Saving and Loading
Alibi Detect includes support for saving and loading detectors to disk. To
save a detector, simply call the save_detector
method and provide a path to a directory (a new
one will be created if it doesn’t exist):
from alibi_detect.od import OutlierVAE
from alibi_detect.saving import save_detector
od = OutlierVAE(...)
filepath = './my_detector/'
save_detector(od, filepath)
To load a previously saved detector, use the load_detector
method and provide it with the path to the detector’s
directory:
from alibi_detect.saving import load_detector
filepath = './my_detector/'
od = load_detector(filepath)
Warning
When loading a saved detector, a warning will be issued if the runtime alibi-detect version is different from the version used to save the detector. It is highly recommended to use the same alibi-detect, Python and dependency versions as were used to save the detector to avoid potential bugs and incompatibilities.
Formats
Detectors can be saved using two formats:
Config format: For drift detectors, by default
save_detector
serializes the detector via a config file namedconfig.toml
, stored infilepath
. The TOML format is human-readable, which makes the config files useful for record keeping, and allows a detector to be edited before it is reloaded. For more details, see Detector Configuration Files.Legacy format: Outlier and adversarial detectors are saved to dill files stored within
filepath
. Drift detectors can also be saved in this legacy format by runningsave_detector
withlegacy=True
. Loading is performed in the same way, by simply runningload_detector(filepath)
.
Supported detectors
The following tables list the current state of save/load support for each detector. Adding full support for the remaining detectors is in the Roadmap.
Detector |
Legacy save/load |
Config save/load |
---|---|---|
✅ |
✅ |
|
❌ |
✅ |
|
❌ |
✅ |
|
❌ |
✅ |
|
✅ |
✅ |
|
❌ |
✅ |
|
✅ |
✅ |
|
✅ |
✅ |
|
✅ |
✅ |
|
❌ |
✅ |
|
❌ |
✅ |
|
❌ |
✅ |
|
❌ |
✅ |
|
❌ |
✅ |
|
❌ |
✅ |
|
❌ |
✅ |
Detector |
Legacy save/load |
Config save/load |
---|---|---|
✅ |
❌ |
|
✅ |
❌ |
|
✅ |
❌ |
|
✅ |
❌ |
|
✅ |
❌ |
|
✅ |
❌ |
|
✅ |
❌ |
|
✅ |
❌ |
|
✅ |
❌ |
|
✅ |
❌ |
Detector |
Legacy save/load |
Config save/load |
---|---|---|
✅ |
❌ |
|
✅ |
❌ |
Supported ML models
Alibi Detect drift detectors offer the option to perform preprocessing with user-defined machine learning models:
model = ... # A TensorFlow model
preprocess_fn = partial(preprocess_drift, model=model, batch_size=128)
cd = MMDDrift(x_ref, backend='tensorflow', p_val=.05, preprocess_fn=preprocess_fn)
Additionally, some detectors are built upon models directly,
for example the Classifier drift detector requires a model
to be passed
as an argument:
cd = ClassifierDrift(x_ref, model, backend='sklearn', p_val=.05, preds_type='probs')
In order for a detector to be saveable and loadable, any models contained within it (or referenced within a detector configuration file) must fall within the family of supported models:
Alibi Detect supports serialization of any TensorFlow model that can be serialized to the HDF5 format. Custom objects should be pre-registered with register_keras_serializable.
PyTorch models are serialized by saving the entire model
using the dill module. Therefore, Alibi Detect should support any PyTorch
model that can be saved and loaded with torch.save(..., pickle_module=dill)
and torch.load(..., pickle_module=dill)
.
Scikit-learn models are serialized using joblib.
Any scikit-learn model that is a subclass of sklearn.base.BaseEstimator
is supported, including
xgboost models following
the scikit-learn API.
Online detectors
Online drift detectors are stateful, with their state updated each timestep t
(each time
.predict()
is called). save_detector()
will save the state of online
detectors to disk if t > 0
. At load time, load_detector()
will load this state.
For example:
from alibi_detect.cd import LSDDDriftOnline
from alibi_detect.saving import save_detector, load_detector
# Init detector (t=0)
dd = LSDDDriftOnline(x_ref, window_size=10, ert=50)
# Run 2 predictions
pred_1 = dd.predict(x_1) # t=1
pred_2 = dd.predict(x_2) # t=2
# Save detector (state will be saved since t>0)
save_detector(dd, filepath)
# Load detector
dd_new = load_detector(filepath) # detector will start at t=2
To save a clean (stateless) detector, it should be reset before saving:
dd.reset_state() # reset to t=0
save_detector(dd, filepath) # save the detector without state