Migrating from 4.x to 5.0
CoBaIR 5.0 splits the library into a config model and the inference, brings the web GUI and renames the package
to lowercase cobair. Code written for 4.x keeps working in 5.x, it only shows DeprecationWarnings.
The deprecated names are removed in 6.0.0
(#199).
Install
5.0 needs Python ≥ 3.10. The dependencies are split into extras:
| 4.x | 5.0 |
|---|---|
pip install CoBaIR |
pip install "cobair[infer]" for the inference, "cobair[all]" for everything |
python start_configurator.py |
pip install "cobair[qt]", then cobair-qt |
new: pip install "cobair[web]", then cobair-web (Python ≥ 3.11) |
pip install cobair alone installs only the config model (cobair.model, needs only pyyaml).
Imports and API
| 4.x (deprecated) | 5.0 |
|---|---|
import CoBaIR |
import cobair |
from CoBaIR.bayes_net import BayesNet |
from cobair.model import Config and from cobair.infer import Inference |
load_config(path) |
Config.load(path) |
BayesNet(config) |
config = Config.from_dict(config), inference = Inference(config) |
bayes_net.add_context(...), edit_context, del_context, add_intention, edit_intention, del_intention, change_context_apriori_value, change_influence_value, add_combined_influence, del_combined_influence, change_decision_threshold |
the same methods on Config |
bayes_net.validate_config() (warnings) |
config.validate() returns the problems as list, config.valid |
bayes_net.save(path) |
config.save(path) |
bayes_net.infer(evidence) |
Inference(config).infer(evidence) |
bayes_net.bind_discretization_function(...) |
inference.bind_discretization_function(...) |
from CoBaIR.configurator import Configurator |
from cobair.qt import Configurator |
Example:
# 4.x
from CoBaIR.bayes_net import BayesNet, load_config
net = BayesNet(load_config('my_scenario.yml'))
intention, threshold, probabilities = net.infer({'speech commands': 'pickup'})
# 5.0
from cobair.model import Config
from cobair.infer import Inference
inference = Inference(Config.load('my_scenario.yml'))
intention, threshold, probabilities = inference.infer({'speech commands': 'pickup'})
Inference.infer returns a named tuple (intention, decision_threshold, probabilities), so the 4.x unpacking
works unchanged. An Inference works on a copy of the config: create a new one after editing the config.
Config files
The YAML format is compatible in both directions: 5.0 reads 4.x files, and 4.x reads files written by 5.0.
Files written by 5.0 start with a comment and the keys format_version and created_with, which 4.x ignores.
Contexts, instantiations and intentions keep their order instead of being sorted alphabetically.
Changed behaviour
These were bugs in 4.x. If your code relied on them, check the results:
- Editing or deleting a context no longer deletes all combined influences. Renaming a context renames it in the combined influences.
- All combinations of a context tuple count in the inference, 4.x only used the first one.
del_combined_influenceremoves only the given combination. - Discretization functions run for every evidence value that is not an instantiation. In 4.x they only ran for unhashable values, so e.g. a float for a bool context was ignored with a warning.
- Renaming a context to the name of another context raises a
ValueErrorinstead of overwriting it. - Apriori probabilities are checked to sum up to 1 with a small tolerance, so
0.1 + 0.2 + 0.7is valid. Integers (e.g. a decision threshold of0) count as numbers.