Option for outputting a detached classifier when training

birdnet_analyzer/gui/train.py

@@ -68,6 +68,7 @@ def start_training(
     upsampling_mode,
     model_format,
     audio_speed,
+    save_detached_classifier,
     progress=gr.Progress(),
 ):
     """Starts the training of a custom classifier.
@@ -104,7 +105,7 @@ def start_training(
         upsampling_mode: Mode for upsampling (repeat, mean, smote).
         model_format: Format to save the trained model (tflite, raven, both).
         audio_speed: Speed factor for audio playback.
-
+        save_detached_classifier: Whether to save the detached classifier.
     Returns:
         Returns a matplotlib.pyplot figure.
     """
@@ -146,6 +147,9 @@ def start_training(
     if cache_mode == "load" and not os.path.isfile(cache_file):
         raise gr.Error(loc.localize("validation-no-cache-file-selected"))
 
+    if model_save_mode == "append" and save_detached_classifier:
+        gr.Warning(loc.localize("training-tab-warning-detached-classifier-append"))
+
     def data_load_progression(num_files, num_total_files, label):
         if progress is not None:
             progress(
@@ -218,6 +222,7 @@ def trial_progression(trial):
             autotune_trials=int(autotune_trials),
             autotune_n_splits=int(autotune_folds),
             autotune_n_repeats=int(autotune_repeats),
+            save_detached_classifier=save_detached_classifier,
         )
     except Exception as e:
         if e.args and len(e.args) > 1:
@@ -338,10 +343,10 @@ def selection_fn():
                 rtl=True,
             )
 
-        with gr.Column():
+        with gr.Column(visible=False) as classifier_settings_column:
             classifier_name = gr.Textbox(
                 "CustomClassifier",
-                visible=False,
+                interactive=True,
                 info=loc.localize("training-tab-classifier-textbox-info"),
             )
             output_format = gr.Radio(
@@ -353,7 +358,17 @@ def selection_fn():
                 value="tflite",
                 label=loc.localize("training-tab-output-format-radio-label"),
                 info=loc.localize("training-tab-output-format-radio-info"),
-                visible=False,
+                interactive=True,
+            )
+            save_detached_classifier_checkbox = gr.Checkbox(
+                False,
+                label=loc.localize(
+                    "training-tab-save-detached-classifier-checkbox-label"
+                ),
+                info=loc.localize(
+                    "training-tab-save-detached-classifier-checkbox-info"
+                ),
+                interactive=True,
             )
 
         def select_classifier_directory_and_update_tb():
@@ -363,8 +378,8 @@ def select_classifier_directory_and_update_tb():
                 return (
                     dir_name,
                     dir_name,
-                    gr.update(label=dir_name, visible=True),
-                    gr.update(visible=True, interactive=True),
+                    gr.update(label=dir_name),
+                    gr.update(visible=True),
                 )
 
             return gr.update(), gr.update(), gr.update(), gr.update()
@@ -375,7 +390,7 @@ def select_classifier_directory_and_update_tb():
                 output_directory_state,
                 selected_output_textbox,
                 classifier_name,
-                output_format,
+                classifier_settings_column,
             ],
             show_progress="hidden",
         )
@@ -869,6 +884,7 @@ def train_and_show_metrics(*args):
                 upsampling_mode,
                 output_format,
                 audio_speed_slider,
+                save_detached_classifier_checkbox,
             ],
             outputs=[train_history_plot, metrics_table],
         )

birdnet_analyzer/lang/de.json

@@ -353,6 +353,8 @@
     "training-tab-output-format-radio-info": "Format des trainierten Klassifikators.",
     "training-tab-output-format-radio-label": "Modellausgabeformat",
     "training-tab-preprocessing-accordion-label": "Vorverarbeitung",
+    "training-tab-save-detached-classifier-checkbox-info": "Speichert den Klassifikatorkopf als eigenständiges Modell, um die Dateigröße zu reduzieren",
+    "training-tab-save-detached-classifier-checkbox-label": "Eigenständigen Klassifikator speichern",
     "training-tab-select-output-button-label": "Ausgabeverzeichnis für Klassifikator auswählen",
     "training-tab-select-output-textbox-placeholder": "Kein Ausgabeverzeichnis ausgewählt",
     "training-tab-start-training-button-label": "Training starten",
@@ -367,6 +369,7 @@
     "training-tab-upsampling-radio-option-smote": "SMOTE",
     "training-tab-upsampling-ratio-slider-info": "Das Mindestverhältnis für eine Minderheitenklasse im Vergleich zur Mehrheitsklasse nach dem Upsampling.",
     "training-tab-upsampling-ratio-slider-label": "Upsampling-Verhältnis",
+    "training-tab-warning-detached-classifier-append": "Der \"Anhängen\"-Modus wird für den abgekoppelten Klassifikator nicht angewendet.",
     "training-tab-use-focal-loss-checkbox-info": "Verwendet Focal Loss während des Trainings. Dies gibt schwierigen Beispielen mehr Gewicht und hilft, Klassenungleichgewichte zu bewältigen.",
     "training-tab-use-focal-loss-checkbox-label": "Focal Loss verwenden",
     "training-tab-use-labelsmoothing-checkbox-info": "Wendet Label Smoothing während des Trainings an. Dies kann die Generalisierung des Modells verbessern, indem es verhindert, dass das Modell zu selbstsicher wird.",

birdnet_analyzer/lang/en.json

@@ -353,6 +353,8 @@
     "training-tab-output-format-radio-info": "Format of the trained classifier.",
     "training-tab-output-format-radio-label": "Model output format",
     "training-tab-preprocessing-accordion-label": "Preprocessing",
+    "training-tab-save-detached-classifier-checkbox-info": "Saves the classifier head as detached model to reduce size",
+    "training-tab-save-detached-classifier-checkbox-label": "Save detached classifier",
     "training-tab-select-output-button-label": "Select classifier output",
     "training-tab-select-output-textbox-placeholder": "No output directory selected",
     "training-tab-start-training-button-label": "Start training",
@@ -367,6 +369,7 @@
     "training-tab-upsampling-radio-option-smote": "SMOTE",
     "training-tab-upsampling-ratio-slider-info": "The minimum ratio for a minority class compared to the majority class after upsampling.",
     "training-tab-upsampling-ratio-slider-label": "Upsampling ratio",
+    "training-tab-warning-detached-classifier-append": "Append mode will not apply to the detached classifier.",
     "training-tab-use-focal-loss-checkbox-info": "Uses focal loss during training. This gives more weight to hard examples and helps handle class imbalance.",
     "training-tab-use-focal-loss-checkbox-label": "Use focal loss",
     "training-tab-use-labelsmoothing-checkbox-info": "Applies label smoothing during training. This can help improve model generalization by preventing the model from becoming too confident.",

birdnet_analyzer/lang/fi.json

@@ -353,6 +353,8 @@
     "training-tab-output-format-radio-info": "Koulutetun luokittelijan muoto.",
     "training-tab-output-format-radio-label": "Mallin tulostemuoto",
     "training-tab-preprocessing-accordion-label": "Esikäsittely",
+    "training-tab-save-detached-classifier-checkbox-info": "Tallentaa luokittimen pään erillisenä mallina koon pienentämiseksi",
+    "training-tab-save-detached-classifier-checkbox-label": "Tallenna erillinen luokitin",
     "training-tab-select-output-button-label": "Valitse luokittelijan tuloste",
     "training-tab-select-output-textbox-placeholder": "Tulostehakemistoa ei valittu",
     "training-tab-start-training-button-label": "Aloita koulutus",
@@ -367,6 +369,7 @@
     "training-tab-upsampling-radio-option-smote": "SMOTE",
     "training-tab-upsampling-ratio-slider-info": "Vähemmistöluokan vähimmäissuhde enemmistöluokkaan verrattuna ylösnäytteistyksen jälkeen.",
     "training-tab-upsampling-ratio-slider-label": "Ylösnäytteistyssuhde",
+    "training-tab-warning-detached-classifier-append": "Liitä-tila ei koske irrotettua luokitinta.",
     "training-tab-use-focal-loss-checkbox-info": "Käyttää focal lossia koulutuksen aikana. Tämä antaa enemmän painoarvoa vaikeille esimerkeille ja auttaa käsittelemään luokkien epätasapainoa.",
     "training-tab-use-focal-loss-checkbox-label": "Käytä focal lossia",
     "training-tab-use-labelsmoothing-checkbox-info": "Soveltaa label smoothingia koulutuksen aikana. Tämä voi parantaa mallin yleistämistä estämällä mallia tulemasta liian itsevarmaksi.",

birdnet_analyzer/lang/fr.json

@@ -353,6 +353,8 @@
     "training-tab-output-format-radio-info": "Format du classificateur formé.",
     "training-tab-output-format-radio-label": "Format de sortie du modèle",
     "training-tab-preprocessing-accordion-label": "Prétraitement",
+    "training-tab-save-detached-classifier-checkbox-info": "Enregistre la tête du classificateur comme modèle détaché pour réduire la taille",
+    "training-tab-save-detached-classifier-checkbox-label": "Enregistrer le classificateur détaché",
     "training-tab-select-output-button-label": "Sélectionner la sortie du classificateur",
     "training-tab-select-output-textbox-placeholder": "Aucun répertoire de sortie sélectionné",
     "training-tab-start-training-button-label": "Commencer l’apprentissage",
@@ -367,6 +369,7 @@
     "training-tab-upsampling-radio-option-smote": "SMOTE",
     "training-tab-upsampling-ratio-slider-info": "Le ratio minimum pour une classe minoritaire par rapport à la classe majoritaire après le rééchantillonnage.",
     "training-tab-upsampling-ratio-slider-label": "Rapport de suréchantillonnage",
+    "training-tab-warning-detached-classifier-append": "Le mode d'ajout ne s'appliquera pas au classificateur détaché.",
     "training-tab-use-focal-loss-checkbox-info": "Utilise la perte focale pendant l'entraînement. Cela donne plus de poids aux exemples difficiles et aide à gérer le déséquilibre des classes.",
     "training-tab-use-focal-loss-checkbox-label": "Utiliser la perte focale",
     "training-tab-use-labelsmoothing-checkbox-info": "Applique le lissage des étiquettes pendant l'entraînement. Cela peut améliorer la généralisation du modèle en empêchant le modèle de devenir trop confiant.",

birdnet_analyzer/lang/id.json

@@ -353,6 +353,8 @@
     "training-tab-output-format-radio-info": "Format klasifikator yang telah dilatih.",
     "training-tab-output-format-radio-label": "Format output model",
     "training-tab-preprocessing-accordion-label": "Praproses",
+    "training-tab-save-detached-classifier-checkbox-info": "Menyimpan kepala pengklasifikasi sebagai model terpisah untuk mengurangi ukuran",
+    "training-tab-save-detached-classifier-checkbox-label": "Simpan pengklasifikasi terpisah",
     "training-tab-select-output-button-label": "Pilih klasifikator output",
     "training-tab-select-output-textbox-placeholder": "Tidak ada direktori output yang dipilih",
     "training-tab-start-training-button-label": "Mulai pelatihan",
@@ -367,6 +369,7 @@
     "training-tab-upsampling-radio-option-smote": "SMOTE",
     "training-tab-upsampling-ratio-slider-info": "Rasio minimum untuk kelas minoritas dibandingkan dengan kelas mayoritas setelah peningkatan sampel.",
     "training-tab-upsampling-ratio-slider-label": "Rasio peningkatan sampel",
+    "training-tab-warning-detached-classifier-append": "Mode tambah tidak akan berlaku untuk pengklasifikasi terlepas.",
     "training-tab-use-focal-loss-checkbox-info": "Menggunakan focal loss selama pelatihan. Ini memberikan bobot lebih pada contoh yang sulit dan membantu menangani ketidakseimbangan kelas.",
     "training-tab-use-focal-loss-checkbox-label": "Gunakan focal loss",
     "training-tab-use-labelsmoothing-checkbox-info": "Menerapkan label smoothing selama pelatihan. Ini dapat membantu meningkatkan generalisasi model dengan mencegah model menjadi terlalu percaya diri.",

birdnet_analyzer/lang/pt-br.json

@@ -353,6 +353,8 @@
     "training-tab-output-format-radio-info": "Formato do classificador treinado.",
     "training-tab-output-format-radio-label": "Formato da saída do modelo",
     "training-tab-preprocessing-accordion-label": "Pré-processamento",
+    "training-tab-save-detached-classifier-checkbox-info": "Salva a cabeça do classificador como modelo desacoplado para reduzir o tamanho",
+    "training-tab-save-detached-classifier-checkbox-label": "Salvar classificador desacoplado",
     "training-tab-select-output-button-label": "Selecionar a saída do classificador",
     "training-tab-select-output-textbox-placeholder": "Nenhum diretório de saída selecionado",
     "training-tab-start-training-button-label": "Comece o treinamento",
@@ -367,6 +369,7 @@
     "training-tab-upsampling-radio-option-smote": "SMOTE",
     "training-tab-upsampling-ratio-slider-info": "A proporção mínima para uma classe minoritária em comparação com a classe majoritária após a amostragem adicional.",
     "training-tab-upsampling-ratio-slider-label": "Taxa de upsampling",
+    "training-tab-warning-detached-classifier-append": "O modo de acréscimo não será aplicado ao classificador desanexado.",
     "training-tab-use-focal-loss-checkbox-info": "Usa perda focal durante o treinamento. Isso dá mais peso a exemplos difíceis e ajuda a lidar com o desequilíbrio de classes.",
     "training-tab-use-focal-loss-checkbox-label": "Usar perda focal",
     "training-tab-use-labelsmoothing-checkbox-info": "Aplica suavização de rótulos durante o treinamento. Isso pode melhorar a generalização do modelo, evitando que ele se torne excessivamente confiante.",

birdnet_analyzer/lang/ru.json

@@ -353,6 +353,8 @@
     "training-tab-output-format-radio-info": "Формат обученного классификатора.",
     "training-tab-output-format-radio-label": "Формат вывода модели",
     "training-tab-preprocessing-accordion-label": "Предобработка",
+    "training-tab-save-detached-classifier-checkbox-info": "Сохраняет голову классификатора как отдельную модель для уменьшения размера",
+    "training-tab-save-detached-classifier-checkbox-label": "Сохранить отдельный классификатор",
     "training-tab-select-output-button-label": "Выбор выходных данных классификатора",
     "training-tab-select-output-textbox-placeholder": "Каталог вывода не выбран",
     "training-tab-start-training-button-label": "Начало обучения",
@@ -367,6 +369,7 @@
     "training-tab-upsampling-radio-option-smote": "SMOTE",
     "training-tab-upsampling-ratio-slider-info": "Минимальный коэффициент для меньшего класса по сравнению с большим классом после повышения выборки.",
     "training-tab-upsampling-ratio-slider-label": "Коэффициент повышения дискретизации",
+    "training-tab-warning-detached-classifier-append": "Режим добавления не применяется к отсоединённому классификатору.",
     "training-tab-use-focal-loss-checkbox-info": "Использует фокальную потерю во время обучения. Это придает больший вес сложным примерам и помогает справляться с дисбалансом классов.",
     "training-tab-use-focal-loss-checkbox-label": "Использовать фокальную потерю",
     "training-tab-use-labelsmoothing-checkbox-info": "Применяет сглаживание меток во время обучения. Это может улучшить обобщение модели, предотвращая излишнюю уверенность модели.",

birdnet_analyzer/lang/se.json

@@ -353,6 +353,8 @@
     "training-tab-output-format-radio-info": "Format för den tränade klassificeraren.",
     "training-tab-output-format-radio-label": "Modell utdataformat",
     "training-tab-preprocessing-accordion-label": "Förbehandling",
+    "training-tab-save-detached-classifier-checkbox-info": "Sparar klassificerarens huvud som en fristående modell för att minska storleken",
+    "training-tab-save-detached-classifier-checkbox-label": "Spara fristående klassificerare",
     "training-tab-select-output-button-label": "Välj klassificerarens utdata",
     "training-tab-select-output-textbox-placeholder": "Ingen utdatakatalog vald",
     "training-tab-start-training-button-label": "Starta träning",
@@ -367,6 +369,7 @@
     "training-tab-upsampling-radio-option-smote": "SMOTE",
     "training-tab-upsampling-ratio-slider-info": "Minsta förhållandet för en minoritetsklass jämfört med majoritetsklass efter upsampling.",
     "training-tab-upsampling-ratio-slider-label": "Upsamplingsförhållande",
+    "training-tab-warning-detached-classifier-append": "Tilläggsläget gäller inte för den fristående klassificeraren.",
     "training-tab-use-focal-loss-checkbox-info": "Använder fokalförlust under träning. Detta ger mer vikt till svåra exempel och hjälper till att hantera klassobalans.",
     "training-tab-use-focal-loss-checkbox-label": "Använd fokalförlust",
     "training-tab-use-labelsmoothing-checkbox-info": "Tillämpa label smoothing under träning. Detta kan förbättra modellens generalisering genom att förhindra att modellen blir för självsäker.",

birdnet_analyzer/lang/tlh.json

@@ -353,6 +353,8 @@
     "training-tab-output-format-radio-info": "Hol chu' qel.",
     "training-tab-output-format-radio-label": "tu'law' Hol",
     "training-tab-preprocessing-accordion-label": "ngoq nung",
+    "training-tab-save-detached-classifier-checkbox-info": "chenmoHwI' nach toD qach mIw nIb chenmoH nap",
+    "training-tab-save-detached-classifier-checkbox-label": "toD chenmoHwI' HaD",
     "training-tab-select-output-button-label": "tu'law' wav",
     "training-tab-select-output-textbox-placeholder": "ngoq Daq wIvlu'be'",
     "training-tab-start-training-button-label": "qeq yItagh",
@@ -367,6 +369,7 @@
     "training-tab-upsampling-radio-option-smote": "SMOTE",
     "training-tab-upsampling-ratio-slider-info": "chu' ratio Seghmey.",
     "training-tab-upsampling-ratio-slider-label": "qochHa' ratio",
+    "training-tab-warning-detached-classifier-append": "chel mIw DetlhHa' ghovwI'vaD lInglaHbe'.",
     "training-tab-use-focal-loss-checkbox-info": "qeq focal loss lo'. qeq QatlhmoH 'ej qeq laH QaQmoH.",
     "training-tab-use-focal-loss-checkbox-label": "focal loss lo'",
     "training-tab-use-labelsmoothing-checkbox-info": "qeq label smoothing lo'. qeq laH QaQmoH 'ej qeq laH QatlhmoH.",

birdnet_analyzer/lang/zh_TW.json

@@ -353,6 +353,8 @@
     "training-tab-output-format-radio-info": "訓練後分類器的格式",
     "training-tab-output-format-radio-label": "模型輸出格式",
     "training-tab-preprocessing-accordion-label": "前處理",
+    "training-tab-save-detached-classifier-checkbox-info": "將分類器頭部儲存為分離模型以縮減大小",
+    "training-tab-save-detached-classifier-checkbox-label": "儲存分離的分類器",
     "training-tab-select-output-button-label": "選擇分類器輸出資料夾",
     "training-tab-select-output-textbox-placeholder": "未選擇輸出目錄",
     "training-tab-start-training-button-label": "開始訓練",
@@ -367,6 +369,7 @@
     "training-tab-upsampling-radio-option-smote": "SMOTE",
     "training-tab-upsampling-ratio-slider-info": "在上取樣後，最少數類別與最多數類別的資料數量最小比例",
     "training-tab-upsampling-ratio-slider-label": "上取樣比例",
+    "training-tab-warning-detached-classifier-append": "附加模式不適用於分離的分類器。",
     "training-tab-use-focal-loss-checkbox-info": "在訓練期間使用焦點損失。這給予困難樣本更多的權重，並有助於處理類別不平衡。",
     "training-tab-use-focal-loss-checkbox-label": "使用焦點損失",
     "training-tab-use-labelsmoothing-checkbox-info": "在訓練期間應用標籤平滑。這可以通過防止模型過於自信來幫助改善模型的泛化能力。",