familienarchiv/scripts/download-kraken-models.sh

#!/bin/bash
set -euo pipefail

# Downloads Kraken HTR models, runs OCR evaluation on sample PDFs, and activates the winner.
#
# Usage:
#   ./scripts/download-kraken-models.sh                # download models + run evaluation
#   ./scripts/download-kraken-models.sh --activate 1   # activate model 1
#   ./scripts/download-kraken-models.sh --activate 2   # activate model 2
#   ./scripts/download-kraken-models.sh --eval-only    # re-run evaluation (models already downloaded)

COMPOSE_SERVICE="ocr-service"
MODEL_DIR="/app/models"
ACTIVE_MODEL="$MODEL_DIR/german_kurrent.mlmodel"

# Kraken 7 uses DOIs to identify models from HTR-United / Zenodo
MODEL_1_DOI="10.5281/zenodo.7933463"
MODEL_1_NAME="german_handwriting"
MODEL_1_DESC="HTR model for German manuscripts (handwritten text recognition)"
MODEL_1_PATH="$MODEL_DIR/$MODEL_1_NAME.mlmodel"

MODEL_2_DOI="10.5281/zenodo.13788177"
MODEL_2_NAME="mccatmus"
MODEL_2_DESC="McCATMuS — generic model for handwritten, printed & typewritten (16th c. onward)"
MODEL_2_PATH="$MODEL_DIR/$MODEL_2_NAME.mlmodel"

EVAL_PDFS=("Eu-0693.pdf" "Eu-0692.pdf" "W-0150.pdf" "W-0575.pdf")
IMPORT_DIR="./import"
OUTPUT_BASE="./ocr-model-evaluation"

# ─── Functions ────────────────────────────────────────────────────────────────

download_model() {
    local doi="$1"
    local dest="$2"
    local name="$3"

    echo "  Downloading $name ($doi)..."

    # kraken get downloads to /root/.local/share/htrmopo/<uuid>/
    # We find the .mlmodel file after download and copy it to our volume
    docker compose run --rm "$COMPOSE_SERVICE" sh -c "
        kraken get $doi 2>&1
        # Find the most recently downloaded .mlmodel and copy to target
        FOUND=\$(find /root/.local/share/htrmopo -name '*.mlmodel' -newer /tmp 2>/dev/null | head -1)
        if [ -n \"\$FOUND\" ]; then
            cp \"\$FOUND\" $dest
            echo \"Saved to $dest\"
        else
            echo 'ERROR: No .mlmodel file found after download'
            exit 1
        fi
    "
}

download_models() {
    echo "Downloading Kraken HTR models into the ocr_models volume..."
    echo ""

    # Create a timestamp marker so we can find newly downloaded files
    docker compose run --rm "$COMPOSE_SERVICE" touch /tmp/.download-marker

    echo "Model 1: $MODEL_1_NAME"
    echo "  $MODEL_1_DESC"
    download_model "$MODEL_1_DOI" "$MODEL_1_PATH" "$MODEL_1_NAME"
    echo ""

    echo "Model 2: $MODEL_2_NAME"
    echo "  $MODEL_2_DESC"
    download_model "$MODEL_2_DOI" "$MODEL_2_PATH" "$MODEL_2_NAME"
    echo ""

    echo "Both models downloaded."
}

run_evaluation() {
    echo ""
    echo "═══════════════════════════════════════════════════════"
    echo "  Running OCR evaluation on ${#EVAL_PDFS[@]} documents"
    echo "═══════════════════════════════════════════════════════"
    echo ""

    local dir_1="$OUTPUT_BASE/$MODEL_1_NAME"
    local dir_2="$OUTPUT_BASE/$MODEL_2_NAME"
    mkdir -p "$dir_1" "$dir_2"

    for pdf in "${EVAL_PDFS[@]}"; do
        local src="$IMPORT_DIR/$pdf"
        local basename="${pdf%.pdf}"

        if [[ ! -f "$src" ]]; then
            echo "SKIP: $src not found"
            continue
        fi

        echo "──── $pdf ────"

        echo "  Model 1: $MODEL_1_NAME ..."
        docker compose run --rm \
            -v "$(cd "$IMPORT_DIR" && pwd):/eval-input:ro" \
            -v "$(cd "$dir_1" && pwd):/eval-output" \
            "$COMPOSE_SERVICE" \
            kraken -f pdf -i "/eval-input/$pdf" "/eval-output/${basename}.txt" segment -bl ocr -m "$MODEL_1_PATH" \
            2>/dev/null || echo "  ⚠ Model 1 failed on $pdf"

        echo "  Model 2: $MODEL_2_NAME ..."
        docker compose run --rm \
            -v "$(cd "$IMPORT_DIR" && pwd):/eval-input:ro" \
            -v "$(cd "$dir_2" && pwd):/eval-output" \
            "$COMPOSE_SERVICE" \
            kraken -f pdf -i "/eval-input/$pdf" "/eval-output/${basename}.txt" segment -bl ocr -m "$MODEL_2_PATH" \
            2>/dev/null || echo "  ⚠ Model 2 failed on $pdf"

        echo ""
    done

    echo "═══════════════════════════════════════════════════════"
    echo "  Evaluation complete. Results:"
    echo ""
    echo "  Model 1 — $MODEL_1_NAME ($MODEL_1_DOI):"
    for f in "$dir_1"/*.txt; do
        [[ -f "$f" ]] && echo "    $(basename "$f"): $(wc -l < "$f") lines, $(wc -c < "$f") bytes"
    done
    echo ""
    echo "  Model 2 — $MODEL_2_NAME ($MODEL_2_DOI):"
    for f in "$dir_2"/*.txt; do
        [[ -f "$f" ]] && echo "    $(basename "$f"): $(wc -l < "$f") lines, $(wc -c < "$f") bytes"
    done
    echo ""
    echo "  Compare outputs:"
    echo "    diff $dir_1/Eu-0693.txt $dir_2/Eu-0693.txt"
    echo "    # or view individually:"
    echo "    cat $dir_1/Eu-0693.txt"
    echo ""
    echo "  Activate the better model:"
    echo "    ./scripts/download-kraken-models.sh --activate 1   # $MODEL_1_NAME"
    echo "    ./scripts/download-kraken-models.sh --activate 2   # $MODEL_2_NAME"
    echo "═══════════════════════════════════════════════════════"
}

activate_model() {
    local choice="$1"
    case "$choice" in
        1)
            echo "Activating model 1: $MODEL_1_NAME"
            docker compose run --rm "$COMPOSE_SERVICE" cp "$MODEL_1_PATH" "$ACTIVE_MODEL"
            ;;
        2)
            echo "Activating model 2: $MODEL_2_NAME"
            docker compose run --rm "$COMPOSE_SERVICE" cp "$MODEL_2_PATH" "$ACTIVE_MODEL"
            ;;
        *)
            echo "Error: --activate expects 1 or 2"
            exit 1
            ;;
    esac

    echo "Active model is now: $ACTIVE_MODEL"
    echo "Restart the OCR service to load the new model:"
    echo "  docker compose restart ocr-service"
}

# ─── Main ─────────────────────────────────────────────────────────────────────

case "${1:-}" in
    --activate)
        activate_model "${2:-}"
        ;;
    --eval-only)
        run_evaluation
        ;;
    *)
        download_models
        run_evaluation
        ;;
esac