AgentSkillsCN

retrain

采用正确的TensorFlow导入顺序，对ML交易模型（XGBoost、LSTM、CNN、Transformer，或全部）进行再训练。

SKILL.md
--- frontmatter
name: retrain
description: Retrain ML trading models (xgboost, lstm, cnn, transformer, or all) with proper TF import ordering
disable-model-invocation: true
argument-hint: "[model: all|xgboost|lstm|cnn|transformer] [options: tune, deploy]"
allowed-tools: Bash(python *), Bash(.venv/bin/python *), Read, Grep, Glob

Model Retraining

Retrain ML models using scripts/retrain_models.py. Parse the arguments from $ARGUMENTS.

Argument Mapping

User saysCLI flags
all (or no model specified)--models all
xgboost or xgb--models xgboost
lstm--models lstm
cnn--models cnn
transformer--models transformer
tune--tune
deploy--deploy
force-deploy--force-deploy
epochs=N--epochs N
trials=N--tune-trials N
status--status
versions--list-versions

For individual model tuning: tune-xgboost, tune-lstm, tune-cnn, tune-transformer map to their respective flags.

Steps

  1. •Parse arguments and build the CLI command
  2. •Show the user the exact command being run
  3. •CRITICAL: Run with .venv/bin/python only (system Python 3.9.6 causes TF mutex deadlocks)
  4. •Run: .venv/bin/python scripts/retrain_models.py <flags>
    • •Timeout: 600 seconds for single model, 1800 seconds for all models or tuning
  5. •After completion, summarize:
    • •Training accuracy and test accuracy for each model
    • •Whether the model improved over production
    • •Whether deployment happened (if --deploy was used)
  6. •If the user asked to retrain the transformer specifically, remind them that it currently has 49.8% accuracy and its ensemble weight is 0.3 (lowered to avoid diluting signal)

Important Notes

  • •TF import order bug: pandas before TF causes model.fit() deadlock on macOS. The scripts already handle this, but always use .venv/bin/python.
  • •XGBoost requires libomp (already installed via Homebrew)
  • •LSTM/CNN training uses MPS GPU acceleration on Apple M4
  • •Transformer is undertrained (49.8% accuracy) — needs more data/tuning before increasing its ensemble weight
  • •Default deployment threshold is 1% accuracy improvement