textvqaSharedVision-1Bguide-8Btext.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="${UPSTREAM_SGL_ROOT:-/root/SGL}"
export PYTHONPATH="${UPSTREAM_SGL_ROOT}:${PYTHONPATH}"
export PYTHONPATH="${UPSTREAM_SGL_ROOT}/eval/vqa:${PYTHONPATH}"

PYTHON_BIN=${PYTHON_BIN:-/root/miniconda3/envs/sgl/bin/python}
DATA_ROOT=${DATA_ROOT:-/root/data}
TEXTVQA_ROOT=${TEXTVQA_ROOT:-"${DATA_ROOT}/textvqa"}
CHECKPOINT_ROOT=${CHECKPOINT_ROOT:-"${DATA_ROOT}"}
GUIDE_CHECKPOINT=${GUIDE_CHECKPOINT:-"${CHECKPOINT_ROOT}/InternVL2-1B"}
LARGE_CHECKPOINT=${LARGE_CHECKPOINT:-"${CHECKPOINT_ROOT}/InternVL2-8B"}
OUT_DIR=${OUT_DIR:-"${REPO_ROOT}/outputs/textvqa_shared_vision_1bguide_8btext"}
RUN_NAME=${RUN_NAME:-textvqa_shared_vision_1bguide_8btext}
LOG_PATH=${LOG_PATH:-"${OUT_DIR}/run.log"}
PRUNE_LAYER=${PRUNE_LAYER:-0.0}
PRUNE_RATIO=${PRUNE_RATIO:-0.4}
CONSISTENCY_TOKEN_RATIO=${CONSISTENCY_TOKEN_RATIO:-0.05}
LIMIT=${LIMIT:-}
REASONING_MODE=${REASONING_MODE:-none}
REASONING_MAX_NEW_TOKENS=${REASONING_MAX_NEW_TOKENS:-64}
REASONING_TEMPERATURE=${REASONING_TEMPERATURE:-0.0}
SAVE_REASONING=${SAVE_REASONING:-0}
GUIDE_REASONING_MODE=${GUIDE_REASONING_MODE:-none}
GUIDE_REASONING_MAX_NEW_TOKENS=${GUIDE_REASONING_MAX_NEW_TOKENS:-1024}
GUIDE_REASONING_TEMPERATURE=${GUIDE_REASONING_TEMPERATURE:-0.0}
GUIDE_REASONING_FILTER_MODE=${GUIDE_REASONING_FILTER_MODE:-none}
GUIDE_ATTENTION_SOURCE=${GUIDE_ATTENTION_SOURCE:-default}
GUIDE_ATTENTION_AGGREGATION_MODE=${GUIDE_ATTENTION_AGGREGATION_MODE:-raw}
GUIDE_QUESTION_ATTENTION_WEIGHT=${GUIDE_QUESTION_ATTENTION_WEIGHT:-1.0}
GUIDE_REASONING_ATTENTION_WEIGHT=${GUIDE_REASONING_ATTENTION_WEIGHT:-1.0}
GUIDE_ANSWER_ATTENTION_WEIGHT=${GUIDE_ANSWER_ATTENTION_WEIGHT:-1.0}
GUIDE_TEXT_MODE=${GUIDE_TEXT_MODE:-none}
GUIDE_TEXT_MAX_NEW_TOKENS=${GUIDE_TEXT_MAX_NEW_TOKENS:-12}
GUIDE_TEXT_TEMPERATURE=${GUIDE_TEXT_TEMPERATURE:-0.0}
PRUNE_SELECTION_MODE=${PRUNE_SELECTION_MODE:-topk}
SEED=${SEED:-0}

mkdir -p "${OUT_DIR}"

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

EXTRA_ARGS=()
if [[ "${REASONING_MODE}" != "none" ]]; then
    EXTRA_ARGS+=(
        --reasoning-mode "${REASONING_MODE}"
        --reasoning-max-new-tokens "${REASONING_MAX_NEW_TOKENS}"
        --reasoning-temperature "${REASONING_TEMPERATURE}"
    )
fi
if [[ "${SAVE_REASONING}" == "1" ]]; then
    EXTRA_ARGS+=(--save-reasoning)
fi
if [[ "${GUIDE_REASONING_MODE}" != "none" ]]; then
    EXTRA_ARGS+=(
        --guide-reasoning-mode "${GUIDE_REASONING_MODE}"
        --guide-reasoning-max-new-tokens "${GUIDE_REASONING_MAX_NEW_TOKENS}"
        --guide-reasoning-temperature "${GUIDE_REASONING_TEMPERATURE}"
        --guide-reasoning-filter-mode "${GUIDE_REASONING_FILTER_MODE}"
        --guide-attention-source "${GUIDE_ATTENTION_SOURCE}"
        --guide-attention-aggregation-mode "${GUIDE_ATTENTION_AGGREGATION_MODE}"
        --guide-reasoning-attention-weight "${GUIDE_REASONING_ATTENTION_WEIGHT}"
        --guide-answer-attention-weight "${GUIDE_ANSWER_ATTENTION_WEIGHT}"
    )
fi
EXTRA_ARGS+=(
    --guide-question-attention-weight "${GUIDE_QUESTION_ATTENTION_WEIGHT}"
    --guide-answer-attention-weight "${GUIDE_ANSWER_ATTENTION_WEIGHT}"
)
if [[ "${GUIDE_TEXT_MODE}" != "none" ]]; then
    EXTRA_ARGS+=(
        --guide-text-mode "${GUIDE_TEXT_MODE}"
        --guide-text-max-new-tokens "${GUIDE_TEXT_MAX_NEW_TOKENS}"
        --guide-text-temperature "${GUIDE_TEXT_TEMPERATURE}"
    )
fi

echo "start_time=$(date '+%Y-%m-%d %H:%M:%S')"
echo "guide_checkpoint=${GUIDE_CHECKPOINT}"
echo "large_checkpoint=${LARGE_CHECKPOINT}"
echo "data_root=${DATA_ROOT}"
echo "textvqa_root=${TEXTVQA_ROOT}"
echo "out_dir=${OUT_DIR}"
echo "run_name=${RUN_NAME}"
echo "prune_layer=${PRUNE_LAYER}"
echo "prune_ratio=${PRUNE_RATIO}"
echo "prune_selection_mode=${PRUNE_SELECTION_MODE}"
echo "consistency_token_ratio=${CONSISTENCY_TOKEN_RATIO}"
echo "limit=${LIMIT:-full}"
echo "seed=${SEED}"
echo "guide_question_attention_weight=${GUIDE_QUESTION_ATTENTION_WEIGHT}"
echo "guide_answer_attention_weight=${GUIDE_ANSWER_ATTENTION_WEIGHT}"
echo "guide_reasoning_mode=${GUIDE_REASONING_MODE}"
echo "guide_reasoning_filter_mode=${GUIDE_REASONING_FILTER_MODE}"
echo "guide_attention_aggregation_mode=${GUIDE_ATTENTION_AGGREGATION_MODE}"
echo "guide_text_mode=${GUIDE_TEXT_MODE}"
echo

CMD=(
    "${PYTHON_BIN}" eval/vqa/run_shared_vision_guided_textvqa.py
    --guide-checkpoint "${GUIDE_CHECKPOINT}"
    --large-checkpoint "${LARGE_CHECKPOINT}"
    --data-root "${DATA_ROOT}"
    --textvqa-root "${TEXTVQA_ROOT}"
    --dynamic
    --out-dir "${OUT_DIR}"
    --run-name "${RUN_NAME}"
    --large-model-prune-layer "${PRUNE_LAYER}"
    --large-model-prune-ratio "${PRUNE_RATIO}"
    --large-model-prune-selection "${PRUNE_SELECTION_MODE}"
    --consistency-token-ratio "${CONSISTENCY_TOKEN_RATIO}"
    --seed "${SEED}"
)

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

"${CMD[@]}" \
    "${EXTRA_ARGS[@]}"