torch-rechub

English

中文

English

中文

Appearance

Experiment Tracking (Lite) ​

Torch-RecHub offers optional experiment tracking with:

  • Weights & Biases (wandb)
  • SwanLab
  • TensorBoardX

Goal: zero overhead by default. If you want tracking, pass model_logger to a Trainer.

Installation ​

bash
# all tracking deps
pip install torch-rechub[tracking]

# or individually
pip install wandb
pip install swanlab
pip install tensorboardX

Quickstart ​

python
from torch_rechub.trainers import CTRTrainer
from torch_rechub.basic.tracking import WandbLogger

logger = WandbLogger(project="my-ctr", name="exp1")  # optional

trainer = CTRTrainer(
    model,
    optimizer_params={"lr": 1e-3},
    n_epoch=10,
    model_logger=logger,   # pass None for zero tracking
)

trainer.fit(train_dl, val_dl)

Multiple loggers ​

model_logger can be a single instance or a list:

python
tb = TensorBoardXLogger(log_dir="./runs/exp1")
wb = WandbLogger(project="my-ctr", name="exp1")
sb = SwanLabLogger(project="my-ctr")

trainer = CTRTrainer(model, model_logger=[tb, wb])

Set API keys in code ​

python
import os
os.environ['WANDB_API_KEY'] = "your API_KEY"
os.environ['SWANLAB_API_KEY'] = "your API_KEY"

Overhead when disabled ​

When model_logger=None, training runs without any tracking calls—no extra cost.

API Overview ​

python
class BaseLogger:
    log_metrics(metrics: Dict[str, Any], step: Optional[int] = None)
    log_hyperparams(params: Dict[str, Any])
    finish()
  • WandbLogger(project, name=None, config=None, tags=None, notes=None, dir=None, …)
  • SwanLabLogger(project=None, experiment_name=None, description=None, config=None, logdir=None, …)
  • TensorBoardXLogger(log_dir, comment="", …)

Trainer behavior (auto-logging) ​

  • train/loss, learning_rate
  • Validation metrics: val/auc (CTR/Match), val/task_i_score (MTL), val/accuracy (Seq)
  • Metrics are logged only when model_logger is provided

FAQ ​

  1. How to skip tracking?
    Do not pass model_logger (default None).

  2. Disable hyperparameter logging?
    Trainers log hyperparameters once at start; you can override/ignore as needed.

  3. A library is missing?
    Instantiating a logger raises a friendly ImportError; training still works.

  4. Custom metrics?
    Keep the logger instance and call log_metrics yourself.

python
logger = WandbLogger(project="demo")
logger.log_metrics({"custom_metric": 0.9}, step=epoch)

Best Practices ​

  • Local quick debug: TensorBoardX
  • Team/remote: WandB or SwanLab
  • Need local + cloud: pass multiple loggers
  • No logger specified: minimal overhead