prior and decoding for joint to be tested

users/raissi/setups/common/TF_factored_hybrid_system.py

@@ -381,15 +381,17 @@ def _compute_returnn_rasr_priors_via_hdf(
 
         return prior_job
 
-    def set_mono_priors_returnn_rasr(
+    def set_single_prior_returnn_rasr(
         self,
         key: str,
         epoch: int,
         train_corpus_key: str,
         dev_corpus_key: str,
+        context_type: PhoneticContext = PhoneticContext.monophone,
         returnn_config: Optional[returnn.ReturnnConfig] = None,
         output_layer_name: str = "output",
-        data_share: float = 0.1,
+        smoothen: bool = False,
+        data_share: float = 0.3,
     ):
         self.set_graph_for_experiment(key)
 
@@ -411,10 +413,26 @@ def set_mono_priors_returnn_rasr(
             share=data_share,
         )
 
+
         job.add_alias(f"priors/{name}/c")
-        tk.register_output(f"priors/{name}/center-state.xml", job.out_prior_xml_file)
 
-        self.experiments[key]["priors"] = [job.out_prior_xml_file]
+        if context_type == PhoneticContext.monophone:
+            p_info = PriorInfo(
+                center_state_prior=PriorConfig(file=job.out_prior_xml_file, scale=0.0),
+            )
+            tk.register_output(f"priors/{name}/center-state.xml", p_info.center_state_prior.file)
+        elif context_type == PhoneticContext.joint_diphone:
+            p_info = PriorInfo(
+                diphone_prior=PriorConfig(file=job.out_prior_xml_file, scale=0.0),
+            )
+            tk.register_output(f"priors/{name}/joint_diphone.xml", p_info.diphone_prior.file)
+        else:
+            raise NotImplementedError("Unknown PhoneticContext, i.e. context_type")
+
+        p_info = smoothen_priors(p_info) if smoothen else p_info
+        self.experiments[key]["priors"] = p_info
+
+
 
     def set_diphone_priors_returnn_rasr(
         self,

users/raissi/setups/common/data/factored_label.py

@@ -42,39 +42,6 @@ def use_word_end(self) -> bool:
         return self == PhonemeStateClasses.word_end
 
 
-@dataclass(eq=True, frozen=True)
-class LabelInfo:
-    n_contexts: int
-    n_states_per_phone: int
-    phoneme_state_classes: PhonemeStateClasses
-    ph_emb_size: int
-    st_emb_size: int
-    state_tying: RasrStateTying
-    add_unknown_phoneme: bool = True
-    sil_id: typing.Optional[int] = None
-
-    def get_n_of_dense_classes(self) -> int:
-        n_contexts = self.n_contexts
-        if not self.add_unknown_phoneme:
-            n_contexts += 1
-        return self.n_states_per_phone * (n_contexts**3) * self.phoneme_state_classes.factor()
-
-    def get_n_state_classes(self) -> int:
-        return self.n_states_per_phone * self.n_contexts * self.phoneme_state_classes.factor()
-
-    @classmethod
-    def default_ls(cls) -> "LabelInfo":
-        return LabelInfo(
-            n_contexts=42,
-            n_states_per_phone=3,
-            ph_emb_size=32,
-            st_emb_size=128,
-            add_unknown_phoneme=True,
-            phoneme_state_classes=PhonemeStateClasses.word_end,
-            state_tying=RasrStateTying.triphone,
-        )
-
-
 class PhoneticContext(Enum):
     """
     These are the implemented models. The string value is the one used in the feature scorer of rasr, except monophone
@@ -126,3 +93,46 @@ def is_triphone(self):
             or self == PhoneticContext.triphone_symmetric
             or self == PhoneticContext.tri_state_transition
         )
+
+@dataclass(eq=True, frozen=True)
+class LabelInfo:
+    n_contexts: int
+    n_states_per_phone: int
+    phoneme_state_classes: PhonemeStateClasses
+    ph_emb_size: int
+    st_emb_size: int
+    state_tying: RasrStateTying
+
+    add_unknown_phoneme: bool = True
+    sil_id: typing.Optional[int] = None
+
+    def get_n_of_dense_classes(self) -> int:
+        if self.state_tying == RasrStateTying.monophone:
+            exp = 1
+        elif self.state_tying == RasrStateTying.diphone:
+            exp = 2
+        elif self.state_tying == RasrStateTying.triphone:
+            exp = 3
+        else:
+            assert False, "cannot compute number of CART classes"
+
+        n_contexts = self.n_contexts
+        if not self.add_unknown_phoneme:
+            n_contexts += 1
+        return self.n_states_per_phone * (n_contexts**exp) * self.phoneme_state_classes.factor()
+
+    def get_n_state_classes(self) -> int:
+        return self.n_states_per_phone * self.n_contexts * self.phoneme_state_classes.factor()
+
+    @classmethod
+    def default_ls(cls) -> "LabelInfo":
+        return LabelInfo(
+            n_contexts=42,
+            n_states_per_phone=3,
+            ph_emb_size=32,
+            st_emb_size=128,
+            sil_id=40,
+            add_unknown_phoneme=True,
+            phoneme_state_classes=PhonemeStateClasses.word_end,
+            state_tying=RasrStateTying.triphone,
+        )

users/raissi/setups/common/decoder/BASE_factored_hybrid_search.py

@@ -57,9 +57,9 @@ def __str__(self):
         return self.value
 
     def get_fs_class(self):
-        if self.factored:
+        if self == self.factored:
             return FactoredHybridFeatureScorer
-        elif self.nn_precomputed:
+        elif self == self.nn_precomputed:
             return rasr.PrecomputedHybridFeatureScorer
         else:
             raise ValueError("Unknown type of Feature Scorer")
@@ -232,10 +232,11 @@ def get_nn_precomputed_feature_scorer(
     if isinstance(prior_info, PriorInfo):
         check_prior_info(context_type=context_type, prior_info=prior_info)
 
+
     return feature_scorer_type.get_fs_class()(
         prior_mixtures=mixtures,
-        prior_file=prior_info.joint_diphone.file,
-        priori_scale=prior_info.joint_diphone.scale,
+        priori_scale=prior_info.diphone_prior.scale,
+        prior_file=prior_info.diphone_prior.file,
     )
 
 
@@ -356,21 +357,21 @@ def set_tf_fs_flow(self):
             raise NotImplementedError
 
     def set_nnprecomputed_tf_fs_flow(self):
-        tf_flow = self.get_nnprecomputed_tf_flow()
+        self.feature_scorer_flow = self.get_nnprecomputed_tf_flow()
+        """
         tf_feature_flow = add_tf_flow_to_base_flow(
             base_flow=self.feature_path,
             tf_flow=tf_flow,
             tf_fwd_input_name=tf_fwd_input_name,
         )
-        self.feature_scorer_flow = tf_feature_flow
+        self.feature_scorer_flow = tf_feature_flow"""
 
     def get_nnprecomputed_tf_flow(self):
         return make_precomputed_hybrid_tf_feature_flow(
             tf_graph=self.graph,
             tf_checkpoint=self.model_path,
             output_layer_name=self.tensor_map.out_joint_diphone,
-            native_ops=self.tf_library,
-            tf_fwd_input_name=tf_fwd_input_name,
+            native_ops=self.library_path,
         )
 
     def set_factored_tf_fs_flow(self):

users/raissi/setups/common/decoder/config.py

@@ -33,14 +33,15 @@ class PriorInfo:
     center state.
     """
 
-    center_state_prior: PriorConfig
+    center_state_prior: Optional[PriorConfig] = None
     left_context_prior: Optional[PriorConfig] = None
     right_context_prior: Optional[PriorConfig] = None
     diphone_prior: Optional[PriorConfig] = None
 
+
     def with_scale(
         self,
-        center: Float,
+        center: Optional[Float] = None,
         left: Optional[Float] = None,
         right: Optional[Float] = None,
         diphone: Optional[Float] = None,
@@ -50,17 +51,18 @@ def with_scale(
 
         left/right/diphone scale must be set if the left/right priors are set.
         """
-
+        assert self.center_state_prior is None or center is not None
         assert self.left_context_prior is None or left is not None
         assert self.right_context_prior is None or right is not None
         assert self.diphone_prior is None or diphone is not None
 
+        center = self.center_state_prior.with_scale(center) if self.center_state_prior is not None else None
         left = self.left_context_prior.with_scale(left) if self.left_context_prior is not None else None
         right = self.right_context_prior.with_scale(right) if self.right_context_prior is not None else None
         diphone = self.diphone_prior.with_scale(diphone) if self.diphone_prior is not None else None
         return dataclasses.replace(
             self,
-            center_state_prior=self.center_state_prior.with_scale(center),
+            center_state_prior=center,
             left_context_prior=left,
             right_context_prior=right,
             diphone_prior=diphone,
@@ -104,6 +106,16 @@ def from_triphone_job(cls, output_dir: Union[str, tk.Path]) -> "PriorInfo":
             left_context_prior=PriorConfig(file=output_dir.join_right("left-context.xml"), scale=0.0),
             right_context_prior=PriorConfig(file=output_dir.join_right("right-context.xml"), scale=0.0),
         )
+    def from_diiphone_job(cls, output_dir: Union[str, tk.Path]) -> "PriorInfo":
+        """
+        Initializes a PriorInfo instance with scale 0.0 from the output directory of
+        a previously run/captured ComputeTriphoneForwardPriorsJob.
+        """
+
+        output_dir = tk.Path(output_dir) if isinstance(output_dir, str) else output_dir
+        return cls(
+            diphone_prior=PriorConfig(file=output_dir.join_right("center-state.xml"), scale=0.0),
+        )
 
 
 PosteriorScales = TypedDict(
@@ -160,11 +172,12 @@ def with_lm_scale(self, scale: Float) -> "SearchParameters":
 
     def with_prior_scale(
         self,
-        center: Float,
+        center: Optional[Float] = None,
         left: Optional[Float] = None,
         right: Optional[Float] = None,
+        diphone: Optional[Float] = None,
     ) -> "SearchParameters":
-        return dataclasses.replace(self, prior_info=self.prior_info.with_scale(center=center, left=left, right=right))
+        return dataclasses.replace(self, prior_info=self.prior_info.with_scale(center=center, left=left, right=right, diphone=diphone))
 
     def with_pron_scale(self, pron_scale: Float) -> "SearchParameters":
         return dataclasses.replace(self, pron_scale=pron_scale)
@@ -232,6 +245,23 @@ def default_triphone(cls, *, priors: PriorInfo) -> "SearchParameters":
             we_pruning_limit=10000,
         )
 
+    @classmethod
+    def default_joint_diphone(cls, *, priors: PriorInfo) -> "SearchParameters":
+        return cls(
+            beam=20,
+            beam_limit=500_000,
+            lm_scale=9.0,
+            tdp_scale=0.4,
+            pron_scale=2.0,
+            prior_info=priors.with_scale(diphone=0.4),
+            tdp_speech=(3.0, 0.0, "infinity", 0.0),
+            tdp_silence=(0.0, 3.0, "infinity", 20.0),
+            tdp_non_word=(0.0, 3.0, "infinity", 20.0),
+            non_word_phonemes="[UNKNOWN]",
+            we_pruning=0.5,
+            we_pruning_limit=10000,
+        )
+
     @classmethod
     def default_cart(cls, *, priors: PriorInfo) -> "SearchParameters":
         return dataclasses.replace(
@@ -249,5 +279,8 @@ def default_for_ctx(cls, context: PhoneticContext, priors: PriorInfo) -> "Search
             return cls.default_diphone(priors=priors)
         elif context == PhoneticContext.triphone_forward:
             return cls.default_triphone(priors=priors)
+        elif context == PhoneticContext.joint_diphone:
+            return cls.default_joint_diphone(priors=priors)
+
         else:
             raise NotImplementedError(f"unimplemented context {context}")