# TensorBoard

> Visualization toolkit for TensorFlow and PyTorch.

- Category: AI & Machine Learning
- Difficulty: beginner
- Components: tensorboard
- Tags: tensorboard, visualization, tensorflow, pytorch
- Source: https://docker.recipes/ai-ml/tensorboard

## Overview

TensorBoard is Google's visualization toolkit originally designed for TensorFlow but now supporting multiple machine learning frameworks including PyTorch. Released in 2015 as part of TensorFlow, TensorBoard has become the de facto standard for ML experiment tracking and visualization, enabling data scientists and ML engineers to debug neural networks, track training metrics, visualize model architectures, and analyze high-dimensional embeddings through an intuitive web interface. This containerized TensorBoard deployment provides a centralized visualization server that can monitor training runs from multiple sources simultaneously, making it invaluable for teams working on machine learning projects where understanding model behavior and performance is critical. Data scientists, ML engineers, and researchers will find this setup particularly valuable for experiment tracking and model debugging. The Docker deployment eliminates installation complexity while providing a consistent environment for visualization across different development and production environments, making it easy to share results with team members and stakeholders.

## Features

- Real-time scalar metrics visualization for tracking loss, accuracy, and custom metrics during model training
- Interactive histogram and distribution plots for monitoring weight and bias changes across training epochs
- Computational graph visualization showing TensorFlow model architecture and operation flow
- High-dimensional embedding visualization using t-SNE and PCA for analyzing word embeddings and feature representations
- Image and audio sample display for computer vision and audio processing model outputs
- Hyperparameter tuning dashboard with parallel coordinates plots for comparing experiment configurations
- Profile tab for analyzing model performance bottlenecks and GPU utilization
- Multi-run comparison interface for evaluating different model architectures and training strategies

## Use cases

- Deep learning research teams tracking experiments across multiple model architectures and hyperparameter configurations
- Computer vision projects monitoring training progress and visualizing convolutional layer activations
- Natural language processing workflows analyzing word embeddings and attention mechanisms
- MLOps pipelines requiring centralized experiment tracking and model performance monitoring
- Educational institutions teaching machine learning concepts through visual model behavior analysis
- Production ML teams debugging model training issues and optimizing neural network performance
- Data science consultancies presenting model development progress and results to clients

## Prerequisites

- Docker and Docker Compose installed with at least 2GB available RAM for TensorFlow container
- Port 6006 available on the host system for TensorBoard web interface access
- Basic understanding of TensorFlow or PyTorch logging mechanisms and summary operations
- Local logs directory structure with ML training outputs in TensorBoard-compatible format
- Familiarity with machine learning metrics and model training concepts for effective visualization interpretation

## docker-compose.yml

```yaml
services:
  tensorboard:
    image: tensorflow/tensorflow:latest
    container_name: tensorboard
    restart: unless-stopped
    command: tensorboard --logdir=/logs --host=0.0.0.0
    volumes:
      - ./logs:/logs
    ports:
      - "6006:6006"
```

## .env

```bash
# Place logs in ./logs directory
```

## Usage notes

- Docs: https://www.tensorflow.org/tensorboard
- Access at http://localhost:6006 - auto-refreshes with new logs
- TensorFlow: tf.summary.create_file_writer('./logs')
- PyTorch: from torch.utils.tensorboard import SummaryWriter
- View scalars, images, histograms, graphs, and embeddings
- Compare runs by organizing subdirectories under ./logs

## Troubleshooting

- No data appears in TensorBoard interface: Ensure your ML code is writing summary data to the ./logs directory and check file permissions
- Container exits with 'logdir does not exist' error: Create the ./logs directory on the host system before starting the container
- TensorBoard shows empty dashboard: Verify your training scripts are using tf.summary.FileWriter or torch.utils.tensorboard.SummaryWriter correctly
- Scalar plots not updating: Check that your training code is calling summary_writer.flush() periodically to write buffered data
- High memory usage warnings: Reduce the number of histogram summaries or increase sampling intervals in your logging code
- Permission denied errors on log files: Ensure the Docker container has read access to the logs directory with proper ownership settings

