mirror of
https://github.com/RYDE-WORK/lnp_ml.git
synced 2026-03-21 09:36:32 +08:00
285 lines
11 KiB
Makefile
285 lines
11 KiB
Makefile
#################################################################################
|
||
# GLOBALS #
|
||
#################################################################################
|
||
|
||
PROJECT_NAME = lnp-ml
|
||
PYTHON_VERSION = 3.8
|
||
PYTHON_INTERPRETER = python
|
||
|
||
# --- CLI flag 变量 ---
|
||
MPNN_FLAG = $(if $(USE_MPNN),--use-mpnn,)
|
||
FREEZE_FLAG = $(if $(FREEZE_BACKBONE),--freeze-backbone,)
|
||
DEVICE_FLAG = $(if $(DEVICE),--device $(DEVICE),)
|
||
SCAFFOLD_SPLIT_FLAG = $(if $(filter 1,$(SCAFFOLD_SPLIT)),--scaffold-split,)
|
||
SEED_FLAG = $(if $(SEED),--seed $(SEED),)
|
||
N_TRIALS_FLAG = $(if $(N_TRIALS),--n-trials $(N_TRIALS),)
|
||
EPOCHS_PER_TRIAL_FLAG = $(if $(EPOCHS_PER_TRIAL),--epochs-per-trial $(EPOCHS_PER_TRIAL),)
|
||
MIN_STRATUM_FLAG = $(if $(MIN_STRATUM_COUNT),--min-stratum-count $(MIN_STRATUM_COUNT),)
|
||
OUTPUT_DIR_FLAG = $(if $(OUTPUT_DIR),--output-dir $(OUTPUT_DIR),)
|
||
USE_SWA_FLAG = $(if $(USE_SWA),--use-swa,)
|
||
INIT_PRETRAIN_FLAG = $(if $(NO_PRETRAIN),,--init-from-pretrain $(or $(INIT_PRETRAIN),models/pretrain_delivery.pt))
|
||
|
||
#################################################################################
|
||
# ENVIRONMENT & CODE QUALITY #
|
||
#################################################################################
|
||
|
||
## Install Python dependencies
|
||
.PHONY: requirements
|
||
requirements:
|
||
pixi install
|
||
|
||
## Set up Python interpreter environment
|
||
.PHONY: create_environment
|
||
create_environment:
|
||
@echo ">>> Pixi environment will be created when running 'make requirements'"
|
||
@echo ">>> Activate with:\npixi shell"
|
||
|
||
## Delete all compiled Python files
|
||
.PHONY: clean
|
||
clean:
|
||
find . -type f -name "*.py[co]" -delete
|
||
find . -type d -name "__pycache__" -delete
|
||
|
||
## Lint using ruff (use `make format` to do formatting)
|
||
.PHONY: lint
|
||
lint:
|
||
ruff format --check
|
||
ruff check
|
||
|
||
## Format source code with ruff
|
||
.PHONY: format
|
||
format:
|
||
ruff check --fix
|
||
ruff format
|
||
|
||
#################################################################################
|
||
# DATA PROCESSING #
|
||
#################################################################################
|
||
|
||
## Preprocess internal data (raw -> interim)
|
||
.PHONY: preprocess
|
||
preprocess: requirements
|
||
$(PYTHON_INTERPRETER) scripts/preprocess_internal.py
|
||
|
||
## Process dataset (interim -> processed)
|
||
.PHONY: data
|
||
data: requirements
|
||
$(PYTHON_INTERPRETER) scripts/process_data.py
|
||
|
||
## Process dataset for final training (interim -> processed/final, train:val=9:1, no test)
|
||
.PHONY: data_final
|
||
data_final: requirements
|
||
$(PYTHON_INTERPRETER) scripts/process_data_final.py
|
||
|
||
## Process external data for pretrain (external -> processed)
|
||
.PHONY: data_pretrain
|
||
data_pretrain: requirements
|
||
$(PYTHON_INTERPRETER) scripts/process_external.py
|
||
|
||
## Process CV data for cross-validation pretrain (external/all_amine_split_for_LiON -> processed/cv)
|
||
.PHONY: data_pretrain_cv
|
||
data_pretrain_cv: requirements
|
||
$(PYTHON_INTERPRETER) scripts/process_external_cv.py
|
||
|
||
## Process internal data with CV splitting (interim -> processed/cv)
|
||
## Use SCAFFOLD_SPLIT=1 to enable amine-based scaffold splitting (default: random shuffle)
|
||
.PHONY: data_cv
|
||
data_cv: requirements
|
||
$(PYTHON_INTERPRETER) scripts/process_data_cv.py $(SCAFFOLD_SPLIT_FLAG)
|
||
|
||
#################################################################################
|
||
# TRAINING #
|
||
#################################################################################
|
||
|
||
## Pretrain on external data (delivery only)
|
||
.PHONY: pretrain
|
||
pretrain: requirements
|
||
$(PYTHON_INTERPRETER) -m lnp_ml.modeling.pretrain main $(MPNN_FLAG) $(DEVICE_FLAG)
|
||
|
||
## Pretrain with cross-validation (5-fold)
|
||
.PHONY: pretrain_cv
|
||
pretrain_cv: requirements
|
||
$(PYTHON_INTERPRETER) -m lnp_ml.modeling.pretrain_cv main $(MPNN_FLAG) $(DEVICE_FLAG)
|
||
|
||
## Train model (multi-task, from scratch)
|
||
.PHONY: train
|
||
train: requirements
|
||
$(PYTHON_INTERPRETER) -m lnp_ml.modeling.train $(MPNN_FLAG) $(DEVICE_FLAG)
|
||
|
||
## Finetune from pretrained checkpoint (use FREEZE_BACKBONE=1 to freeze backbone)
|
||
.PHONY: finetune
|
||
finetune: requirements
|
||
$(PYTHON_INTERPRETER) -m lnp_ml.modeling.train --init-from-pretrain models/pretrain_delivery.pt $(FREEZE_FLAG) $(MPNN_FLAG) $(DEVICE_FLAG)
|
||
|
||
## Final training using all data (train:val=9:1, no test set), with pretrained weights
|
||
.PHONY: train_final
|
||
train_final: requirements
|
||
$(PYTHON_INTERPRETER) -m lnp_ml.modeling.train \
|
||
--train-path data/processed/final/train.parquet \
|
||
--val-path data/processed/final/val.parquet \
|
||
--output-dir models/final \
|
||
--init-from-pretrain models/pretrain_delivery.pt \
|
||
$(FREEZE_FLAG) $(MPNN_FLAG) $(DEVICE_FLAG)
|
||
|
||
## Train with cross-validation on internal data only (5-fold, amine-based split)
|
||
.PHONY: train_cv
|
||
train_cv: requirements
|
||
$(PYTHON_INTERPRETER) -m lnp_ml.modeling.train_cv main $(FREEZE_FLAG) $(MPNN_FLAG) $(DEVICE_FLAG)
|
||
|
||
## Finetune with cross-validation on internal data (5-fold) with pretrained weights
|
||
.PHONY: finetune_cv
|
||
finetune_cv: requirements
|
||
$(PYTHON_INTERPRETER) -m lnp_ml.modeling.train_cv main --init-from-pretrain models/pretrain_delivery.pt $(FREEZE_FLAG) $(MPNN_FLAG) $(DEVICE_FLAG)
|
||
|
||
#################################################################################
|
||
# EVALUATION #
|
||
#################################################################################
|
||
|
||
## Evaluate pretrain model (delivery metrics)
|
||
.PHONY: test_pretrain
|
||
test_pretrain: requirements
|
||
$(PYTHON_INTERPRETER) -m lnp_ml.modeling.pretrain test $(MPNN_FLAG) $(DEVICE_FLAG)
|
||
|
||
## Evaluate CV pretrain models on test sets (auto-detects MPNN from checkpoint)
|
||
.PHONY: test_pretrain_cv
|
||
test_pretrain_cv: requirements
|
||
$(PYTHON_INTERPRETER) -m lnp_ml.modeling.pretrain_cv test $(DEVICE_FLAG)
|
||
|
||
## Evaluate CV finetuned models on test sets (auto-detects MPNN from checkpoint)
|
||
.PHONY: test_cv
|
||
test_cv: requirements
|
||
$(PYTHON_INTERPRETER) -m lnp_ml.modeling.train_cv test $(DEVICE_FLAG)
|
||
|
||
## Test model on test set (with detailed metrics, auto-detects MPNN from checkpoint)
|
||
.PHONY: test
|
||
test: requirements
|
||
$(PYTHON_INTERPRETER) -m lnp_ml.modeling.predict test $(DEVICE_FLAG)
|
||
|
||
## Run predictions
|
||
.PHONY: predict
|
||
predict: requirements
|
||
$(PYTHON_INTERPRETER) -m lnp_ml.modeling.predict $(DEVICE_FLAG)
|
||
|
||
#################################################################################
|
||
# HYPERPARAMETER TUNING #
|
||
#################################################################################
|
||
# 通用参数：
|
||
# SEED 随机种子 (默认: 42)
|
||
# N_TRIALS Optuna 试验数 (默认: 20)
|
||
# EPOCHS_PER_TRIAL 每个试验的最大 epoch (默认: 30)
|
||
# MIN_STRATUM_COUNT 复合分层标签的最小样本数 (默认: 5)
|
||
# OUTPUT_DIR 输出目录 (根据命令有不同默认值)
|
||
# INIT_PRETRAIN 预训练权重路径 (默认: models/pretrain_delivery.pt)
|
||
# NO_PRETRAIN=1 禁用预训练权重
|
||
|
||
## Train with hyperparameter tuning
|
||
.PHONY: tune
|
||
tune: requirements
|
||
$(PYTHON_INTERPRETER) -m lnp_ml.modeling.train --tune $(MPNN_FLAG) $(DEVICE_FLAG)
|
||
|
||
## Nested CV with Optuna: outer 5-fold (test) + inner 3-fold (tune)
|
||
## 用于模型评估：外层 5-fold 产生无偏性能估计，内层 3-fold 做超参搜索
|
||
## 使用示例: make nested_cv_tune DEVICE=cuda N_TRIALS=30
|
||
.PHONY: nested_cv_tune
|
||
nested_cv_tune: requirements
|
||
$(PYTHON_INTERPRETER) -m lnp_ml.modeling.nested_cv_optuna \
|
||
$(DEVICE_FLAG) $(MPNN_FLAG) $(SEED_FLAG) $(INIT_PRETRAIN_FLAG) \
|
||
$(N_TRIALS_FLAG) $(EPOCHS_PER_TRIAL_FLAG) $(MIN_STRATUM_FLAG) $(OUTPUT_DIR_FLAG)
|
||
|
||
## Final training with Optuna: 3-fold CV tune + full data train
|
||
## 用于最终模型训练：3-fold 调参后用全量数据训练（无 early-stop）
|
||
## 使用示例: make final_optuna DEVICE=cuda N_TRIALS=30 USE_SWA=1
|
||
.PHONY: final_optuna
|
||
final_optuna: requirements
|
||
$(PYTHON_INTERPRETER) -m lnp_ml.modeling.final_train_optuna_cv \
|
||
$(DEVICE_FLAG) $(MPNN_FLAG) $(SEED_FLAG) $(INIT_PRETRAIN_FLAG) \
|
||
$(N_TRIALS_FLAG) $(EPOCHS_PER_TRIAL_FLAG) $(MIN_STRATUM_FLAG) $(OUTPUT_DIR_FLAG) $(USE_SWA_FLAG)
|
||
|
||
#################################################################################
|
||
# SERVING & DEPLOYMENT #
|
||
#################################################################################
|
||
|
||
## Formulation optimization: find optimal LNP formulation for target organ
|
||
## Usage: make optimize SMILES="CC(C)..." ORGAN=liver
|
||
.PHONY: optimize
|
||
optimize: requirements
|
||
$(PYTHON_INTERPRETER) -m app.optimize --smiles "$(SMILES)" --organ $(ORGAN) $(DEVICE_FLAG)
|
||
|
||
## Start FastAPI backend server (port 8000)
|
||
.PHONY: api
|
||
api: requirements
|
||
uvicorn app.api:app --host 0.0.0.0 --port 8000 --reload
|
||
|
||
## Start Streamlit frontend app (port 8501)
|
||
.PHONY: webapp
|
||
webapp: requirements
|
||
streamlit run app/app.py --server.port 8501
|
||
|
||
## Start both API and webapp (run in separate terminals)
|
||
.PHONY: serve
|
||
serve:
|
||
@echo "请在两个终端分别运行:"
|
||
@echo " 终端 1: make api"
|
||
@echo " 终端 2: make webapp"
|
||
@echo ""
|
||
@echo "然后访问: http://localhost:8501"
|
||
|
||
#################################################################################
|
||
# DOCKER #
|
||
#################################################################################
|
||
|
||
## Build Docker images
|
||
.PHONY: docker-build
|
||
docker-build:
|
||
docker compose build
|
||
|
||
## Start all services with Docker Compose
|
||
.PHONY: docker-up
|
||
docker-up:
|
||
docker compose up -d
|
||
|
||
## Stop all Docker services
|
||
.PHONY: docker-down
|
||
docker-down:
|
||
docker compose down
|
||
|
||
## View Docker logs
|
||
.PHONY: docker-logs
|
||
docker-logs:
|
||
docker compose logs -f
|
||
|
||
## Build and start all services
|
||
.PHONY: docker-serve
|
||
docker-serve: docker-build docker-up
|
||
@echo ""
|
||
@echo "🚀 服务已启动!"
|
||
@echo " - API: http://localhost:8000"
|
||
@echo " - Web 应用: http://localhost:8501"
|
||
@echo ""
|
||
@echo "查看日志: make docker-logs"
|
||
@echo "停止服务: make docker-down"
|
||
|
||
## Clean Docker resources (images, volumes, etc.)
|
||
.PHONY: docker-clean
|
||
docker-clean:
|
||
docker compose down -v --rmi local
|
||
docker system prune -f
|
||
|
||
#################################################################################
|
||
# Self Documenting Commands #
|
||
#################################################################################
|
||
|
||
.DEFAULT_GOAL := help
|
||
|
||
define PRINT_HELP_PYSCRIPT
|
||
import re, sys; \
|
||
lines = '\n'.join([line for line in sys.stdin]); \
|
||
matches = re.findall(r'\n## (.*)\n[\s\S]+?\n([a-zA-Z_-]+):', lines); \
|
||
print('Available rules:\n'); \
|
||
print('\n'.join(['{:25}{}'.format(*reversed(match)) for match in matches]))
|
||
endef
|
||
export PRINT_HELP_PYSCRIPT
|
||
|
||
help:
|
||
@$(PYTHON_INTERPRETER) -c "${PRINT_HELP_PYSCRIPT}" < $(MAKEFILE_LIST)
|