TensorFlow

​

Overview

• Run it directly to get a managed TensorFlow environment for ML/DL development
• Inherit it in your own ML applications to add training and inference capabilities

​

What this template manages

• TensorFlow container (tensorflow/tensorflow image)
• Jupyter notebook server (optional)
• GPU acceleration support (with tensorflow-gpu)
• Python environment with TensorFlow libraries
• Model serving with TensorFlow Serving

​

Quick start (run directly)

1. Load templates

monk load MANIFEST

2. Run TensorFlow with defaults

monk run tensorflow/tensorflow

3. Access Jupyter notebook (if enabled)

4. Customize configuration (recommended via inheritance)

• Preferred: inherit and replace variables with secret("...") as shown below.
• Alternative: fork/clone and edit the variables in tensorflow/tensorflow.yml, then monk load MANIFEST and run.

​

Configuration

variables:
  tensorflow-image-tag: "latest"      # container image tag
  jupyter-enabled: "true"             # enable Jupyter notebook
  jupyter-port: "8888"                # Jupyter port
  gpu-support: "false"                # use tensorflow-gpu image
  python-version: "3.9"               # Python version

​

Use by inheritance (recommended for ML apps)

namespace: myapp
ml-trainer:
  defines: runnable
  inherits: tensorflow/tensorflow
  variables:
    gpu-support: "true"
    tensorflow-image-tag: "latest-gpu"
  files:
    training-script:
      container: tensorflow
      path: /workspace/train.py
      contents: |
        import tensorflow as tf
        # Your training code here
api:
  defines: runnable
  containers:
    api:
      image: myorg/api
  connections:
    ml-service:
      runnable: ml-trainer
      service: tensorflow
  variables:
    ml-host:
      value: <- connection-hostname("ml-service")

monk run myapp/api

​

Ports and connectivity

• Service: tensorflow (configurable ports)
• Jupyter: TCP port 8888 (if enabled)
• TensorFlow Serving: TCP port 8501 (REST API)
• gRPC: TCP port 8500 (for serving)
• From other runnables in the same process group, use connection-hostname("\<connection-name>") to resolve the TensorFlow host.

​

Persistence and configuration

• Notebooks path: ${monk-volume-path}/tensorflow/notebooks
• Models path: ${monk-volume-path}/tensorflow/models
• Data path: ${monk-volume-path}/tensorflow/data
• You can mount additional configuration files and datasets to the container as needed.

​

Features

• Complete ML/DL framework
• Keras high-level API
• Eager execution for intuitive development
• GPU and TPU acceleration
• TensorFlow Serving for production inference
• TensorFlow Lite for mobile and embedded
• TensorFlow.js for browser-based ML
• Distributed training support
• Model optimization and quantization

​

Use cases

• Image classification and object detection
• Natural language processing
• Time series forecasting
• Recommendation systems
• Reinforcement learning
• Generative models (GANs, VAEs)
• Any deep learning task

​

GPU Support

variables:
  tensorflow-image-tag: "latest-gpu"
  gpu-support: "true"

• NVIDIA GPU with CUDA support
• NVIDIA drivers installed
• Docker GPU runtime configured

​

Model Serving

• REST and gRPC APIs
• Model versioning
• Batching for efficiency
• Hot-reloading of models

​

Related templates

• See other ML/AI templates in this repository for complementary services.
• Combine with monitoring tools for observability of training jobs and model serving.

​

Troubleshooting

• For GPU support, verify CUDA compatibility with TensorFlow version and ensure NVIDIA drivers are properly installed.
• Ensure sufficient RAM/VRAM for your models. Monitor GPU memory usage to avoid OOM errors.
• Ensure host volumes are writable by the container user.
• Check logs:

monk logs -l 500 -f tensorflow/tensorflow

• Check Jupyter logs for token:

monk logs -l 100 tensorflow/tensorflow | grep token

• Verify TensorFlow installation inside the container:

import tensorflow as tf
print(tf.__version__)
print(tf.config.list_physical_devices('GPU'))