run_textvqa_shared_vision_ablation.sh

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

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}"

PYTHON_BIN=${PYTHON_BIN:-/home/yf/.conda/envs/sgl/bin/python}
GPU_ID=${GPU_ID:-0}
GUIDE_CHECKPOINT=${GUIDE_CHECKPOINT:-/home/yf/snap/data/yf/InternVL2-2B}
LARGE_CHECKPOINT=${LARGE_CHECKPOINT:-/home/yf/snap/data/yf/InternVL2-8B}
DATA_ROOT=${DATA_ROOT:-/home/yf/snap/data/yf}
TEXTVQA_ROOT=${TEXTVQA_ROOT:-/home/yf/snap/data/yf/textvqa}
OUT_ROOT=${OUT_ROOT:-"${REPO_ROOT}/outputs/shared_vision_ablation"}
RUN_PREFIX=${RUN_PREFIX:-textvqa_shared_vision_ablation}
PRUNE_LAYER=${PRUNE_LAYER:-0.0}
PRUNE_RATIOS=${PRUNE_RATIOS:-0.4}
CONSISTENCY_TOKEN_RATIO=${CONSISTENCY_TOKEN_RATIO:-0.05}
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=${SEED:-0}
LIMIT=${LIMIT:-}
LOG_EVERY=${LOG_EVERY:-20}
DYNAMIC=${DYNAMIC:-1}
SAVE_REASONING=${SAVE_REASONING:-0}
SAVE_VISUAL_TOKEN_IMPORTANCE=${SAVE_VISUAL_TOKEN_IMPORTANCE:-0}
VISUAL_TOKEN_IMPORTANCE_TOPK=${VISUAL_TOKEN_IMPORTANCE_TOPK:-16}
USE_FLASH_ATTN=${USE_FLASH_ATTN:-0}
AUTO=${AUTO:-0}
LOAD_IN_8BIT=${LOAD_IN_8BIT:-0}
LOAD_IN_4BIT=${LOAD_IN_4BIT:-0}
ABLATION_MODES=${ABLATION_MODES:-question_only answer_only}
COMBINED_QUESTION_WEIGHT=${COMBINED_QUESTION_WEIGHT:-1.0}
COMBINED_ANSWER_WEIGHT=${COMBINED_ANSWER_WEIGHT:-1.0}

build_mode_args() {
    local mode="$1"
    case "${mode}" in
        baseline)
            printf '%s\n' \
                --guide-question-attention-weight 1.0 \
                --guide-answer-attention-weight 1.0
            ;;
        question_only)
            printf '%s\n' \
                --guide-question-attention-weight 1.0 \
                --guide-answer-attention-weight 0.0
            ;;
        answer_only)
            printf '%s\n' \
                --guide-question-attention-weight 0.0 \
                --guide-answer-attention-weight 1.0
            ;;
        combined)
            printf '%s\n' \
                --guide-question-attention-weight "${COMBINED_QUESTION_WEIGHT}" \
                --guide-answer-attention-weight "${COMBINED_ANSWER_WEIGHT}"
            ;;
        *)
            echo "Unsupported ablation mode: ${mode}" >&2
            return 1
            ;;
    esac
}

run_one() {
    local mode="$1"
    local prune_ratio="$2"
    local ratio_tag
    ratio_tag="$(printf '%s' "${prune_ratio}" | tr '.' 'p')"
    local run_name="${RUN_PREFIX}_${mode}_prune${ratio_tag}"
    local out_dir="${OUT_ROOT}/${run_name}"
    local log_path="${out_dir}/${run_name}.log"

    mkdir -p "${out_dir}"

    mapfile -t mode_args < <(build_mode_args "${mode}")

    {
        echo "start_time=$(date '+%Y-%m-%d %H:%M:%S')"
        echo "mode=${mode}"
        echo "gpu_id=${GPU_ID}"
        echo "guide_checkpoint=${GUIDE_CHECKPOINT}"
        echo "large_checkpoint=${LARGE_CHECKPOINT}"
        echo "data_root=${DATA_ROOT}"
        echo "textvqa_root=${TEXTVQA_ROOT}"
        echo "prune_layer=${PRUNE_LAYER}"
        echo "prune_ratio=${prune_ratio}"
        echo "prune_selection_mode=${PRUNE_SELECTION_MODE}"
        echo "consistency_token_ratio=${CONSISTENCY_TOKEN_RATIO}"
        echo "guide_text_mode=${GUIDE_TEXT_MODE}"
        echo "guide_text_max_new_tokens=${GUIDE_TEXT_MAX_NEW_TOKENS}"
        echo "guide_text_temperature=${GUIDE_TEXT_TEMPERATURE}"
        echo "combined_question_weight=${COMBINED_QUESTION_WEIGHT}"
        echo "combined_answer_weight=${COMBINED_ANSWER_WEIGHT}"
        echo "seed=${SEED}"
        if [[ -n "${LIMIT}" ]]; then
            echo "limit=${LIMIT}"
        fi
        echo
    } >> "${log_path}"

    local extra_args=()
    if [[ "${DYNAMIC}" == "1" ]]; then
        extra_args+=(--dynamic)
    fi
    if [[ -n "${LIMIT}" ]]; then
        extra_args+=(--limit "${LIMIT}")
    fi
    if [[ "${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 "${GUIDE_TEXT_TEMPERATURE}"
        )
    fi
    if [[ "${SAVE_REASONING}" == "1" ]]; then
        extra_args+=(--save-reasoning)
    fi
    if [[ "${SAVE_VISUAL_TOKEN_IMPORTANCE}" == "1" ]]; then
        extra_args+=(
            --save-visual-token-importance
            --visual-token-importance-topk "${VISUAL_TOKEN_IMPORTANCE_TOPK}"
        )
    fi
    if [[ "${USE_FLASH_ATTN}" == "1" ]]; then
        extra_args+=(--use-flash-attn)
    fi
    if [[ "${AUTO}" == "1" ]]; then
        extra_args+=(--auto)
    fi
    if [[ "${LOAD_IN_8BIT}" == "1" ]]; then
        extra_args+=(--load-in-8bit)
    fi
    if [[ "${LOAD_IN_4BIT}" == "1" ]]; then
        extra_args+=(--load-in-4bit)
    fi

    CUDA_VISIBLE_DEVICES="${GPU_ID}" \
    "${PYTHON_BIN}" "${REPO_ROOT}/eval/vqa/run_shared_vision_guided_textvqa.py" \
        --guide-checkpoint "${GUIDE_CHECKPOINT}" \
        --large-checkpoint "${LARGE_CHECKPOINT}" \
        --data-root "${DATA_ROOT}" \
        --textvqa-root "${TEXTVQA_ROOT}" \
        --out-dir "${out_dir}" \
        --run-name "${run_name}" \
        --log-every "${LOG_EVERY}" \
        --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 "${SEED}" \
        "${extra_args[@]}" \
        "${mode_args[@]}" \
        >> "${log_path}" 2>&1
}

IFS=' ' read -r -a ablation_modes <<< "${ABLATION_MODES}"
IFS=' ' read -r -a prune_ratios <<< "${PRUNE_RATIOS}"

for prune_ratio in "${prune_ratios[@]}"; do
    for mode in "${ablation_modes[@]}"; do
        echo "Running mode=${mode} prune_ratio=${prune_ratio}"
        run_one "${mode}" "${prune_ratio}"
    done
done

echo
echo "Ablation runs finished. Summaries under: ${OUT_ROOT}"
echo "To compare results:"
echo "\"${PYTHON_BIN}\" \"${REPO_ROOT}/tools/summarize_shared_vision_ablation.py\" --input-dir \"${OUT_ROOT}\""