What it does

TensorStream lets you build and run machine learning models using Ruby syntax and TensorFlow-like dataflow graphs. It ships with two backends: a pure-Ruby evaluator that works across MRI, JRuby, and TruffleRuby, and an optional OpenCL evaluator for GPU/CPU acceleration on supported hardware. You can train on OpenCL, then deploy the trained model on plain Ruby.

The interesting bit

The project isn’t a wrapper around Python TensorFlow — it’s a ground-up reimplementation of the graph model, auto-differentiation, and core ops (tf.add, tf.matmul, placeholders, variables, etc.) in Ruby itself. The .t method on integers, floats, and arrays to cast them into tensors is a nice Ruby-ish touch.

Key highlights

• Pure Ruby backend runs on MRI, JRuby, and TruffleRuby; OpenCL backend requires MRI only
• Supports auto-differentiation via formal derivation, with optimizers including Adam, RMSProp, Adagrad, and momentum variants
• Experimental import/export of TensorFlow pbtext files for model portability
• Includes GraphML visualization output (yED-compatible) since TensorBoard isn’t supported
• OpenCL kernels are generic and work across Intel, AMD, and Nvidia hardware

Caveats

• OpenCL is basically required for deep learning or image tasks; the README bluntly states pure Ruby is “too slow even with JIT speedups”
• Scalar-only workloads (like the included linear regression sample) actually run slower on OpenCL due to overhead
• TensorBoard not supported; visualization is limited to GraphML dumps
• “Not all ops are available” — compatibility with TensorFlow is partial, and the Python-to-Ruby porting guide highlights naming and parameter convention differences

Verdict

Worth a look if you’re committed to a Ruby stack and need lightweight model deployment without Python dependencies. Skip it if you need full TensorFlow compatibility, mature GPU acceleration (CUDA isn’t mentioned), or serious deep learning performance.