1
Fork 0

feat(experiment-logger): add standalone logging framework

Create reusable experiment logging package with:
- UnifiedLogger for multi-backend logging (WandB, disk, stdout)
- Automatic checkpoint and model saving with metadata
- WandB artifact upload support
- Graceful degradation when WandB unavailable
- Comprehensive API documentation

This is a standalone, project-agnostic library that can be reused
across different ML projects.
This commit is contained in:
Tibo De Peuter 2026-03-31 19:49:04 +00:00
parent c27f5fcdf2
commit 3ce107a560
4 changed files with 495 additions and 0 deletions

View file

@ -0,0 +1,145 @@
# Experiment Logger
A lightweight, standalone logging framework for machine learning experiments with multi-backend support.
## Features
- **Multi-backend logging**: Simultaneously log to WandB, local disk (JSON), and stdout
- **Data preservation**: All metrics saved locally, even if WandB is unavailable
- **Checkpoint management**: Save model checkpoints with metadata
- **WandB integration**: Optional artifact upload for model versioning
- **Graceful degradation**: Works without WandB installed
- **Simple API**: Minimal configuration required
## Installation
This package is included in the project. To use it in your code:
```python
from experiment_logger import UnifiedLogger
```
## Quick Start
```python
from experiment_logger import UnifiedLogger
# Initialize logger
logger = UnifiedLogger(
run_name="my_experiment",
config={"learning_rate": 0.001, "batch_size": 32},
project_name="MyProject",
entity="my-wandb-username", # Optional
use_wandb=True, # Set to False to disable WandB
)
# Log metrics
for step in range(100):
logger.log({
"loss": 1.0 / (step + 1),
"accuracy": step * 0.01,
}, step=step)
# Save checkpoint
logger.save_checkpoint(
params=model_params,
step=100,
metadata={"epoch": 1, "val_acc": 0.95},
)
# Save final model
logger.save_final_model(
params=final_params,
metadata={"final_accuracy": 0.98},
)
# Finalize (flushes remaining metrics)
logger.finish()
```
## Context Manager
Use as a context manager for automatic cleanup:
```python
with UnifiedLogger(run_name="my_exp", config={}) as logger:
logger.log({"metric": 1.0})
# Automatically calls finish() on exit
```
## Configuration
### Constructor Parameters
- `run_name` (str): Unique name for this run
- `config` (dict): Configuration dictionary with hyperparameters
- `project_name` (str): WandB project name (default: "PPO-Modularity")
- `entity` (str, optional): WandB entity (team/user name)
- `base_dir` (str): Base directory for local storage (default: "runs")
- `use_wandb` (bool): Enable WandB logging (default: True)
- `save_code` (bool): Save code to WandB (default: True)
### Directory Structure
```
runs/
└── my_experiment/
├── config.json # Saved configuration
├── metrics/
│ └── metrics.jsonl # Line-delimited JSON metrics
├── checkpoints/
│ ├── checkpoint_step_100.flax
│ └── checkpoint_step_100_metadata.json
└── final_model.flax
```
## API Reference
### `log(metrics, step=None, commit=True)`
Log metrics to all backends.
**Parameters:**
- `metrics` (dict): Dictionary of metric name -> value
- `step` (int, optional): Global step counter (auto-incremented if None)
- `commit` (bool): Whether to commit to WandB immediately
### `save_checkpoint(params, step, prefix="checkpoint", metadata=None)`
Save model checkpoint to disk and optionally to WandB.
**Parameters:**
- `params`: Model parameters (Flax params or any serializable object)
- `step` (int): Current training step
- `prefix` (str): Prefix for checkpoint filename
- `metadata` (dict, optional): Additional metadata to save
### `save_final_model(params, metadata=None)`
Save the final trained model.
**Parameters:**
- `params`: Model parameters
- `metadata` (dict, optional): Metadata about the final model
### `finish()`
Finalize logging and cleanup. Flushes remaining metrics to disk.
## Usage in Projects
This logger is designed to be:
- **Project-agnostic**: Use in any ML project, not just this one
- **Framework-agnostic**: Works with JAX, PyTorch, TensorFlow, etc.
- **Minimal dependencies**: Only requires `wandb` (optional), `flax` (for serialization), and `numpy`
## Design Philosophy
1. **Never lose data**: All metrics saved locally, regardless of WandB availability
2. **Simple API**: Minimal boilerplate, easy to integrate
3. **Fail gracefully**: Missing WandB shouldn't break experiments
4. **Reproducibility**: Save full configuration with every run
## License
Part of the 2026SEL3-project-BrittleStar repository.

View file

@ -0,0 +1,11 @@
"""Unified logging framework for machine learning experiments.
This package provides a unified interface for logging to multiple backends
(WandB, disk, stdout) simultaneously, ensuring no data loss.
"""
from experiment_logger.unified_logger import UnifiedLogger
from experiment_logger.wandb_utils import finish_wandb, init_wandb
__all__ = ["UnifiedLogger", "init_wandb", "finish_wandb"]
__version__ = "0.1.0"

View file

@ -0,0 +1,270 @@
"""Unified logger that writes to multiple backends simultaneously.
This logger ensures all experimental data is preserved by writing to:
1. Weights & Biases (when available)
2. Local disk (JSON files, model checkpoints)
3. stdout (for real-time monitoring)
"""
import json
import logging
import time
from pathlib import Path
from typing import Any, Dict, Optional
import flax
import numpy as np
from experiment_logger.wandb_utils import finish_wandb, init_wandb
logger = logging.getLogger(__name__)
class UnifiedLogger:
"""Unified logger for scientific experiments with redundant backup."""
def __init__(
self,
run_name: str,
config: Dict[str, Any],
project_name: str = "PPO-Modularity",
entity: Optional[str] = None,
base_dir: str = "runs",
use_wandb: bool = True,
save_code: bool = True,
):
"""Initialize the unified logger.
Args:
run_name: Unique name for this run
config: Configuration dictionary with hyperparameters
project_name: WandB project name
entity: WandB entity (team/user name)
base_dir: Base directory for local storage
use_wandb: Whether to use WandB logging
save_code: Whether to save code to WandB
"""
self.run_name = run_name
self.config = config
self.use_wandb = use_wandb
self.wandb_available = False
self.wandb_run = None
# Setup local storage
self.run_dir = Path(base_dir) / run_name
self.run_dir.mkdir(parents=True, exist_ok=True)
self.checkpoints_dir = self.run_dir / "checkpoints"
self.checkpoints_dir.mkdir(exist_ok=True)
self.metrics_dir = self.run_dir / "metrics"
self.metrics_dir.mkdir(exist_ok=True)
self.config_file = self.run_dir / "config.json"
# Save config to disk
self._save_config()
# Initialize WandB if requested
if self.use_wandb:
self._init_wandb(project_name, entity, save_code)
# Initialize metrics storage
self.metrics_buffer: list[Dict[str, Any]] = []
self.step_counter = 0
logger.info(f"Initialized for run: {run_name}")
logger.info(f"Local storage: {self.run_dir.absolute()}")
logger.info(f"WandB logging: {self.wandb_available}")
def _init_wandb(self, project_name: str, entity: Optional[str], save_code: bool):
"""Initialize Weights & Biases logging."""
self.wandb_run = init_wandb(
project=project_name,
entity=entity,
name=self.run_name,
config=self.config,
save_code=save_code,
resume="allow",
)
self.wandb_available = self.wandb_run is not None
def _save_config(self):
"""Save configuration to disk."""
try:
with open(self.config_file, "w") as f:
json.dump(self.config, f, indent=2)
logger.info(f"Config saved to {self.config_file}")
except Exception as e:
logger.error(f"Error saving config: {e}")
def log(self, metrics: Dict[str, Any], step: Optional[int] = None, commit: bool = True):
"""Log metrics to all backends.
Args:
metrics: Dictionary of metric name -> value
step: Global step counter (auto-incremented if None)
commit: Whether to commit to WandB immediately
"""
if step is None:
step = self.step_counter
self.step_counter += 1
# Add timestamp
metrics_with_metadata = {
"step": step,
"timestamp": time.time(),
**metrics,
}
# Log to stdout
self._log_to_stdout(metrics_with_metadata)
# Log to WandB
if self.wandb_available:
try:
self.wandb_run.log(metrics, step=step, commit=commit)
except Exception as e:
logger.warning(f"WandB logging failed: {e}")
# Buffer for disk storage
self.metrics_buffer.append(metrics_with_metadata)
# Periodically flush to disk
if len(self.metrics_buffer) >= 100:
self._flush_metrics()
def _log_to_stdout(self, metrics: Dict[str, Any]):
"""Log metrics to stdout for real-time monitoring."""
step = metrics.get("step", "?")
metric_str = ", ".join(
f"{k}={v:.6f}" if isinstance(v, (float, np.floating)) else f"{k}={v}"
for k, v in metrics.items()
if k not in ["step", "timestamp"]
)
logger.info(f"[Step {step}] {metric_str}")
def _flush_metrics(self):
"""Flush buffered metrics to disk."""
if not self.metrics_buffer:
return
try:
metrics_file = self.metrics_dir / "metrics.jsonl"
with open(metrics_file, "a") as f:
for metric in self.metrics_buffer:
f.write(json.dumps(metric) + "\n")
self.metrics_buffer.clear()
except Exception as e:
logger.error(f"Error flushing metrics: {e}")
def save_checkpoint(
self,
params: Any,
step: int,
prefix: str = "checkpoint",
metadata: Optional[Dict[str, Any]] = None,
):
"""Save model checkpoint to disk and optionally to WandB.
Args:
params: Model parameters (Flax params or any serializable object)
step: Current training step
prefix: Prefix for checkpoint filename
metadata: Additional metadata to save with checkpoint
"""
checkpoint_name = f"{prefix}_step_{step}.flax"
checkpoint_path = self.checkpoints_dir / checkpoint_name
try:
# Save to disk using Flax serialization
with open(checkpoint_path, "wb") as f:
f.write(flax.serialization.to_bytes(params))
# Save metadata if provided
if metadata:
metadata_path = self.checkpoints_dir / f"{prefix}_step_{step}_metadata.json"
with open(metadata_path, "w") as f:
json.dump(metadata, f, indent=2)
logger.info(f"Checkpoint saved: {checkpoint_path}")
# Log to WandB as artifact
if self.wandb_available:
try:
import wandb
artifact = wandb.Artifact(
name=f"{self.run_name}_{prefix}",
type="model",
metadata=metadata or {},
)
artifact.add_file(str(checkpoint_path))
if metadata:
artifact.add_file(str(metadata_path))
self.wandb_run.log_artifact(artifact)
logger.info("Checkpoint uploaded to WandB")
except Exception as e:
logger.warning(f"Could not upload checkpoint to WandB: {e}")
except Exception as e:
logger.error(f"Error saving checkpoint: {e}")
def save_final_model(self, params: Any, metadata: Optional[Dict[str, Any]] = None):
"""Save the final trained model.
Args:
params: Model parameters
metadata: Additional metadata about the final model
"""
final_model_path = self.run_dir / "final_model.flax"
try:
with open(final_model_path, "wb") as f:
f.write(flax.serialization.to_bytes(params))
if metadata:
metadata_path = self.run_dir / "final_model_metadata.json"
with open(metadata_path, "w") as f:
json.dump(metadata, f, indent=2)
logger.info(f"Final model saved: {final_model_path}")
# Log to WandB
if self.wandb_available:
try:
import wandb
artifact = wandb.Artifact(
name=f"{self.run_name}_final_model",
type="model",
metadata=metadata or {},
)
artifact.add_file(str(final_model_path))
if metadata:
artifact.add_file(str(metadata_path))
self.wandb_run.log_artifact(artifact)
except Exception as e:
logger.warning(f"Could not upload final model to WandB: {e}")
except Exception as e:
logger.error(f"Error saving final model: {e}")
def finish(self):
"""Finalize logging and cleanup."""
# Flush remaining metrics
self._flush_metrics()
logger.info(f"Run complete. Results saved to: {self.run_dir.absolute()}")
# Finish WandB run
if self.wandb_available:
finish_wandb()
def __enter__(self):
"""Context manager entry."""
return self
def __exit__(self, exc_type, exc_val, exc_tb):
"""Context manager exit."""
self.finish()

View file

@ -0,0 +1,69 @@
"""Centralized WandB initialization utilities."""
import logging
from typing import Any, Dict, Optional
logger = logging.getLogger(__name__)
def init_wandb(
project: str,
config: Dict[str, Any],
name: Optional[str] = None,
entity: Optional[str] = None,
sync_tensorboard: bool = False,
save_code: bool = True,
resume: str = "allow",
**kwargs,
):
"""Initialize WandB with standardized settings.
This function provides a centralized way to initialize WandB across different
scripts, ensuring consistent configuration and error handling.
Args:
project: WandB project name
config: Configuration dictionary to log
name: Run name (auto-generated if None)
entity: WandB entity (team/user name)
sync_tensorboard: Whether to sync tensorboard logs
save_code: Whether to save code snapshots
resume: Resume strategy ("allow", "must", "never", "auto")
**kwargs: Additional arguments to pass to wandb.init()
Returns:
wandb.Run object if successful, None otherwise
"""
try:
import wandb
run = wandb.init(
project=project,
entity=entity,
name=name,
config=config,
sync_tensorboard=sync_tensorboard,
save_code=save_code,
resume=resume,
**kwargs,
)
logger.info(f"WandB initialized successfully for project '{project}', run '{run.name}'")
return run
except ImportError:
logger.warning("WandB not installed. Skipping WandB initialization.")
return None
except Exception as e:
logger.error(f"Failed to initialize WandB: {e}")
return None
def finish_wandb():
"""Safely finish the current WandB run."""
try:
import wandb
if wandb.run is not None:
wandb.finish()
logger.info("WandB run finished successfully")
except Exception as e:
logger.warning(f"Error finishing WandB run: {e}")