Weights Biases Career Development IT

W&B MLOps

Weights Biases Career Development IT MLOps Experiment Tracking Model Registry Hyperparameter Sweep Artifacts Reports Portfolio Resume Production ML

Experiment Tracking

# === W&B Experiment Tracking ===

# pip install wandb torch torchvision

# import wandb
# import torch
# import torch.nn as nn
#
# # Initialize
# wandb.init(
#     project="image-classifier",
#     name="resnet50-baseline",
#     config={
#         "architecture": "ResNet50",
#         "dataset": "CIFAR-10",
#         "epochs": 50,
#         "batch_size": 32,
#         "learning_rate": 0.001,
#         "optimizer": "Adam",
#         "weight_decay": 1e-4,
#     }
# )
# config = wandb.config
#
# # Training Loop
# for epoch in range(config.epochs):
#     model.train()
#     for batch_idx, (data, target) in enumerate(train_loader):
#         optimizer.zero_grad()
#         output = model(data)
#         loss = criterion(output, target)
#         loss.backward()
#         optimizer.step()
#
#         wandb.log({
#             "train/loss": loss.item(),
#             "train/batch": batch_idx,
#             "epoch": epoch,
#         })
#
#     # Validation
#     val_loss, val_acc = evaluate(model, val_loader)
#     wandb.log({
#         "val/loss": val_loss,
#         "val/accuracy": val_acc,
#         "epoch": epoch,
#         "learning_rate": optimizer.param_groups[0]["lr"],
#     })
#
#     # Log model checkpoint
#     if val_acc > best_acc:
#         wandb.save("best_model.pth")
#         best_acc = val_acc
#
# wandb.finish()

from dataclasses import dataclass

@dataclass
class ExperimentRun:
    run_name: str
    model: str
    lr: float
    batch_size: int
    val_acc: float
    val_loss: float
    epochs: int
    duration: str

runs = [
    ExperimentRun("resnet50-baseline", "ResNet50", 0.001, 32, 0.923, 0.245, 50, "2h 15m"),
    ExperimentRun("resnet50-lr-high", "ResNet50", 0.01, 32, 0.891, 0.312, 50, "2h 10m"),
    ExperimentRun("resnet50-batch64", "ResNet50", 0.001, 64, 0.918, 0.258, 50, "1h 45m"),
    ExperimentRun("efficientnet-b0", "EfficientNet-B0", 0.001, 32, 0.935, 0.198, 50, "3h 20m"),
    ExperimentRun("vit-small", "ViT-Small", 0.0003, 32, 0.941, 0.182, 100, "5h 40m"),
]

print("=== Experiment Runs ===")
for r in runs:
    print(f"  [{r.run_name}] Model: {r.model}")
    print(f"    LR: {r.lr} | BS: {r.batch_size} | Epochs: {r.epochs}")
    print(f"    Val Acc: {r.val_acc:.3f} | Val Loss: {r.val_loss:.3f} | Time: {r.duration}")

Hyperparameter Sweep

# === W&B Sweep Configuration ===

# sweep_config = {
#     "method": "bayes",  # grid, random, bayes
#     "metric": {
#         "name": "val/accuracy",
#         "goal": "maximize"
#     },
#     "parameters": {
#         "learning_rate": {
#             "min": 0.0001,
#             "max": 0.01,
#             "distribution": "log_uniform_values"
#         },
#         "batch_size": {
#             "values": [16, 32, 64, 128]
#         },
#         "optimizer": {
#             "values": ["adam", "sgd", "adamw"]
#         },
#         "weight_decay": {
#             "min": 1e-5,
#             "max": 1e-2,
#             "distribution": "log_uniform_values"
#         },
#         "dropout": {
#             "min": 0.1,
#             "max": 0.5
#         }
#     },
#     "early_terminate": {
#         "type": "hyperband",
#         "min_iter": 10,
#         "eta": 3
#     }
# }
#
# sweep_id = wandb.sweep(sweep_config, project="image-classifier")
# wandb.agent(sweep_id, function=train, count=50)

@dataclass
class SweepResult:
    rank: int
    lr: float
    batch_size: int
    optimizer: str
    dropout: float
    val_acc: float

results = [
    SweepResult(1, 0.00032, 32, "adamw", 0.25, 0.945),
    SweepResult(2, 0.00045, 64, "adam", 0.20, 0.941),
    SweepResult(3, 0.00028, 32, "adamw", 0.30, 0.939),
    SweepResult(4, 0.001, 32, "adam", 0.15, 0.935),
    SweepResult(5, 0.0005, 128, "sgd", 0.20, 0.928),
]

print("\n=== Top 5 Sweep Results ===")
for r in results:
    print(f"  Rank {r.rank}: Val Acc {r.val_acc:.3f}")
    print(f"    LR: {r.lr} | BS: {r.batch_size} | Opt: {r.optimizer} | Drop: {r.dropout}")

Career Portfolio

# === Building ML Career Portfolio ===

@dataclass
class PortfolioProject:
    project: str
    tech_stack: str
    wandb_features: str
    skill_demonstrated: str
    portfolio_url: str

projects = [
    PortfolioProject("Image Classification", "PyTorch ResNet EfficientNet",
        "Experiment Tracking, Sweep, Model Registry",
        "Deep Learning, Hyperparameter Tuning, MLOps",
        "wandb.ai/username/image-classifier"),
    PortfolioProject("NLP Sentiment Analysis", "HuggingFace Transformers BERT",
        "Fine-tuning Tracking, Dataset Versioning",
        "NLP, Transfer Learning, Data Pipeline",
        "wandb.ai/username/sentiment"),
    PortfolioProject("Object Detection", "YOLOv8 COCO Dataset",
        "mAP Tracking, Confusion Matrix, Media Panel",
        "Computer Vision, Model Evaluation",
        "wandb.ai/username/yolo-detection"),
    PortfolioProject("Time Series Forecast", "Prophet LSTM Transformer",
        "Multi-model Comparison, Reports",
        "Time Series, Feature Engineering",
        "wandb.ai/username/forecasting"),
]

print("ML Career Portfolio:")
for p in projects:
    print(f"  [{p.project}] Stack: {p.tech_stack}")
    print(f"    W&B: {p.wandb_features}")
    print(f"    Skills: {p.skill_demonstrated}")
    print(f"    URL: {p.portfolio_url}")

career_path = {
    "Junior ML Engineer": "W&B basics, Experiment Tracking, 2-3 projects",
    "ML Engineer": "Sweeps, Model Registry, CI/CD, 5+ projects",
    "Senior ML Engineer": "Full MLOps pipeline, Team collaboration, Reports",
    "ML Architect": "Platform design, W&B Server self-hosted, Enterprise",
    "Research Scientist": "Advanced experiments, Paper-quality Reports",
}

print(f"\n\nCareer Path:")
for k, v in career_path.items():
    print(f"  [{k}]: {v}")

เคล็ดลับ

• wandb.init: ตั้ง Project Name และ Config ทุกครั้ง
• Sweep: ใช้ Bayes Method สำหรับ Sweep ประสิทธิภาพดีกว่า Random
• Report: สร้าง W&B Report แชร์ผลงานแทน Notebook
• Portfolio: ใช้ W&B Public Profile เป็น ML Portfolio
• Community: เข้าร่วม W&B Community เรียนรู้จากู้คืนอื่น

การนำความรู้ไปประยุกต์ใช้งานจริง

แหล่งเรียนรู้ที่แนะนำ ได้แก่ Official Documentation ที่อัพเดทล่าสุดเสมอ Online Course จาก Coursera Udemy edX ช่อง YouTube คุณภาพทั้งไทยและอังกฤษ และ Community อย่าง Discord Reddit Stack Overflow ที่ช่วยแลกเปลี่ยนประสบการณ์กับนักพัฒนาทั่วโลก

สรุป

Weights Biases Career Development IT MLOps Experiment Tracking Model Registry Hyperparameter Sweep Portfolio Resume Reports Community Production ML