textvqa1B-largeonly.sh

#!/usr/bin/env bash
set -euo pipefail
set -x

SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
REPO_ROOT="${SCRIPT_DIR}"
cd "${REPO_ROOT}"

export PYTHONPATH="${REPO_ROOT}:${PYTHONPATH:-}"
export UPSTREAM_SGL_ROOT="${UPSTREAM_SGL_ROOT:-/home/yf/snap/SGL}"
export PYTHONPATH="${UPSTREAM_SGL_ROOT}:${PYTHONPATH}"
export PYTHONPATH="${UPSTREAM_SGL_ROOT}/eval/vqa:${PYTHONPATH}"

PYTHON_BIN=${PYTHON_BIN:-python}
DATA_ROOT=${DATA_ROOT:-/home/yf/snap/data/yf}
TEXTVQA_ROOT=${TEXTVQA_ROOT:-"${DATA_ROOT}/textvqa"}
CHECKPOINT_ROOT=${CHECKPOINT_ROOT:-"${DATA_ROOT}"}
CHECKPOINT=${CHECKPOINT:-"${CHECKPOINT_ROOT}/InternVL2-1B"}
OUT_DIR=${OUT_DIR:-"${REPO_ROOT}/outputs/textvqa_largeonly_1b"}
GPUS_PER_MODEL=${GPUS_PER_MODEL:-1}
LIMIT=${LIMIT:-}
LOG_PATH=${LOG_PATH:-"${OUT_DIR}/run.log"}

mkdir -p "${OUT_DIR}"

if [[ "${ENABLE_RUN_LOG:-1}" == "1" ]]; then
    exec > >(tee -a "${LOG_PATH}") 2>&1
fi

CMD=(
    "${PYTHON_BIN}" eval/vqa/run_single_model_native.py
    --checkpoint "${CHECKPOINT}"
    --mode textvqa_eval
    --dataset textvqa_val
    --data-root "${DATA_ROOT}"
    --train-file "${TEXTVQA_ROOT}/textvqa_train.jsonl"
    --test-file "${TEXTVQA_ROOT}/textvqa_val.jsonl"
    --annotation-file "${TEXTVQA_ROOT}/textvqa_val_annotations.json"
    --dynamic
    --out-dir "${OUT_DIR}"
    --run-name textvqa_val_1b
    --gpus-per-model "${GPUS_PER_MODEL}"
)

if [[ -n "${LIMIT}" ]]; then
    CMD+=(--limit "${LIMIT}")
fi

echo "start_time=$(date '+%Y-%m-%d %H:%M:%S')"
echo "checkpoint=${CHECKPOINT}"
echo "data_root=${DATA_ROOT}"
echo "textvqa_root=${TEXTVQA_ROOT}"
echo "out_dir=${OUT_DIR}"
echo "gpus_per_model=${GPUS_PER_MODEL}"
echo "limit=${LIMIT:-full}"
echo

"${CMD[@]}"