neurenix export

Export a trained model to various formats for deployment

The export command converts a trained Neurenix model to different formats for deployment across various platforms and frameworks.

Usage

neurenix export --model <model_file> --format <format> [options]

Options

Option	Type	Required	Default	Description
--model	string	Yes	-	Path to the model file (.nrx format)
--format	string	Yes	-	Export format (see formats below)
--output	string	No	Auto-generated	Output file or directory path
--optimize	flag	No	false	Optimize the exported model
--quantize	string	No	none	Quantization type (int8, fp16, none)

Supported Formats

Format	Extension	Description	Use Case
onnx	.onnx	Open Neural Network Exchange	Cross-platform deployment
torchscript	.pt	PyTorch TorchScript	PyTorch production
tensorflow	_tf/	TensorFlow SavedModel	TensorFlow serving
tflite	.tflite	TensorFlow Lite	Mobile and edge devices
wasm	.wasm	WebAssembly	Browser and edge deployment
c	.c	C source code	Embedded systems

Examples

Export to ONNX

neurenix export --model models/model.nrx --format onnx

Loading model from models/model.nrx...
Exporting model to onnx format...
Model exported to models/model.onnx

Export to TorchScript

neurenix export --model models/model.nrx --format torchscript

Loading model from models/model.nrx...
Exporting model to torchscript format...
Model exported to models/model.pt

Export to TensorFlow

neurenix export --model models/model.nrx --format tensorflow

Loading model from models/model.nrx...
Exporting model to tensorflow format...
Model exported to models/model_tf

Export to TensorFlow Lite

neurenix export --model models/model.nrx --format tflite

Loading model from models/model.nrx...
Exporting model to tflite format...
Model exported to models/model.tflite

Export to WebAssembly

neurenix export --model models/model.nrx --format wasm

Loading model from models/model.nrx...
Exporting model to wasm format...
Model exported to models/model.wasm

Export to C

neurenix export --model models/model.nrx --format c

Loading model from models/model.nrx...
Exporting model to c format...
Model exported to models/model.c

Custom output path

neurenix export \
  --model models/model.nrx \
  --format onnx \
  --output deployment/production.onnx

Loading model from models/model.nrx...
Exporting model to onnx format...
Model exported to deployment/production.onnx

Export with optimization

neurenix export \
  --model models/model.nrx \
  --format onnx \
  --optimize

Loading model from models/model.nrx...
Exporting model to onnx format...
Optimizing model...
Model exported to models/model.onnx

Export with quantization

neurenix export \
  --model models/model.nrx \
  --format tflite \
  --quantize int8

Loading model from models/model.nrx...
Exporting model to tflite format...
Quantizing model to int8...
Model exported to models/model.tflite

Export with all optimizations

neurenix export \
  --model models/model.nrx \
  --format onnx \
  --optimize \
  --quantize fp16 \
  --output deployment/optimized.onnx

Loading model from models/model.nrx...
Exporting model to onnx format...
Optimizing model...
Quantizing model to fp16...
Model exported to deployment/optimized.onnx

Quantization Options

INT8 Quantization

Reduces model size by ~75% with minimal accuracy loss:

neurenix export \
  --model models/model.nrx \
  --format tflite \
  --quantize int8

Benefits:

• Smaller model size (4x reduction)
• Faster inference on compatible hardware
• Lower memory usage

Trade-offs:

• Small accuracy degradation (typically less than 1%)
• Requires calibration data for best results

FP16 Quantization

Reduces model size by ~50% with negligible accuracy loss:

neurenix export \
  --model models/model.nrx \
  --format onnx \
  --quantize fp16

Benefits:

• Smaller model size (2x reduction)
• Faster inference on GPUs
• Minimal accuracy loss

Trade-offs:

• Less size reduction than INT8
• Requires FP16-capable hardware for speedup

Format-Specific Details

ONNX (.onnx)

Best for: Cross-platform deployment, inference optimization

neurenix export --model models/model.nrx --format onnx

Usage:

import onnxruntime as ort

session = ort.InferenceSession("models/model.onnx")
outputs = session.run(None, {"input": input_data})

TorchScript (.pt)

Best for: PyTorch production environments

neurenix export --model models/model.nrx --format torchscript

Usage:

import torch

model = torch.jit.load("models/model.pt")
output = model(input_tensor)

TensorFlow Lite (.tflite)

Best for: Mobile apps (Android/iOS), edge devices

neurenix export --model models/model.nrx --format tflite --quantize int8

Usage:

import tensorflow as tf

interpreter = tf.lite.Interpreter(model_path="models/model.tflite")
interpreter.allocate_tensors()

WebAssembly (.wasm)

Best for: Browser-based inference, edge computing

neurenix export --model models/model.nrx --format wasm --optimize

Usage:

const model = await loadWasmModel('models/model.wasm');
const output = model.predict(inputData);

C (.c)

Best for: Embedded systems, microcontrollers

neurenix export --model models/model.nrx --format c

Usage:

#include "model.c"

float input[INPUT_SIZE];
float output[OUTPUT_SIZE];
model_predict(input, output);

Error Handling

Model not found

neurenix export --model nonexistent.nrx --format onnx

Error: Model file 'nonexistent.nrx' not found.

Invalid format

neurenix export --model models/model.nrx --format invalid

usage: neurenix export [<args>]
neurenix export: error: argument --format: invalid choice: 'invalid' 
(choose from 'onnx', 'torchscript', 'tensorflow', 'tflite', 'wasm', 'c')

Export error

neurenix export --model models/incompatible.nrx --format tflite

Error exporting model: Unsupported operation for TFLite export

Best Practices

1. Test exported models

Always verify exported models match original performance:

# Export model
neurenix export --model models/model.nrx --format onnx

# Test original model
neurenix eval --model models/model.nrx --data data/test.csv

# Test exported model with your deployment framework
# Compare metrics to ensure accuracy is maintained

2. Choose appropriate format for target platform

# Mobile deployment
neurenix export --model models/model.nrx --format tflite --quantize int8

# Web deployment
neurenix export --model models/model.nrx --format wasm --optimize

# Cloud deployment
neurenix export --model models/model.nrx --format onnx --optimize

# Embedded systems
neurenix export --model models/model.nrx --format c

3. Optimize for production

Always use optimization for production deployments:

neurenix export \
  --model models/model.nrx \
  --format onnx \
  --optimize \
  --quantize fp16

4. Version exported models

neurenix export \
  --model models/model_v1.0.nrx \
  --format onnx \
  --output deployment/model_v1.0.onnx

5. Document export settings

Create a script to reproduce exports:

#!/bin/bash
# export_production.sh

neurenix export \
  --model models/production.nrx \
  --format onnx \
  --optimize \
  --quantize fp16 \
  --output deployment/production.onnx

echo "Export completed: deployment/production.onnx"

Deployment Workflow

# 1. Train model
neurenix run train.py

# 2. Evaluate performance
neurenix eval --model models/model.nrx --data data/test.csv

# 3. Export for deployment
neurenix export \
  --model models/model.nrx \
  --format onnx \
  --optimize \
  --quantize fp16

# 4. Verify exported model
# Use your deployment framework to test

# 5. Deploy
neurenix serve --model models/model.nrx --port 8000

See Also

• Serve command - Serve models as APIs
• Eval command - Evaluate models
• Optimize command - Optimize models
• Run command - Train models