Time Series and Language

Install

Making the Environment

1. Clone this repo.
2. Build the environment using make create_environment. This can be very slow with stock Conda. If you have Mamba installed it will be much faster.

Note If you're on the UW Slurm Cluster you will need to first load CUDA by running module load cuda/12.2

Getting Data

First, install the huggingface CLI:

 pip install -U "huggingface_hub[cli]"

Next, login:

huggingface-cli login

Finally, run this command to save the data to the appropriate directory:

huggingface-cli download mikeam/time-series-reasoning --repo-type dataset --local-dir data/processed

If you're on the UW Klone cluster then you just need to run make data_on_klone. This will link the bdata directory to the project folder.

Running Jobs:

This project was designed to be run from the command line. Here's an example command:

python src/models/cli.py fit \
    --model="src.models.models.LLaVA" \
    --model.hf_name_or_path="liuhaotian/llava-llama-2-7b-chat-lightning-lora-preview" \
    --model.model_base="meta-llama/Llama-2-7b-chat-hf" \
    --model.batch_size="1" \
    --data="configs/tasks/llms_and_ts/ts2desc_mcq.yaml" \
    --trainer.max_epochs="10" \
    --trainer.log_every_n_steps="100" \
    --trainer.precision="bf16" \
    --trainer.limit_train_batches="100" \
    --optimizer.lr="0.0001" \
    --early_stopping_patience="10" \
    --checkpoint_metric="val/loss" \
    --checkpoint_mode="min" \
    --no_wandb

There's a few things to notice about this command:

1. We're able to pass arguments directly to the model (e.g. model.hf_name_or_path). This is possible because we inherit the LightningModule class, which plays nicely with the LightningCLI.
2. We can also configure the Lightning Trainer.
3. The (optional) --no_wandb flag runs the experiment without logging to Weights and Biases.