gemma3n_audio_encoder.py

from transformers import Gemma3nAudioEncoder, Gemma3nConfig
from transformers import AutoFeatureExtractor, PreTrainedModel
from transformers.models.gemma3n.modeling_gemma3n import Gemma3nMultimodalEmbedder

class Audio(PreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.audio_tower = Gemma3nAudioEncoder(config.audio_config)
        self.embed_audio = Gemma3nMultimodalEmbedder(config.audio_config, config.text_config)

class GemmaAudio(PreTrainedModel):
    config_class = Gemma3nConfig
    
    def __init__(self, config):
        super().__init__(config)
        self.model = Audio(config)

    def forward(self, input_features, input_features_mask, **kwargs):
        output = self.model.audio_tower(
            input_features, ~input_features_mask,
        )
        project = self.model.embed_audio(inputs_embeds = output[0])
        return project, output[1]