如何使用Google開源模型SpeciesNet

path = kagglehub.model_download("google/speciesnet/keras/v4.0.0a")

model_file = os.path.join(path, "always_crop_99710272_22x8_v12_epoch_00148.keras")

species_model = tf.keras.models.load_model(model_file)

species_model.summary()

path = kagglehub.model_download("google/speciesnet/keras/v4.0.0a")

model_file = os.path.join(path, "always_crop_99710272_22x8_v12_epoch_00148.keras")

print(model_file);

NEW_CLASS_NAME = "taiwan_black_bear"
LABELS_TXT_PATH =  os.path.join(path, "always_crop_99710272_22x8_v12_epoch_00148.labels.txt")
NEW_DATASET_DIR = "dataset/taiwan_black_bear" 
IMAGE_SIZE = (480, 480)
BATCH_SIZE = 8
EPOCHS = 10


print("[INFO] Loading base model...")
base_model = load_model(model_file)

print("[INFO] Extracting old dense weights...")
old_dense_layer = base_model.layers[-1]
old_weights, old_bias = old_dense_layer.get_weights()

original_class_count = old_weights.shape[1]
new_class_count = original_class_count + 1

print(f"[INFO] Expanding output layer from {original_class_count} to {new_class_count} classes...")

new_weights = np.concatenate(
    [old_weights, np.random.normal(size=(old_weights.shape[0], 1))],
    axis=1
)
new_bias = np.concatenate(
    [old_bias, np.array([0.0])]
)

x = base_model.layers[-2].output  # Dropout 輸出層
new_output = layers.Dense(new_class_count, activation='softmax', name="expanded_output")(x)

model = Model(inputs=base_model.input, outputs=new_output)


model.get_layer("expanded_output").set_weights([new_weights, new_bias])

for layer in model.layers[:-1]:
    layer.trainable = False

print("[INFO] Loading new class images...")
datagen = ImageDataGenerator(rescale=1.0/255)

def generate_bear_dataset(path):
    class_label = new_class_count - 1  # taiwan_black_bear = 2498
    filenames = [os.path.join(path, f) for f in os.listdir(path) if f.lower().endswith((".jpg", ".png"))]
    images = []
    labels = []
    for f in filenames:
        img = tf.keras.utils.load_img(f, target_size=IMAGE_SIZE)
        img_array = tf.keras.utils.img_to_array(img) / 255.0
        images.append(img_array)
        labels.append(class_label)
    return np.array(images), np.array(labels)

X_train, y_train = generate_bear_dataset(NEW_DATASET_DIR)
print(f"[INFO] Loaded {len(X_train)} images for class '{NEW_CLASS_NAME}'.")

model.compile(optimizer=tf.keras.optimizers.Adam(1e-4),
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])

print("[INFO] Training on new class only...")
model.fit(X_train, y_train, batch_size=BATCH_SIZE, epochs=EPOCHS)

print("[INFO] Saving new model as 'speciesnet_with_bear.keras'...")
model.save("speciesnet_with_bear.keras")

new_uuid = str(uuid.uuid4())

NEW_LABEL_LINE = f"{new_uuid};mammalia;carnivora;ursidae;ursus;thibetanus;taiwan_black_bear"

with open(LABELS_TXT_PATH, "r", encoding="utf-8") as f:
    lines = f.readlines()
    if any("taiwan_black_bear" in line for line in lines):
        print("[INFO] Label for 'taiwan_black_bear' already exists in labels.txt, skipping append.")
    else:
        backup_path = LABELS_TXT_PATH + ".backup"
        if not os.path.exists(backup_path):
            with open(backup_path, "w", encoding="utf-8") as backup_file:
                backup_file.writelines(lines)
            print(f"[INFO] Backup created: {backup_path}")

        with open(LABELS_TXT_PATH, "a", encoding="utf-8") as f_append:
            f_append.write("\n" + NEW_LABEL_LINE)
            print(f"[INFO] Appended new class to labels.txt: {new_uuid}")