#!/bin/bash
# =============================================================================
# gradient_ascent_reverse.sh — toilet→bathroom causal influence map via probe gradient.
#
# REVERSED direction: steer = toilet probe, readout = bathroom probe.
# Images: toilet-only (toilet=1, bathroom=0).
#
# For each TOILET-ONLY image: nudge the residual stream at layer l along
# ĝ = ∂ score_toilet_l/∂ h_l (unit gradient ascent in the toilet direction),
# propagate, and measure Δ bathroom score at every layer l' >= l.
# Averaged → heatmap (intervention layer l × readout l').
#
# Tunables (override inline, e.g. DEVICE_ID=1 bash <script>):
#   DEVICE_ID       GPU id                                  (default 0)
#   DTYPE           bfloat16 | float16 | float32            (default bfloat16)
#   BATH_PROBE      steered-direction checkpoint = TOILET probe in reverse
#   TOILET_PROBE    measured-readout  checkpoint = BATHROOM probe in reverse
#   IMAGE_FOLDER    folder of images
#   SAMPLES_JSON    samples.json path
#   BASE_PROMPT     prompt key in samples.json
#   STEER_NAME      display name of steered concept (plots)   (default toilet)
#   READOUT_NAME    display name of readout concept (plots)   (default bathroom)
#   QUESTION        prompt question
#   NUM_IMAGES      cap on images; 0 = ALL toilet-only        (default 0)
#   ALPHAS          steps as FRACTION of ‖h_l‖               (default "0.05 0.1 0.2 0.4 0.8")
#   HOOK_TYPE       pre | mid | post                          (default post)
#   MAX_NEW_TOKENS  caption generation length                 (default 64)
#   MAX_SEQ_TOKENS  caption tokens kept for the probe         (default 64)
#   FORCED_TEXT     if set, force ASSISTANT answer to this string
#   OUT / OUT_JSON  output heatmap PNG / matrix JSON
# =============================================================================

export HF_HOME="/data/caotue/hf_cache"
export HF_DATASETS_CACHE="/data/caotue/hf_cache/datasets"
export TORCH_HOME="/data/caotue/torch_cache"
export TMPDIR="/data/caotue/tmp"

DEVICE_ID="${DEVICE_ID:-0}"
DTYPE="${DTYPE:-bfloat16}"
# REVERSED: BATH_PROBE = toilet probe (steer), TOILET_PROBE = bathroom probe (readout)
BATH_PROBE="${BATH_PROBE:-/data/caotue/latent_probes/seqprobes_toilet/post/seqprobe.pth}"
TOILET_PROBE="${TOILET_PROBE:-/data/caotue/latent_probes/seqprobes_bathroom/post/seqprobe.pth}"
IMAGE_FOLDER="${IMAGE_FOLDER:-/data/caotue/CC3M-Dataset/cc3m_images}"
SAMPLES_JSON="${SAMPLES_JSON:-mechanistic_interp/toilet_bathroom/samples.json}"
BASE_PROMPT="${BASE_PROMPT:-Describe this image.}"
STEER_NAME="${STEER_NAME:-toilet}"
READOUT_NAME="${READOUT_NAME:-bathroom}"
BASE_MENTIONS="${BASE_MENTIONS:-any}"
QUESTION="${QUESTION:-Describe this image.}"
FORCED_TEXT="${FORCED_TEXT:-}"
NUM_IMAGES="${NUM_IMAGES:-0}"
ALPHAS="${ALPHAS:-0.05 0.1 0.2 0.4 0.8}"
HOOK_TYPE="${HOOK_TYPE:-post}"
MAX_NEW_TOKENS="${MAX_NEW_TOKENS:-64}"
MAX_SEQ_TOKENS="${MAX_SEQ_TOKENS:-64}"
OUT="${OUT:-mechanistic_interp/graph/gradient_ascent_toilet2bath.png}"
OUT_JSON="${OUT_JSON:-mechanistic_interp/graph/gradient_ascent_toilet2bath.json}"

REPO_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
cd "${REPO_ROOT}"
export PYTHONPATH="$(cd .. && pwd):$(pwd):${PYTHONPATH:-}"

ARGS=(
    --device_id      "${DEVICE_ID}"
    --dtype          "${DTYPE}"
    --bath_probe     "${BATH_PROBE}"
    --toilet_probe   "${TOILET_PROBE}"
    --image_folder   "${IMAGE_FOLDER}"
    --samples_json   "${SAMPLES_JSON}"
    --base_prompt    "${BASE_PROMPT}"
    --base_mentions  "${BASE_MENTIONS}"
    --question       "${QUESTION}"
    --num_images     "${NUM_IMAGES}"
    --alphas         ${ALPHAS}
    --hook_type      "${HOOK_TYPE}"
    --max_new_tokens "${MAX_NEW_TOKENS}"
    --max_seq_tokens "${MAX_SEQ_TOKENS}"
    --steer_name     "${STEER_NAME}"
    --readout_name   "${READOUT_NAME}"
    --out            "${OUT}"
    --out_json       "${OUT_JSON}"
)
if [ -n "${FORCED_TEXT}" ]; then
    ARGS+=(--forced_text "${FORCED_TEXT}")
fi

python -m mechanistic_interp.gradient_ascent_reverse "${ARGS[@]}" "$@"