Machine learning experiments can become difficult to manage very quickly. Models are trained with different datasets. Hyperparameters keep changing. Training metrics are stored in notebooks, spreadsheets, screenshots, or local files. After a few experiments, it becomes hard to remember which model performed best and why.<\/p>\n\n\n\n
Weights & Biases solves this problem by giving machine learning teams a structured way to track experiments, compare model runs, version datasets, log artifacts, monitor metrics, and manage model development workflows.<\/p>\n\n\n\n
This Weights & Biases tutorial explains what W&B is, why it is used, how to install it, how to track experiments, how to log metrics, how to use artifacts, how to run sweeps, and how to manage model outputs. It is written for beginners who understand basic Python, machine learning, training loops, and model evaluation.<\/p>\n\n\n\n
Weights & Biases, also called W&B, is a machine learning platform used to track, visualize, compare, and manage ML experiments. It records training metrics, model configurations, system usage, datasets, model files, and experiment outputs in one dashboard.<\/p>\n\n\n\n
A machine learning project may have hundreds of experiments. Each experiment may use a different learning rate, batch size, optimizer, dataset version, or model architecture. W&B helps teams understand which run produced the best result and what settings were used. At a basic level, W&B connects your Python training script with an online or private workspace. Once connected, it automatically stores training information so teams can review results later.<\/p>\n\n\n\n
Weights & Biases is used because machine learning development needs traceability. A model accuracy number is not enough. Teams need to know the dataset, code version, training settings, loss curve, evaluation metric, and output files behind that number.<\/p>\n\n\n\n
Weights & Biases has a few important building blocks. A run stores one experiment. A project groups related runs. A dashboard visualizes metrics and charts. Artifacts version datasets, models, and outputs. Sweeps automate hyperparameter search.<\/p>\n\n\n\n
Weights & Biases can be installed with pip.<\/p>\n\n\n\n
pip install wandb<\/code><\/pre>\n\n\n\nAfter installation, import the package in your Python script or notebook.<\/p>\n\n\n\n
import wandb<\/code><\/pre>\n\n\n\nThis allows your training code to connect with the W&B workspace.<\/p>\n\n\n\n
2. Log in to W&B<\/strong><\/h3>\n\n\n\nBefore tracking experiments, log in with your W&B account.<\/p>\n\n\n\n
wandb.login()<\/code><\/pre>\n\n\n\nThis connects your local environment with your W&B account. In team environments, the same setup can be used inside notebooks, training servers, or ML pipelines.<\/a><\/p>\n\n\n\n3. Initialize a W&B Run<\/strong><\/h3>\n\n\n\nA run is the core unit of experiment tracking in W&B. Each training job should usually start with wandb.init().<\/p>\n\n\n\n
import wandb\n\nrun = wandb.init(\n\n project=\"image-classification-demo\",\n\n name=\"baseline-cnn\",\n\n config={\n\n \"learning_rate\": 0.001,\n\n \"epochs\": 10,\n\n \"batch_size\": 32,\n\n \"optimizer\": \"adam\"\n\n }\n\n)<\/code><\/pre>\n\n\n\nThe project name groups related experiments. The run name helps identify a specific experiment. The config stores training settings.<\/p>\n\n\n\n
Build industry-ready AI and machine learning skills with HCL GUVI\u2019s <\/em>Artificial Intelligence and Machine Learning Program<\/em><\/a>. Learn Python, SQL, ML, MLOps, Generative AI, Agentic AI, and real-world model development through live online classes, Intel-designed curriculum, 1:1 doubt sessions, industry-grade projects, capstone learning, and placement assistance with 1000+ hiring partners.<\/em><\/p>\n\n\n\n4. Log Training Metrics<\/strong><\/h3>\n\n\n\nMetrics can be logged during training with wandb.log().<\/p>\n\n\n\n
for epoch in range(10):\n\n train_loss = 0.45 - epoch * 0.02\n\n val_accuracy = 0.70 + epoch * 0.02\n\n wandb.log({\n\n \"epoch\": epoch,\n\n \"train_loss\": train_loss,\n\n \"val_accuracy\": val_accuracy\n\n })<\/code><\/pre>\n\n\n\nThis creates real-time charts for loss, accuracy, and other values. Data scientists<\/a> can compare curves across multiple runs instead of manually checking notebook outputs.<\/p>\n\n\n\n5. Track Hyperparameters<\/strong><\/h3>\n\n\n\nHyperparameters are stored inside the run config. This helps teams reproduce results later.<\/p>\n\n\n\n
config = wandb.config\n\nlearning_rate = config.learning_rate\n\nepochs = config.epochs\n\nbatch_size = config.batch_size<\/code><\/pre>\n\n\n\nThis is useful when several experiments use different learning rates, optimizers, dropout values, or model sizes.<\/p>\n\n\n\n
Build strong foundations in machine learning and MLOps with HCL GUVI\u2019s <\/em>Machine Learning Course. <\/em><\/a>Learn core concepts like model building, training workflows, data handling, and real-world ML applications through structured, hands-on training designed for aspiring ML engineers.<\/em><\/p>\n\n\n\n6. Log Model Files with Artifacts<\/strong><\/h3>\n\n\n\nArtifacts help version files used or produced by ML experiments. These files may include datasets, model checkpoints, tokenizers, feature files, or prediction outputs.<\/p>\n\n\n\n
artifact = wandb.Artifact(\n\n name=\"cnn-model\",\n\n type=\"model\"\n\n)\n\nartifact.add_file(\"model.pt\")\n\nwandb.log_artifact(artifact)<\/code><\/pre>\n\n\n\nArtifacts make model tracking more reliable because each output can be linked to the run that created it.<\/p>\n\n\n\n
7. Log Predictions with W&B Tables<\/strong><\/h3>\n\n\n\nW&B Tables help inspect predictions, labels, scores, images, and text outputs.<\/p>\n\n\n\n
table = wandb.Table(\n\n columns=[\"image_id\", \"actual_label\", \"predicted_label\", \"confidence\"]\n\n)\n\ntable.add_data(\"img_001\", \"cat\", \"cat\", 0.94)\n\ntable.add_data(\"img_002\", \"dog\", \"cat\", 0.61)\n\nwandb.log({\"prediction_table\": table})<\/code><\/pre>\n\n\n\nTables are useful for error analysis. A team can quickly find misclassified examples and understand where the model fails.<\/p>\n\n\n\n
8. Run Hyperparameter Sweeps<\/strong><\/h3>\n\n\n\nSweeps help automate hyperparameter tuning. Instead of manually running many experiments, teams define a search space.<\/p>\n\n\n\n
sweep_config = {\n\n \"method\": \"random\",\n\n \"metric\": {\n\n \"name\": \"val_accuracy\",\n\n \"goal\": \"maximize\"\n\n },\n\n \"parameters\": {\n\n \"learning_rate\": {\n\n \"values\": [0.001, 0.0005, 0.0001]\n\n },\n\n \"batch_size\": {\n\n \"values\": [16, 32, 64]\n\n }\n\n }\n\n}\n\nsweep_id = wandb.sweep(\n\n sweep_config,\n\n project=\"image-classification-demo\"\n\n)\n\nA sweep agent can then run experiments using different combinations.\n\nwandb.agent(sweep_id, function=train_model, count=5)<\/code><\/pre>\n\n\n\nSweeps are useful for improving model performance without manual experiment tracking.<\/p>\n\n\n\n
9. Compare Experiments<\/strong><\/h3>\n\n\n\nThe W&B dashboard allows teams to compare runs by accuracy, loss, runtime, hyperparameters, and system metrics. This helps answer practical questions.<\/p>\n\n\n\n
\n- Which model performed best?<\/li>\n\n\n\n
- Which learning rate worked better?<\/li>\n\n\n\n
- Which run overfitted?<\/li>\n\n\n\n
- Which dataset version was used?<\/li>\n\n\n\n
- Which checkpoint should be promoted?<\/li>\n<\/ul>\n\n\n\n