Where is the training script

[Lankhonba2024]

Hi,
I am not able to locate the training script, which python file do i need to run to train gemma 3n and which one should i run for inference for my own custom video.
Thanks

[pratik376]

Hi,

Thanks for your question.

I believe, at the moment, the main Gemma repository does not include a standalone training script (e.g., train.py) for training Gemma 3n from scratch. The released code primarily supports model architecture, tokenization, and inference utilities rather than the full pretraining pipeline.

If you're looking to fine-tune Gemma models, you may want to consider using the HuggingFace Transformers ecosystem, where Gemma checkpoints are integrated and easier to fine-tune using standard training loops (e.g., with PEFT/LoRA).

For inference with custom video using Gemma 3n (multimodal), there is no direct script to pass a raw video file. Typically, the workflow would involve:

Extracting frames from the video,

Converting frames into image inputs (or embeddings),

Passing them into the model using the appropriate multimodal tokens (e.g., <start_of_image>).

You can refer to the tokenizer and sampler utilities in gm/text for examples of how text generation is handled.

If you can clarify whether you're aiming for full training or fine-tuning, I’d be happy to provide more specific guidance.

Thanks!

[sonali-kumari1]

Hi @Lankhonba2024,

For training and fine-tuning you can refer fine_tune_gemma3n_on_t4.ipynb notebook in the Hugging Face recipes repository as a starting point, though it would likely need adaptation for video data. Regarding inference, since the model generally processes video as a sequence of image tokens, you would likely need to pass your video as a list of extracted frames. Thanks!