third_party/InternVideo/InternVideo2/single_modality/README.md

# Single-modality

## Installation

Please follow the installation instructions in [INSTALL](./INSTALL.md).

## Datasets

You can find the dataset instructions in [DATASET](./DATASET.md).

## Model ZOO

You can find all the models and the scripts in [MODEL_ZOO](./MODEL_ZOO.md).

## Pre-Training

We use [InternVL](https://github.com/OpenGVLab/InternVL/) and [VideoMAEv2](https://github.com/OpenGVLab/VideoMAEv2) pretrained models as teachers by default

For training, you can simply run the pretraining scripts in `scripts/pretraining` as follows:
```shell
bash ./scripts/pretraining/1B_pt.sh
```

:warning: **Notes:**
1. Chage `DATA_PATH` to your data path before running the scripts.
2. `--sampling_rate` is set to 1 for **sprase sampling**.
3. The latest checkpoint will be automatically saved while training, thus we use a large `--save_ckpt_freq`.
4. For InternVideo2-1B and 6B, we use InternVL-C-13B and VideoMAEv2-g.


## Finetuning

For finetuning, you can simply run the pretraining scripts in `scripts/finetuning` as follows:
```shell
bash ./scripts/finetuning/full_tuning/k400/1B_ft_k710_ft_k400_f8.sh
```

:warning: **Notes:**
1. Chage `DATA_PATH` And `PREFIX` to your data path before running the scripts.
2. Chage `MODEL_PATH` to your model path.
3. Set `--use_checkpoint` and `--checkpoint_num` to save GPU memory.
4. The best checkpoint will be automatically evaluated with `--test_best`.
5. Set `--test_num_segment` and `--test_num_crop` for different evaluation strategies.
6. To only run evaluation, just set `--eval`.