_kwargs=dict(margin=1),
... training_loop='slcwa',
... training_kwargs=dict(num_epochs=100, batch_size=128),
... negative_sampler='basic',
... negative_sampler_kwargs=dict(num_negs_per_pos=1),
... evaluator_kwargs=dict(filtered=True),
... evaluation_kwargs=dict(batch_size=128),
... stopper='early',
... stopper_kwargs=dict(frequency=5, patience=2, relative_delta=0.002),
... )
... }
... with open('config.json', 'w') as file:
... json.dump(config, file, indent=2)
... hpo_pipeline_result = hpo_pipeline_from_path('config.json')