Add PickOptimalParametersJob (#491)

* Add PickOptimalParametersJob

Job that takes pickleable parameters and
a tk.Variable value, and just returns the paramaters
corresponding to the highest or lowest value.

Co-authored-by: Albert Zeyer <[email protected]>

Author: JackTemaki

tools/parameter_tuning.py

@@ -0,0 +1,60 @@
+from sisyphus import Job, Task, tk
+from typing import Any, Literal, Sequence, Union
+
+import numpy as np
+
+from i6_core.util import instanciate_delayed
+
+
+class GetOptimalParametersAsVariableJob(Job):
+    """
+    Pick a set of optimal parameters based on their assigned (dynamic) score value.
+    Each optimal parameter is outputted individually to be accessible in the Sisyphus manager.
+
+    Can be used to e.g. pick best lm-scale and prior scale to a corresponding ScliteJob.out_wer.
+    """
+
+    def __init__(
+        self,
+        *,
+        parameters: Sequence[Sequence[Any]],
+        values: Sequence[tk.Variable],
+        mode: Union[Literal["maximize"], Literal["minimize"]],
+    ):
+        """
+        :param parameters: parameters[best_idx] will be written to self.out_optimal_parameters
+             as Sisyphus output variables.
+             parameters[best_idx] (and thus self.out_optimal_parameters) is a sequence of fixed length,
+             to allow to index into it.
+             Thus, len(parameters[i]) must be the same for all i.
+        :param values: best_idx = argmax(values) or argmin(values).
+              Must have len(values) == len(parameters).
+              Some calculations might be done using DelayedOps math beforehand.
+        :param mode: "minimize" or "maximize"
+        """
+        assert len(parameters) == len(values)
+        for param in parameters[1:]:
+            assert len(param) == len(parameters[0]), "all entries should have the same number of parameters"
+        assert mode in ["minimize", "maximize"]
+        self.parameters = parameters
+        self.values = values
+        self.mode = mode
+        self.num_parameters = len(parameters[0])
+
+        self.out_optimal_parameters = [self.output_var("param_%i" % i, pickle=True) for i in range(self.num_parameters)]
+
+    def tasks(self):
+        yield Task("run", mini_task=True)
+
+    def run(self):
+        values = instanciate_delayed(self.values)
+
+        if self.mode == "minimize":
+            index = np.argmin(values)
+        else:
+            index = np.argmax(values)
+
+        best_parameters = self.parameters[index]
+
+        for i, param in enumerate(best_parameters):
+            self.out_optimal_parameters[i].set(param)