scripts/train_cta_ablation.sh

#!/usr/bin/env bash
# Train a single CTA ablation variant.
#
# Usage:
#   bash scripts/train_cta_ablation.sh <VARIANT> [extra hydra overrides...]
#
# Examples:
#   bash scripts/train_cta_ablation.sh A1_full
#   bash scripts/train_cta_ablation.sh A2_no_asym
#   bash scripts/train_cta_ablation.sh B2_all_samples
#   bash scripts/train_cta_ablation.sh D4_mlp_predictor
#
# The output directory will be:
#   outputs/cta_ablation_<VARIANT>_<timestamp>/
set -euo pipefail
cd "$(dirname "$0")/.."

# load .env if present
[ -f .env ] && set -a && . ./.env && set +a

# NCCL optimization settings to prevent timeout errors
export NCCL_TIMEOUT=3600                    # 1 hour timeout (default 30min)
export NCCL_ASYNC_ERROR_HANDLING=1         # Enable async error handling
export NCCL_MAX_NCHANNELS=4                # Reduce number of channels for better stability
export NCCL_MIN_NCHANNELS=4                # Set minimum channels
export NCCL_BUFFSIZE=4194304               # 4MB buffer size

# Additional optimization settings
export CUDA_LAUNCH_BLOCKING=0              # Non-blocking CUDA operations
export NCCL_P2P_DISABLE=0                  # Enable peer-to-peer communication

VARIANT="${1:-A1_full}"
shift || true

echo "[ablation] running variant: ${VARIANT}"
echo "[ablation] NCCL settings: TIMEOUT=${NCCL_TIMEOUT}s, ASYNC_ERROR_HANDLING=${NCCL_ASYNC_ERROR_HANDLING}"

python3 src/train.py \
    method=cta_ablation \
    data=fairtalking \
    trainer=ddp \
    backbone=timesformer \
    method.ablation_variant="${VARIANT}" \
    experiment_name="cta_ablation_${VARIANT}" \
    "$@"