fix: support mlg-norm and eds-pseudo again

Author: percevalw

edsnlp/__init__.py

@@ -44,6 +44,10 @@ def find_spec(self, fullname, path, target=None):  # pragma: no cover
             new_name = "edsnlp.metrics" + fullname[14:]
             spec = importlib.util.spec_from_loader(fullname, AliasLoader(new_name))
             return spec
+        if fullname.startswith("edsnlp.metrics.span_classification"):
+            new_name = "edsnlp.metrics.span_attributes" + fullname[34:]
+            spec = importlib.util.spec_from_loader(fullname, AliasLoader(new_name))
+            return spec
         if "span_qualifier" in fullname.split("."):
             new_name = fullname.replace("span_qualifier", "span_classifier")
             spec = importlib.util.spec_from_loader(fullname, AliasLoader(new_name))

edsnlp/data/converters.py

@@ -43,6 +43,9 @@
 
 _DEFAULT_TOKENIZER = None
 
+# For backward compatibility
+SequenceStr = AsList[str]
+
 
 def without_filename(d):
     d.pop(FILENAME, None)

edsnlp/metrics/span_attributes.py

@@ -11,11 +11,12 @@
 def span_attribute_metric(
     examples: Examples,
     span_getter: SpanGetterArg,
-    attributes: Attributes,
+    attributes: Attributes = None,
     include_falsy: bool = False,
     default_values: Dict = {},
     micro_key: str = "micro",
     filter_expr: Optional[str] = None,
+    **kwargs: Any,
 ):
     """
     Scores the attributes predictions between a list of gold and predicted spans.
@@ -47,6 +48,23 @@ def span_attribute_metric(
     -------
     Dict[str, float]
     """
+    if "qualifiers" in kwargs:
+        warnings.warn(
+            "The `qualifiers` argument of span_attribute_metric() is "
+            "deprecated. Use `attributes` instead.",
+            DeprecationWarning,
+        )
+        assert attributes is None
+        attributes = kwargs.pop("qualifiers")
+    if attributes is None:
+        raise TypeError(
+            "span_attribute_metric() missing 1 required argument: 'attributes'"
+        )
+    if kwargs:
+        raise TypeError(
+            f"span_attribute_metric() got unexpected keyword arguments: "
+            f"{', '.join(kwargs.keys())}"
+        )
     examples = make_examples(examples)
     if filter_expr is not None:
         filter_fn = eval(f"lambda doc: {filter_expr}")

edsnlp/pipes/trainable/embeddings/span_pooler/span_pooler.py

@@ -129,7 +129,9 @@ def preprocess(
             "same."
         )
         aligned_contexts = (
-            [[c] for c in contexts] if pre_aligned else align_spans(contexts, spans)
+            [[c] for c in contexts]
+            if pre_aligned
+            else align_spans(contexts, spans, sort_by_overlap=True)
         )
         for i, (span, ctx) in enumerate(zip(spans, aligned_contexts)):
             if len(ctx) == 0 or ctx[0].start > span.start or ctx[0].end < span.end:
@@ -196,8 +198,9 @@ def forward(self, batch: SpanPoolerBatchInput) -> SpanPoolerBatchOutput:
         """
         device = next(self.parameters()).device
         if len(batch["begins"]) == 0:
+            span_embeds = torch.empty(0, self.output_size, device=device)
             return {
-                "embeddings": torch.empty(0, self.output_size, device=device),
+                "embeddings": batch["begins"].with_data(span_embeds),
             }
 
         embeds = self.embedding(batch["embedding"])["embeddings"]

edsnlp/pipes/trainable/span_classifier/span_classifier.py

@@ -194,7 +194,7 @@ def __init__(
         embedding: SpanEmbeddingComponent,
         attributes: AttributesArg = None,
         qualifiers: AttributesArg = None,
-        span_getter: SpanGetterArg = {"ents": True},
+        span_getter: SpanGetterArg = None,
         context_getter: Optional[SpanGetterArg] = None,
         values: Optional[Dict[str, List[Any]]] = None,
         keep_none: bool = False,
@@ -217,6 +217,7 @@ def __init__(
             sub_span_getter is not None and span_getter is None
         ):  # pragma: no cover  # noqa: E501
             span_getter = sub_span_getter
+        span_getter = span_getter or {"ents": True}
         sub_context_getter = getattr(embedding, "context_getter", None)
         if (
             sub_context_getter is not None and context_getter is None

edsnlp/pipes/trainable/span_linker/span_linker.py

@@ -278,7 +278,7 @@ def __init__(
         rescale: float = 20,
         threshold: float = 0.5,
         attribute: str = "cui",
-        span_getter: SpanGetterArg = {"ents": True},
+        span_getter: SpanGetterArg = None,
         context_getter: Optional[SpanGetterArg] = None,
         reference_mode: Literal["concept", "synonym"] = "concept",
         probability_mode: Literal["softmax", "sigmoid"] = "sigmoid",
@@ -289,6 +289,7 @@ def __init__(
         sub_span_getter = getattr(embedding, "span_getter", None)
         if sub_span_getter is not None and span_getter is None:  # pragma: no cover
             span_getter = sub_span_getter
+        span_getter = span_getter or {"ents": True}
         sub_context_getter = getattr(embedding, "context_getter", None)
         if (
             sub_context_getter is not None and context_getter is None
@@ -309,7 +310,7 @@ def __init__(
         self.reference_mode = reference_mode
         self.probability_mode = probability_mode
         self.init_weights = init_weights
-        self.context_getter: SpanGetter = context_getter or span_getter
+        self.context_getter: SpanGetter = context_getter or self.span_getter
         with warnings.catch_warnings():
             warnings.simplefilter("ignore", UserWarning)
             self.classifier = Metric(

edsnlp/train.py

@@ -1,7 +1,7 @@
 from confit import Cli
 
-from edsnlp.core.registries import registry
-from edsnlp.training import train
+from edsnlp.training.trainer import *  # noqa: F403
+from edsnlp.training.trainer import registry, train
 
 app = Cli(pretty_exceptions_show_locals=False)
 train_command = app.command(name="train", registry=registry)(train)