#!/bin/bash # Run evaluation set # Usage: bash scripts/eval.sh --lora path/to/lora.safetensors --eval expressive --output eval_output/ SCRIPT_DIR="$(cd "$(dirname "$0")/.." && pwd)" LORA="" EVAL_TYPE="short" OUTPUT_DIR="eval_output" GPUS=8 while [[ $# -gt 0 ]]; do case $1 in --lora) LORA="$2"; shift 2;; --eval) EVAL_TYPE="$2"; shift 2;; --output) OUTPUT_DIR="$2"; shift 2;; --gpus) GPUS="$2"; shift 2;; *) echo "Unknown arg: $1"; exit 1;; esac done case $EVAL_TYPE in short) CONFIG="$SCRIPT_DIR/evals/eval_short.txt";; long) CONFIG="$SCRIPT_DIR/evals/eval_long.txt";; stress) CONFIG="$SCRIPT_DIR/evals/eval_stress.txt";; expressive) CONFIG="$SCRIPT_DIR/evals/eval_expressive.txt";; multilang) CONFIG="$SCRIPT_DIR/evals/eval_multilang.txt";; *) echo "Unknown eval type: $EVAL_TYPE (use: short, long, stress, expressive, multilang)"; exit 1;; esac mkdir -p "$OUTPUT_DIR"/{generated,refs} gpu=0 mapfile -t LINES < <(grep -v '^#' "$CONFIG" | grep '|') for line in "${LINES[@]}"; do name=$(echo "$line" | cut -d'|' -f1 | xargs) ref=$(echo "$line" | cut -d'|' -f2 | xargs) prompt=$(echo "$line" | cut -d'|' -f3-) prompt="${prompt## }" [ -z "$name" ] && continue [[ "$ref" == *.mp3 ]] || [[ "$ref" == *.flac ]] && \ ffmpeg -y -i "$ref" "$OUTPUT_DIR/refs/${name}.wav" -loglevel quiet 2>/dev/null || \ cp "$ref" "$OUTPUT_DIR/refs/${name}.wav" 2>/dev/null LORA_FLAG="" [ -n "$LORA" ] && LORA_FLAG="--lora $LORA" CUDA_VISIBLE_DEVICES=$gpu python "$SCRIPT_DIR/src/inference.py" \ --checkpoint ltx-2.3-22b-dev-audio-only.safetensors \ --full-checkpoint ltx-2.3-22b-dev.safetensors \ $LORA_FLAG --voice-sample "$ref" --prompt "$prompt" \ --output "$OUTPUT_DIR/generated/${name}.wav" \ --cfg-scale 2.5 --stg-scale 1.5 --rescale-scale 0 --modality-scale 1 \ --duration-multiplier 1.1 --seed 42 2>&1 | grep "Output" & gpu=$(( (gpu + 1) % GPUS )) if [ $gpu -eq 0 ]; then wait; fi done wait echo "Eval done: $(ls "$OUTPUT_DIR/generated/"*.wav 2>/dev/null | wc -l) samples in $OUTPUT_DIR/"