how_to_run.txt

# 1. Evaluate using provided fine-tuned model results

# Note: Rename `Dockerfile_eval` to `Dockerfile` before running this command.

# Build the Docker image required for running evaluation.
# It takes approximately 20-25 minutes.
docker build -t emosum .

# Run the docker image
docker run --gpus all -it emosum


# 2. Full pipeline: training + adapter training + evaluation

# Place `Dockerfile_for_train_and_test` in the `Emosum` directory
# Note: Rename `Dockerfile_for_train_and_test` to `Dockerfile` before running this command.

# Build the Docker image required for running evaluation.
docker build -t emosum .

# Run the docker image with volume mounting
docker run --gpus all -it \
  -v $(pwd):/workspace \
  emosum

# Activate the conda environment inside the Docker container
source /opt/conda/etc/profile.d/conda.sh
conda activate emosum

# Change permission to make shell scripts executable
chmod +x train_finetuning.sh train_adapter.sh peft_eval.sh

# Clone required pretrained models from Hugging Face
git clone https://huggingface.co/facebook/bart-large
git clone https://huggingface.co/j-hartmann/emotion-english-roberta-large

# Download large model files using Git LFS
cd bart-large && git lfs pull
cd ../emotion-english-roberta-large && git lfs pull
cd ../

# Step 1. Full fine-tuning of the base model
GPU_DEVICE=0
FFT_OUTPUT=output/fft

./train_finetuning.sh $GPU_DEVICE $FFT_OUTPUT

# Step 2. Train LoRA adapter on top of a specific full fine-tuning checkpoint
GPU_DEVICE=0
# Here, replace 'checkpoint-N' with the actual checkpoint directory name
# Important: If the script fails, ensure the directory name follows the format of `(some_path)/checkpoint-N` (not '[Emotional_...]' or other formats)
# For a quick sanity check, you may also use `FFT_OUTPUT=./ours/BART16`
FFT_OUTPUT=output/fft/checkpoint-N
LORA_OUTPUT=output/lora

./train_adapter.sh $GPU_DEVICE $LORA_OUTPUT $FFT_OUTPUT

# Step 3. Evaluate the trained model with LoRA adapter
GPU_DEVICE=0
# Here, replace 'checkpoint-N' with the actual checkpoint directory name
# Important: If the script fails, ensure the directory name follows the format of `(some_path)/checkpoint-N` (not '[Emotional_...]' or other formats)
# For a quick sanity check, you may also use `BASE_MODEL=./ours/BART16`
BASE_MODEL=output/fft/checkpoint-N
# For a quick sanity check, you may also use `LORA=./ours/lora/summarization_adapter`
LORA=output/lora/summarization_adapter
OUTPUT=eval_result

./peft_eval.sh $GPU_DEVICE $BASE_MODEL $LORA $OUTPUT