default HF jobs to amp off

3caced1 verified 5 months ago

5 kB

#!/bin/bash
set -euo pipefail

PRESET="${1:?usage: $0 <preset>}"
REPO_URL="${SOURCE_REPO_URL:-https://huggingface.co/vedatonuryilmaz/deepgenopix}"
REPO_BRANCH="${REPO_BRANCH:-}"
DATA_DATASET="${DATA_DATASET:-vedatonuryilmaz/te-seqdata-v1}"
ARTIFACT_REPO="${ARTIFACT_REPO:?must be set}"

if [ "${DEEPGENOPIX_ALREADY_CLONED:-0}" = "1" ]; then
  echo "using existing DeepGenopix checkout at $(pwd)"
else
  if [ -n "$REPO_BRANCH" ]; then
    git clone --branch "$REPO_BRANCH" --depth 1 "$REPO_URL" deepgenopix
  else
    git clone --depth 1 "$REPO_URL" deepgenopix
  fi
  cd deepgenopix
fi
uv sync --extra hf

if [ -z "${DATA_REVISION:-}" ]; then
  DATA_REVISION="$(uv run python -c "from huggingface_hub import HfApi; print(HfApi().dataset_info('$DATA_DATASET').sha)")"
  echo "WARNING: DATA_REVISION was not set; using mutable dataset HEAD $DATA_REVISION" >&2
  echo "WARNING: production HF jobs should pin DATA_REVISION to the immutable split-input commit" >&2
fi
DATA_REVISION_SHORT="${DATA_REVISION:0:12}"
echo "using DATA_DATASET=$DATA_DATASET revision=$DATA_REVISION"

mkdir -p data/raw/te_split
uv run python -c "
from huggingface_hub import snapshot_download
snapshot_download(
    repo_id='$DATA_DATASET',
    repo_type='dataset',
    revision='$DATA_REVISION',
    local_dir='data/raw/te_split',
    allow_patterns=['train/*.parquet', 'val/*.parquet', 'test/*.parquet', 'split_summary.json'],
)
"

if [ "${RESUME:-0}" = "1" ]; then
  uv run python scripts/hf_artifacts.py pull --repo "$ARTIFACT_REPO" --preset "$PRESET" --dest data/output
fi

export DEEPGENOPIX_HF_SYNC=1
export DEEPGENOPIX_HF_ARTIFACT_REPO="$ARTIFACT_REPO"
export DEEPGENOPIX_NUM_WORKERS=4
export DEEPGENOPIX_USE_AMP="${DEEPGENOPIX_USE_AMP:-0}"
echo "using DEEPGENOPIX_USE_AMP=$DEEPGENOPIX_USE_AMP"

MODE=$([ "${RESUME:-0}" = "1" ] && echo resume || echo train)

# Local processed directory signature used by the trainer.
LOCAL_PROCESSED_SIGNATURE="$(uv run python -c "
from deepgenopix.notebook_support import build_experiment_config
cfg = build_experiment_config('$PRESET')
parts = [
    f'px{cfg.bp_per_pixel}',
    f'ps{cfg.pixel_stride_bp}',
    f'fl{cfg.flank_up_bp}-{cfg.flank_down_bp}',
    f'vf{str(cfg.val_frac).replace(\".\", \"p\")}',
    f'tf{str(cfg.test_frac).replace(\".\", \"p\")}',
    f'seed{cfg.split_seed}',
    f'mf{cfg.min_family_size}',
]
print('_'.join(parts))
")"

# HF dataset cache key. DATA_REVISION must mean the immutable split-input
# revision, not the mutable dataset repo HEAD after LMDB/cache uploads.
PROCESSED_SIGNATURE="${LOCAL_PROCESSED_SIGNATURE}_ds${DATA_REVISION_SHORT}"
PIXEL_STRIDE="$(uv run python -c "from deepgenopix.notebook_support import build_experiment_config; print(build_experiment_config('$PRESET').pixel_stride_bp)")"
PROCESSED_DIR="data/processed/te_visuals/$LOCAL_PROCESSED_SIGNATURE"
mkdir -p "$PROCESSED_DIR"

# Try to pull an existing ETL cache (LMDB + registry + classes) by signature.
# Cache hit ⇒ skip `deepgenopix etl` entirely; cache miss ⇒ materialize and
# push the result so the next job hits the cache.
if [ "${FORCE_ETL_REBUILD:-0}" = "1" ]; then
  echo "FORCE_ETL_REBUILD=1; skipping LMDB cache pull for $PROCESSED_SIGNATURE"
  rm -f "$PROCESSED_DIR/tensors.lmdb" "$PROCESSED_DIR/tensors.lmdb-lock" "$PROCESSED_DIR/registry.csv" "$PROCESSED_DIR/classes.json"
else
  uv run python scripts/hf_artifacts.py pull-etl \
    --repo "$DATA_DATASET" \
    --signature "$PROCESSED_SIGNATURE" \
    --processed-root "$PROCESSED_DIR" || true
fi

if [ -f "$PROCESSED_DIR/tensors.lmdb" ] && [ -f "$PROCESSED_DIR/registry.csv" ] && [ -f "$PROCESSED_DIR/classes.json" ] && [ -f "$PROCESSED_DIR/etl_cache_manifest.json" ]; then
  echo "etl cache hit for $PROCESSED_SIGNATURE"
else
  echo "etl cache miss for $PROCESSED_SIGNATURE — materializing"
  uv run deepgenopix etl --split-root data/raw/te_split --output-dir "$PROCESSED_DIR" --pixel-stride-bp "$PIXEL_STRIDE"
  uv run python -c "
import json
from pathlib import Path
manifest = {
    'data_dataset': '$DATA_DATASET',
    'data_revision': '$DATA_REVISION',
    'data_revision_short': '$DATA_REVISION_SHORT',
    'processed_signature': '$PROCESSED_SIGNATURE',
    'preset': '$PRESET',
    'pixel_stride_bp': int('$PIXEL_STRIDE'),
    'split_files': [
        'train/te_seqdata.parquet',
        'val/te_seqdata.parquet',
        'test/te_seqdata.parquet',
        'split_summary.json',
    ],
    'source_policy': 'split parquet only; root te_seqdata.parquet is intentionally ignored',
}
Path('$PROCESSED_DIR/etl_cache_manifest.json').write_text(json.dumps(manifest, indent=2))
"
  uv run python scripts/hf_artifacts.py push-etl \
    --repo "$DATA_DATASET" \
    --signature "$PROCESSED_SIGNATURE" \
    --processed-root "$PROCESSED_DIR" \
    --message "etl cache from preset $PRESET" || true
fi

uv run deepgenopix train --preset "$PRESET" --run-mode "$MODE" --no-etl --no-export-embeddings --json
uv run python scripts/hf_artifacts.py push --repo "$ARTIFACT_REPO" --preset "$PRESET" --source data/output

	#!/bin/bash
	set -euo pipefail

	PRESET="${1:?usage: $0 <preset>}"
	REPO_URL="${SOURCE_REPO_URL:-https://huggingface.co/vedatonuryilmaz/deepgenopix}"
	REPO_BRANCH="${REPO_BRANCH:-}"
	DATA_DATASET="${DATA_DATASET:-vedatonuryilmaz/te-seqdata-v1}"
	ARTIFACT_REPO="${ARTIFACT_REPO:?must be set}"

	if [ "${DEEPGENOPIX_ALREADY_CLONED:-0}" = "1" ]; then
	echo "using existing DeepGenopix checkout at $(pwd)"
	else
	if [ -n "$REPO_BRANCH" ]; then
	git clone --branch "$REPO_BRANCH" --depth 1 "$REPO_URL" deepgenopix
	else
	git clone --depth 1 "$REPO_URL" deepgenopix
	fi
	cd deepgenopix
	fi
	uv sync --extra hf

	if [ -z "${DATA_REVISION:-}" ]; then
	DATA_REVISION="$(uv run python -c "from huggingface_hub import HfApi; print(HfApi().dataset_info('$DATA_DATASET').sha)")"
	echo "WARNING: DATA_REVISION was not set; using mutable dataset HEAD $DATA_REVISION" >&2
	echo "WARNING: production HF jobs should pin DATA_REVISION to the immutable split-input commit" >&2
	fi
	DATA_REVISION_SHORT="${DATA_REVISION:0:12}"
	echo "using DATA_DATASET=$DATA_DATASET revision=$DATA_REVISION"

	mkdir -p data/raw/te_split
	uv run python -c "
	from huggingface_hub import snapshot_download
	snapshot_download(
	repo_id='$DATA_DATASET',
	repo_type='dataset',
	revision='$DATA_REVISION',
	local_dir='data/raw/te_split',
	allow_patterns=['train/.parquet', 'val/.parquet', 'test/*.parquet', 'split_summary.json'],
	)
	"

	if [ "${RESUME:-0}" = "1" ]; then
	uv run python scripts/hf_artifacts.py pull --repo "$ARTIFACT_REPO" --preset "$PRESET" --dest data/output
	fi

	export DEEPGENOPIX_HF_SYNC=1
	export DEEPGENOPIX_HF_ARTIFACT_REPO="$ARTIFACT_REPO"
	export DEEPGENOPIX_NUM_WORKERS=4
	export DEEPGENOPIX_USE_AMP="${DEEPGENOPIX_USE_AMP:-0}"
	echo "using DEEPGENOPIX_USE_AMP=$DEEPGENOPIX_USE_AMP"

	MODE=$([ "${RESUME:-0}" = "1" ] && echo resume \|\| echo train)

	# Local processed directory signature used by the trainer.
	LOCAL_PROCESSED_SIGNATURE="$(uv run python -c "
	from deepgenopix.notebook_support import build_experiment_config
	cfg = build_experiment_config('$PRESET')
	parts = [
	f'px{cfg.bp_per_pixel}',
	f'ps{cfg.pixel_stride_bp}',
	f'fl{cfg.flank_up_bp}-{cfg.flank_down_bp}',
	f'vf{str(cfg.val_frac).replace(\".\", \"p\")}',
	f'tf{str(cfg.test_frac).replace(\".\", \"p\")}',
	f'seed{cfg.split_seed}',
	f'mf{cfg.min_family_size}',
	]
	print('_'.join(parts))
	")"

	# HF dataset cache key. DATA_REVISION must mean the immutable split-input
	# revision, not the mutable dataset repo HEAD after LMDB/cache uploads.
	PROCESSED_SIGNATURE="${LOCAL_PROCESSED_SIGNATURE}_ds${DATA_REVISION_SHORT}"
	PIXEL_STRIDE="$(uv run python -c "from deepgenopix.notebook_support import build_experiment_config; print(build_experiment_config('$PRESET').pixel_stride_bp)")"
	PROCESSED_DIR="data/processed/te_visuals/$LOCAL_PROCESSED_SIGNATURE"
	mkdir -p "$PROCESSED_DIR"

	# Try to pull an existing ETL cache (LMDB + registry + classes) by signature.
	# Cache hit ⇒ skip `deepgenopix etl` entirely; cache miss ⇒ materialize and
	# push the result so the next job hits the cache.
	if [ "${FORCE_ETL_REBUILD:-0}" = "1" ]; then
	echo "FORCE_ETL_REBUILD=1; skipping LMDB cache pull for $PROCESSED_SIGNATURE"
	rm -f "$PROCESSED_DIR/tensors.lmdb" "$PROCESSED_DIR/tensors.lmdb-lock" "$PROCESSED_DIR/registry.csv" "$PROCESSED_DIR/classes.json"
	else
	uv run python scripts/hf_artifacts.py pull-etl \
	--repo "$DATA_DATASET" \
	--signature "$PROCESSED_SIGNATURE" \
	--processed-root "$PROCESSED_DIR" \|\| true
	fi

	if [ -f "$PROCESSED_DIR/tensors.lmdb" ] && [ -f "$PROCESSED_DIR/registry.csv" ] && [ -f "$PROCESSED_DIR/classes.json" ] && [ -f "$PROCESSED_DIR/etl_cache_manifest.json" ]; then
	echo "etl cache hit for $PROCESSED_SIGNATURE"
	else
	echo "etl cache miss for $PROCESSED_SIGNATURE — materializing"
	uv run deepgenopix etl --split-root data/raw/te_split --output-dir "$PROCESSED_DIR" --pixel-stride-bp "$PIXEL_STRIDE"
	uv run python -c "
	import json
	from pathlib import Path
	manifest = {
	'data_dataset': '$DATA_DATASET',
	'data_revision': '$DATA_REVISION',
	'data_revision_short': '$DATA_REVISION_SHORT',
	'processed_signature': '$PROCESSED_SIGNATURE',
	'preset': '$PRESET',
	'pixel_stride_bp': int('$PIXEL_STRIDE'),
	'split_files': [
	'train/te_seqdata.parquet',
	'val/te_seqdata.parquet',
	'test/te_seqdata.parquet',
	'split_summary.json',
	],
	'source_policy': 'split parquet only; root te_seqdata.parquet is intentionally ignored',
	}
	Path('$PROCESSED_DIR/etl_cache_manifest.json').write_text(json.dumps(manifest, indent=2))
	"
	uv run python scripts/hf_artifacts.py push-etl \
	--repo "$DATA_DATASET" \
	--signature "$PROCESSED_SIGNATURE" \
	--processed-root "$PROCESSED_DIR" \
	--message "etl cache from preset $PRESET" \|\| true
	fi

	uv run deepgenopix train --preset "$PRESET" --run-mode "$MODE" --no-etl --no-export-embeddings --json
	uv run python scripts/hf_artifacts.py push --repo "$ARTIFACT_REPO" --preset "$PRESET" --source data/output