Download data to todi in video2dataset format.

[kdu4108]

Download HowTo100M and Hd-vila datasets in video2dataset format using the scripts/methods described in https://github.com/swiss-ai/video2dataset/tree/main/swiss_ai.

This will be downloaded in this format:

 ├── 00000.tar
 |     ├── 00000.mp4
 |     ├── 00000.txt
 |     ├── 00000.json
 |     ├── 00001.mp4
 |     ├── 00001.txt
 |     ├── 00001.json
 |     └── ...
 |     ├── 10000.mp4
 |     ├── 10000.txt
 |     ├── 10000.json
 ├── 00001.tar
 |     ├── 10001.mp4
 |     ├── 10001.txt
 |     ├── 10001.json
 │     ...
 ...

Child issue of #3.

[kdu4108]

Update: we have built a Docker image for containerized runs with v2d here: https://github.com/swiss-ai/containers.

We have done small runs to test the right setting for the num_tasks and num_cpus_per_task settings and found that tasks=2 and cpus=128 is optimal (compared to more tasks and fewer cpus).

To download the data fully, we will need to follow these steps:

1. Allocate and enter a compute node
   a. Run salloc -t240 to allocate compute, and note the jobid for that allocation.
2. From login node, create a container based on the Docker image we just created and run the job, e.g. for howto100m:

srun --overlap --jobid <JOBID> --environment=/store/swissai/a08/containers/v2d/v2d.toml --container-workdir=$PWD video2dataset --url_list="/store/swissai/a08/data/raw/howto100m/v2d/howto100m_v2d.csv" --config="/store/swissai/a08/kdu/video2dataset/swiss_ai/configs/download_todi.yaml" --output_folder="/store/swissai/a08/data/raw/howto100m/v2d" --input_format="csv" --output_format="webdataset" --url_col="video_link" --encode_formats="{'video': 'mp4', 'audio':'m4a'}"

The exact download commands for each dataset can be found in the respective readmes of https://github.com/swiss-ai/video2dataset/tree/kdu/todi/swiss_ai/datasets.

[kdu4108]

Most of the code to do this is in the video2dataset repo, so progress and code will be documented in this PR swiss-ai/video2dataset#5.