Onnx Feature Scorer

rasr/feature_scorer.py

@@ -7,6 +7,7 @@
     "ReturnnScorer",
     "InvAlignmentPassThroughFeatureScorer",
     "PrecomputedHybridFeatureScorer",
+    "OnnxFeatureScorer",
 ]
 
 from sisyphus import *
@@ -15,6 +16,8 @@
 
 import os
 
+from typing import Union, Dict
+
 from .config import *
 from i6_core.util import get_returnn_root
 
@@ -126,25 +129,54 @@ def __init__(self, prior_mixtures, scale=1.0, priori_scale=0.0, prior_file=None)
         self.config.normalize_mixture_weights = False
 
 
-class TFLabelContextFeatureScorer(FeatureScorer):
+class OnnxFeatureScorer(FeatureScorer):
     def __init__(
         self,
-        fs_tf_config,
-        contextPriorFile,
-        diphonePriorFile,
-        prior_mixtures,
-        prior_scale,
+        *,
+        mixtures: tk.Path,
+        model: tk.Path,
+        io_map: Dict[str, str],
+        label_log_posterior_scale: float = 1.0,
+        label_prior_scale: float = 0.7,
+        label_log_prior_file: tk.Path = None,
+        apply_log_on_output: bool = False,
+        negate_output: bool = True,
+        intra_op_threads: int = 1,
+        inter_op_threads: int = 1,
+        **kwargs,
     ):
-        super().__init__()
+        """
+        :param mixtures: path to a *.mix file e.g. output of either EstimateMixturesJob or CreateDummyMixturesJob
+        :param model: path of a model e.g. output of ExportPyTorchModelToOnnxJob
+        :param io_map: mapping between internal rasr identifiers and the model related input/output. Default key values
+                            are "features" and "output", and optionally "features-size", e.g.
+                            io_map = {"features": "data", "output": "classes"}
+        :param label_log_posterior_scale: scales for the log probability of a label e.g. 1.0 is recommended
+        :param label_prior_scale: scale for the prior log probability of a label reasonable e.g. values in [0.1, 0.7] interval
+        :param label_log_prior_file: xml file containing log prior probabilities e.g. estimated from the model via povey method
+        :param apply_log_on_output: whether to apply the log-function on the output, usefull if the model outputs softmax instead of log-softmax
+        :param negate_output: whether negate output (because the model outputs log softmax and not negative log softmax
+        :param intra_op_threads: Onnxruntime session's number of parallel threads within each operator
+        :param inter_op_threads: Onnxruntime session's number of parallel threads between operators used only for parallel execution mode
+        """
+        super().__init__(*args, **kwargs)
 
-        self.config = RasrConfig()
-        self.config.feature_scorer_type = "tf-label-context-scorer"
-        self.config.file = prior_mixtures
-        self.config.num_label_contexts = 46
-        self.config.prior_scale = prior_scale
-        self.config.context_prior = contextPriorFile
-        self.config.diphone_prior = diphonePriorFile
-        self.config.normalize_mixture_weights = False
-        self.config.loader = fs_tf_config.loader
-        self.config.input_map = fs_tf_config.input_map
-        self.config.output_map = fs_tf_config.output_map
+        self.config.feature_scorer_type = "onnx-feature-scorer"
+        self.config.file = mixtures
+        self.config.scale = label_log_posterior_scale
+        self.config.priori_scale = label_prior_scale
+        if label_log_prior_file is not None:
+            self.config.prior_file = label_log_prior_file
+
+        self.config.session.file = model
+
+        if apply_log_on_output:
+            self.config.apply_log_on_output = apply_log_on_output
+        if not negate_output:
+            self.config.negate_output = negate_output
+
+        self.post_config.session.intra_op_num_threads = intra_op_threads
+        self.post_config.session.inter_op_num_threads = inter_op_threads
+
+        for k, v in io_map.items():
+            self.config.io_map[k] = v