Update params

birdnet_analyzer/evaluation/__init__.py

@@ -33,7 +33,8 @@ class EvaluationResult(NamedTuple):
     Besides the metrics and the tensors the assessment ran on, it carries the context
     a caller needs to explain the numbers: which recordings had predictions but no
     annotations, and which classes were empty (and therefore excluded from any
-    aggregate score).
+    aggregate score). ``sample_data`` is the per-window matrix the metrics were derived
+    from -- the same table the GUI offers as the "data table" download.
     """
 
     metrics_df: "pd.DataFrame"
@@ -43,6 +44,7 @@ class EvaluationResult(NamedTuple):
     classes: tuple[str, ...]
     unmatched_recordings: tuple[str, ...]
     empty_classes: tuple[str, ...]
+    sample_data: "pd.DataFrame"
 
 
 def process_data(
@@ -56,7 +58,7 @@ def process_data(
     columns_predictions: dict[str, str] | None = None,
     selected_classes: Sequence[str] | None = None,
     selected_recordings: list[str] | None = None,
-    metrics_list: tuple[str, ...] = ("accuracy", "precision", "recall"),
+    metrics_list: tuple[str, ...] = ("auroc", "precision", "recall", "f1", "ap"),
     threshold: float = 0.1,
     class_wise: bool = False,
     averaging: Literal["macro", "micro", "weighted"] = "macro",
@@ -176,9 +178,33 @@ def process_data(
         classes=classes,
         unmatched_recordings=tuple(sorted(processor.unmatched_prediction_files)),
         empty_classes=pa.empty_classes(labels),
+        sample_data=processor.get_sample_data(),
     )
 
 
+# The metrics the PerformanceAssessor understands; also the CLI's allowed choices.
+VALID_METRICS = ("accuracy", "recall", "precision", "f1", "ap", "auroc")
+
+
+def _threshold_arg(value: str) -> float:
+    """argparse type for ``--threshold``: a float strictly between 0 and 1.
+
+    Fails at the parser with a clear message instead of deep inside the
+    PerformanceAssessor, which requires ``0 < threshold < 1``.
+    """
+    try:
+        number = float(value)
+    except ValueError:
+        raise argparse.ArgumentTypeError(
+            f"threshold must be a number between 0 and 1 (exclusive), got {value!r}"
+        ) from None
+    if not 0 < number < 1:
+        raise argparse.ArgumentTypeError(
+            f"threshold must be between 0 and 1 (exclusive), got {value}"
+        )
+    return number
+
+
 def main():
     """
     Entry point for the script. Parses command-line arguments and orchestrates the
@@ -232,11 +258,16 @@ def main():
     parser.add_argument(
         "--metrics",
         nargs="+",
-        default=["accuracy", "precision", "recall"],
-        help="List of metrics",
+        choices=VALID_METRICS,
+        default=["auroc", "precision", "recall", "f1", "ap"],
+        help="List of metrics (accuracy is excluded by default; in this sample-based "
+        "multilabel setting it is dominated by true negatives and misleadingly high)",
     )
     parser.add_argument(
-        "--threshold", type=float, default=0.1, help="Threshold value (0-1)"
+        "--threshold",
+        type=_threshold_arg,
+        default=0.1,
+        help="Threshold value, strictly between 0 and 1",
     )
     parser.add_argument(
         "--class_wise", action="store_true", help="Calculate class-wise metrics"
@@ -301,9 +332,18 @@ def main():
             ", ".join(result.empty_classes),
         )
 
-    # Create output directory if needed
-    if args.output_dir and not os.path.exists(args.output_dir):
-        os.makedirs(args.output_dir)
+    # Write the result tables (and later the plots) when an output directory is given.
+    # These mirror the GUI's "results table" and "data table" downloads.
+    if args.output_dir:
+        os.makedirs(args.output_dir, exist_ok=True)
+
+        results_table_path = os.path.join(args.output_dir, "results_table.csv")
+        result.metrics_df.to_csv(results_table_path, index=True)
+        logger.info("Saved results table to %s", results_table_path)
+
+        data_table_path = os.path.join(args.output_dir, "data_table.csv")
+        result.sample_data.to_csv(data_table_path, index=False)
+        logger.info("Saved data table to %s", data_table_path)
 
     # Generate plots if specified
     if args.plot_metrics:

birdnet_analyzer/gui/multi_file.py

@@ -12,6 +12,7 @@ def _output_type_map():
         loc.localize("multi-tab-output-type-raven-label"): "table",
         loc.localize("multi-tab-output-type-audacity-label"): "audacity",
         loc.localize("multi-tab-output-type-csv-label"): "csv",
+        loc.localize("multi-tab-output-type-parquet-label"): "parquet",
         loc.localize("multi-tab-output-type-kaleidoscope-label"): "kaleidoscope",
     }
 
@@ -249,7 +250,7 @@ def select_directory_wrapper():
                 gr.CheckboxGroup,
                 choices=list(_additional_columns_map().items()),
                 value=[],
-                visible="csv" in output_type_radio.value,
+                visible=bool({"csv", "parquet"} & set(output_type_radio.value)),
                 label=loc.localize("multi-tab-additional-columns-checkbox-label"),
                 info=loc.localize("multi-tab-additional-columns-checkbox-info"),
             )
@@ -297,7 +298,7 @@ def select_directory_wrapper():
         ]
 
         def show_additional_columns(values):
-            return gr.update(visible="csv" in values)
+            return gr.update(visible=bool({"csv", "parquet"} & set(values)))
 
         start_batch_analysis_btn.click(
             run_batch_analysis, inputs=inputs, outputs=result_grid

birdnet_analyzer/gui/train.py

@@ -70,6 +70,7 @@ def start_training(
     upsampling_mode,
     model_formats,
     audio_speed,
+    threads,
     progress=gr.Progress(),
 ):
     """Starts the training of a custom classifier.
@@ -106,7 +107,8 @@ def start_training(
         upsampling_mode: Mode for upsampling (repeat, mean, smote).
         model_formats: Formats to save the trained model (tflite, raven, detached).
         audio_speed: Speed factor for audio playback.
-        save_detached_classifier: Whether to save the detached classifier.
+        threads: Number of parallel CPU threads for loading and decoding the training
+            audio. Ignored when training from a cache file.
     Returns:
         Returns a matplotlib.pyplot figure.
     """
@@ -212,7 +214,9 @@ def trial_progression(trial):
             else None,
             dropout=max(0.0, min(1.0, float(dropout))),
             overlap=max(0.0, min(2.9, float(crop_overlap))),
-            threads=max(1, multiprocessing.cpu_count()),
+            threads=int(threads)
+            if threads and int(threads) > 0
+            else max(1, multiprocessing.cpu_count()),
             on_epoch_end=epoch_progression,
             on_trial_result=trial_progression,
             on_data_load_end=data_load_progression,
@@ -534,6 +538,7 @@ def on_cache_mode_change(value):
                     gr.update(interactive=value != "load"),
                     gr.update(interactive=value != "load"),
                     gr.update(interactive=value != "load"),
+                    gr.update(interactive=value != "load"),
                 )
 
         with (
@@ -620,6 +625,18 @@ def on_cache_mode_change(value):
                     interactive=crops_segments and not uses_cache_file,
                 )
 
+                threads_number = state.persist(
+                    "threads_number",
+                    gr.Number,
+                    value=multiprocessing.cpu_count(),
+                    minimum=1,
+                    step=1,
+                    precision=0,
+                    interactive=not uses_cache_file,
+                    label=loc.localize("training-tab-threads-number-label"),
+                    info=loc.localize("training-tab-threads-number-info"),
+                )
+
             def on_crop_select(new_crop_mode):
                 # Make overlap slider visible for both "segments" and "smart" crop modes
                 return gr.update(
@@ -644,6 +661,7 @@ def on_crop_select(new_crop_mode):
                     audio_speed_slider,
                     crop_mode,
                     crop_overlap,
+                    threads_number,
                 ],
                 show_progress="hidden",
             )
@@ -684,8 +702,13 @@ def on_crop_select(new_crop_mode):
                     info=loc.localize("training-tab-autotune-repeats-number-info"),
                 )
 
+        # The custom-parameters accordion stays visible while autotuning; every field
+        # the tuner searches is disabled instead (its value is only a seed and gets
+        # overridden), so users can still see the values but not waste time tuning them
+        # by hand. "epochs" is not autotuned and stays editable -- it applies to every
+        # trial.
         with (
-            gr.Group(visible=not autotunes) as custom_params,
+            gr.Group(),
             gr.Accordion(
                 open=False,
                 label=loc.localize("training-tab-custom-params-accordion-label"),
@@ -707,6 +730,7 @@ def on_crop_select(new_crop_mode):
                     value=32,
                     minimum=1,
                     step=8,
+                    interactive=not autotunes,
                     label=loc.localize("training-tab-batchsize-number-label"),
                     info=loc.localize("training-tab-batchsize-number-info"),
                 )
@@ -716,6 +740,7 @@ def on_crop_select(new_crop_mode):
                     value=0.0001,
                     minimum=0.0001,
                     step=0.0001,
+                    interactive=not autotunes,
                     label=loc.localize("training-tab-learningrate-number-label"),
                     info=loc.localize("training-tab-learningrate-number-info"),
                 )
@@ -727,6 +752,7 @@ def on_crop_select(new_crop_mode):
                     value=0,
                     minimum=0,
                     step=64,
+                    interactive=not autotunes,
                     label=loc.localize("training-tab-hiddenunits-number-label"),
                     info=loc.localize("training-tab-hiddenunits-number-info"),
                 )
@@ -737,13 +763,15 @@ def on_crop_select(new_crop_mode):
                     minimum=0.0,
                     maximum=0.9,
                     step=0.1,
+                    interactive=not autotunes,
                     label=loc.localize("training-tab-dropout-number-label"),
                     info=loc.localize("training-tab-dropout-number-info"),
                 )
                 use_label_smoothing = state.persist(
                     "use_label_smoothing_checkbox",
                     gr.Checkbox,
                     value=False,
+                    interactive=not autotunes,
                     label=loc.localize(
                         "training-tab-use-labelsmoothing-checkbox-label"
                     ),
@@ -774,6 +802,7 @@ def on_crop_select(new_crop_mode):
                         ),
                     ],
                     value="repeat",
+                    interactive=not autotunes,
                     label=loc.localize("training-tab-upsampling-radio-label"),
                     info=loc.localize("training-tab-upsampling-radio-info"),
                 )
@@ -784,6 +813,7 @@ def on_crop_select(new_crop_mode):
                     maximum=1.0,
                     value=0.0,
                     step=0.05,
+                    interactive=not autotunes,
                     label=loc.localize("training-tab-upsampling-ratio-slider-label"),
                     info=loc.localize("training-tab-upsampling-ratio-slider-info"),
                 )
@@ -793,6 +823,7 @@ def on_crop_select(new_crop_mode):
                     "use_mixup_checkbox",
                     gr.Checkbox,
                     value=False,
+                    interactive=not autotunes,
                     label=loc.localize("training-tab-use-mixup-checkbox-label"),
                     info=loc.localize("training-tab-use-mixup-checkbox-info"),
                     show_label=True,
@@ -801,24 +832,23 @@ def on_crop_select(new_crop_mode):
                     "use_focal_loss_checkbox",
                     gr.Checkbox,
                     value=False,
+                    interactive=not autotunes,
                     label=loc.localize("training-tab-use-focal-loss-checkbox-label"),
                     info=loc.localize("training-tab-use-focal-loss-checkbox-info"),
                     show_label=True,
                 )
 
-            with gr.Row(
-                visible=bool(use_focal_loss.value) and not autotunes
-            ) as focal_loss_params:
+            with gr.Row(visible=bool(use_focal_loss.value)) as focal_loss_params:
                 focal_loss_gamma = state.persist(
                     "focal_loss_gamma_slider",
                     gr.Slider,
                     minimum=0.5,
                     maximum=5.0,
                     value=2.0,
                     step=0.1,
+                    interactive=not autotunes,
                     label=loc.localize("training-tab-focal-loss-gamma-slider-label"),
                     info=loc.localize("training-tab-focal-loss-gamma-slider-info"),
-                    interactive=True,
                 )
                 focal_loss_alpha = state.persist(
                     "focal_loss_alpha_slider",
@@ -827,9 +857,9 @@ def on_crop_select(new_crop_mode):
                     maximum=0.9,
                     value=0.25,
                     step=0.05,
+                    interactive=not autotunes,
                     label=loc.localize("training-tab-focal-loss-alpha-slider-label"),
                     info=loc.localize("training-tab-focal-loss-alpha-slider-info"),
-                    interactive=True,
                 )
 
         def on_focal_loss_change(value):
@@ -842,17 +872,33 @@ def on_focal_loss_change(value):
             show_progress="hidden",
         )
 
+        # Every hyperparameter the tuner searches. When autotuning is on these are
+        # disabled (their values are only seeds and get overridden); "epochs" is not
+        # tuned and stays editable.
+        autotuned_params = [
+            batch_size_number,
+            learning_rate_number,
+            hidden_units_number,
+            dropout_number,
+            use_label_smoothing,
+            upsampling_mode,
+            upsampling_ratio,
+            use_mixup,
+            use_focal_loss,
+            focal_loss_gamma,
+            focal_loss_alpha,
+        ]
+
         def on_autotune_change(value):
-            return (
-                gr.update(visible=not value),
-                gr.update(visible=value),
-                gr.update(visible=not value and use_focal_loss.value),
-            )
+            return [
+                gr.update(visible=value),  # autotune_params
+                *[gr.update(interactive=not value) for _ in autotuned_params],
+            ]
 
         autotune_cb.change(
             on_autotune_change,
             inputs=autotune_cb,
-            outputs=[custom_params, autotune_params, focal_loss_params],
+            outputs=[autotune_params, *autotuned_params],
             show_progress="hidden",
         )
 
@@ -966,6 +1012,7 @@ def train_and_show_metrics(*args):
                 upsampling_mode,
                 output_formats,
                 audio_speed_slider,
+                threads_number,
             ],
             outputs=[train_history_plot, metrics_table],
         )

birdnet_analyzer/lang/de.json

@@ -59,7 +59,7 @@
     "embeddings-tab-select-input-directory-textbox-placeholder": "Kein Eingabeverzeichnis ausgewählt",
     "embeddings-tab-start-button-label": "Embeddings extrahieren",
     "embeddings-tab-title": "Embeddings",
-    "eval-tab-accuracy-checkbox-info": "Accuracy misst den Prozentsatz der richtigen Vorhersagen des Modells.",
+    "eval-tab-accuracy-checkbox-info": "Accuracy ist der Anteil korrekter Vorhersagen. Da die meisten Fenster keinen Ruf enthalten, dominieren echte Negative und die Accuracy bleibt irreführend hoch – daher ist sie standardmäßig deaktiviert.",
     "eval-tab-annotation-col-accordion-label": "Annotationsspalten",
     "eval-tab-annotation-selection-button-label": "Verzeichnis mit Annotationen auswählen",
     "eval-tab-annotation-selection-textbox-placeholder": "Kein Annotationsverzeichnis ausgewählt",
@@ -191,6 +191,7 @@
     "multi-tab-output-type-audacity-label": "Audacity",
     "multi-tab-output-type-csv-label": "CSV",
     "multi-tab-output-type-kaleidoscope-label": "Kaleidoscope",
+    "multi-tab-output-type-parquet-label": "Parquet",
     "multi-tab-output-type-raven-label": "Raven-Auswahltabelle",
     "multi-tab-result-dataframe-column-invalid-file-header": "Ungültige Audiodateien",
     "multi-tab-result-dataframe-column-success-header": "Alle Dateien wurden analysiert!",
@@ -423,6 +424,8 @@
     "training-tab-start-training-button-label": "Training starten",
     "training-tab-test-data-selection-button-label": "Testdaten auswählen (optional)",
     "training-tab-test-data-selection-textbox-placeholder": "Kein Testdatenverzeichnis ausgewählt",
+    "training-tab-threads-number-info": "Anzahl paralleler CPU-Threads zum Laden und Dekodieren der Trainingsaudios. Mehr Threads können das Laden beschleunigen, brauchen aber mehr CPU und Speicher. Wird beim Training aus einer Cache-Datei ignoriert.",
+    "training-tab-threads-number-label": "Threads zum Laden der Daten",
     "training-tab-title": "Trainieren",
     "training-tab-upsampling-radio-info": "Balancieren Sie die Trainingsdaten durch Upsampling von Minderheitenklassen aus.",
     "training-tab-upsampling-radio-label": "Upsampling-Modus",