evaluate.sh

#!/usr/bin/env bash
# Evaluate a trained checkpoint across every sparse shell / gradient-count setting.
#
# `experiment.json` is written next to the checkpoint during training and is
# used to rebuild the exact model architecture, so no architecture flags need
# to be repeated here.

set -euo pipefail

# Training writes to <checkpoint_path>/spatialAttLiteStrong-mask-wb/<experiment_name>/,
# where the best checkpoint is named
#   checkpoints/best_checkpoint_epoch_<epoch>_MSE_<val_metric>.pth
run_dir="./checkpoints/spatialAttLiteStrong-mask-wb/<experiment_name>"

checkpoint="$(ls -t "$run_dir"/checkpoints/best_checkpoint_epoch_*.pth 2>/dev/null | head -1)"
experiment_config="$run_dir/experiment.json"
data_path="./data/HCP100_compact"
output="./results/sparse_test.json"

if [[ -z "$checkpoint" ]]; then
    echo "No best checkpoint found under $run_dir/checkpoints/" >&2
    echo "Set run_dir at the top of this script to your experiment directory." >&2
    exit 1
fi

python evaluation/evaluate_sparse.py \
    --checkpoint "$checkpoint" \
    --experiment_config "$experiment_config" \
    --data_path "$data_path" \
    --output "$output" \
    --device cuda \
    --batch_size 512 \
    --num_workers 8 \
    --amp 1 \
    "$@"