File size: 4,090 Bytes
a2ffd07 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 | #!/bin/bash
# =============================================================================
# random_gradient_ascent.sh — RANDOM-direction control for gradient_ascent.
#
# For each BATHROOM-ONLY image: steer the residual at layer l along a RANDOM unit
# direction (not the bathroom-probe gradient), step alpha*‖h_l‖*ĝ (or alpha*ĝ with
# HNORM=0), and measure Δ toilet score at every layer l'>=l. Averaged → heatmap.
# Null model for the bathroom→toilet effect: shows what any perturbation of
# comparable magnitude does.
#
# Tunables (override inline, e.g. NUM_IMAGES=100 HNORM=0 bash <script>):
# DEVICE_ID GPU id (default 0)
# DTYPE bfloat16 | float16 | float32 (default bfloat16)
# VARIANT base | lora | nullu | efuf (default base)
# TOILET_PROBE toilet SequenceLayerProbes (readout)
# IMAGE_FOLDER folder of images
# SAMPLES_JSON samples.json with base_mentions_object flags
# BASE_MENTIONS false | true | any (default false)
# NUM_IMAGES cap on images (0=ALL) (default 100)
# SEED RNG seed (image subsample + random dirs) (default 0)
# ALPHAS steps as FRACTION of ‖h_l‖ (or absolute if HNORM=0) (default "0.05 0.1 0.2 0.4 0.8")
# HNORM 1 = scale step by ‖h_l‖ (default); 0 = --no-hnorm (absolute step)
# HOOK_TYPE pre | mid | post (default post)
# PLOT_MODE normal | meannorm (default meannorm)
# FORCED_TEXT forced ASSISTANT answer (empty = generated caption)
# 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}"
VARIANT="${VARIANT:-base}"
TOILET_PROBE="${TOILET_PROBE:-/data/caotue/latent_probes/seqprobes_4variant_toilet/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="${BASE_PROMPT:-Describe this image.}"
BASE_MENTIONS="${BASE_MENTIONS:-false}"
QUESTION="${QUESTION:-Describe this image.}"
FORCED_TEXT="${FORCED_TEXT:-This image features a bathroom with a}"
NUM_IMAGES="${NUM_IMAGES:-100}"
SEED="${SEED:-0}"
ALPHAS="${ALPHAS:-0.05 0.1 0.2 0.4 0.8}"
HNORM="${HNORM:-1}"
HOOK_TYPE="${HOOK_TYPE:-post}"
PLOT_MODE="${PLOT_MODE:-meannorm}"
MAX_NEW_TOKENS="${MAX_NEW_TOKENS:-64}"
MAX_SEQ_TOKENS="${MAX_SEQ_TOKENS:-64}"
OUT="${OUT:-mechanistic_interp/graph/random_gradient_ascent_bath2toilet.png}"
OUT_JSON="${OUT_JSON:-mechanistic_interp/graph/random_gradient_ascent_bath2toilet.json}"
REPO_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
cd "${REPO_ROOT}"
export PYTHONPATH="$(cd .. && pwd):$(pwd):${PYTHONPATH:-}"
ARGS=(
--device_id "${DEVICE_ID}"
--dtype "${DTYPE}"
--variant "${VARIANT}"
--toilet_probe "${TOILET_PROBE}"
--image_folder "${IMAGE_FOLDER}"
--samples_json "${SAMPLES_JSON}"
--base_prompt "${BASE_PROMPT}"
--base_mentions "${BASE_MENTIONS}"
--question "${QUESTION}"
--num_images "${NUM_IMAGES}"
--seed "${SEED}"
--alphas ${ALPHAS}
--hook_type "${HOOK_TYPE}"
--plot_mode "${PLOT_MODE}"
--max_new_tokens "${MAX_NEW_TOKENS}"
--max_seq_tokens "${MAX_SEQ_TOKENS}"
--out "${OUT}"
--out_json "${OUT_JSON}"
)
# ‖h‖ scaling toggle: HNORM=0 → --no-hnorm (absolute step).
if [ "${HNORM}" = "0" ] || [ "${HNORM}" = "false" ]; then
ARGS+=(--no-hnorm)
fi
# Only pass --forced_text when non-empty (empty = use generated captions).
if [ -n "${FORCED_TEXT}" ]; then
ARGS+=(--forced_text "${FORCED_TEXT}")
fi
python -m mechanistic_interp.random_gradient_ascent "${ARGS[@]}" "$@"
|