Koios Model Inference Standards

This section outlines how to properly format and deliver input data to Koios for AI model inference. Koios is designed to work with .tflite models, using timestamped time-series data from your industrial or commercial systems.

By following these guidelines, you ensure your model performs correctly and consistently within the Koios platform.

Supported Model Format

Koios requires all models to be in .tflite format for deployment. Below are the most common paths to convert models built in other frameworks

Model Conversions to .tflite

Tensorflow (Keras —> .tflite)

✅ Supported:

• Keras (.h5)

• SavedModel format

🔧 Tool:

• tf.lite.TFLiteConverter (TensorFlow to TFLite)

Conversion Example

import tensorflow as tf

# From SavedModel
converter = tf.lite.TFLiteConverter.from_saved_model("path/to/saved_model")
tflite_model = converter.convert()

# Save to file
with open("model.tflite", "wb") as f:
    f.write(tflite_model)

PyTorch —> TFLite

⌛ Steps:

1. Export PyTorch model to ONNX format.

2. Convert ONNX to TensorFlow SavedModel.

3. Convert SavedModel to TFLite.

🛠️ Tools:

• torch.onnx.export (PyTorch to ONNX)

• onnx-tf or onnx2tf (ONNX to TensorFlow)

• tf.lite.TFLiteConverter (TensorFlow to TFLite)

Example Workflow

# Step 1: PyTorch to ONNX
import torch
import torchvision.models as models

model = models.resnet18(pretrained=True)
model.eval()
dummy_input = torch.randn(1, 3, 224, 224)
torch.onnx.export(model, dummy_input, "model.onnx")

# Step 2: ONNX to TensorFlow
# Terminal command using onnx-tf:
# pip install onnx-tf
onnx-tf convert -i model.onnx -o tf_model

# Step 3: TensorFlow to TFLite (as in previous example)

Input Data Format

Koios expects a 2D array for model input, shaped as follows:

Input Shape: [1, T, N]
# Where:
#  - 1 = Sample
#  - T = Number of time steps (depth of history, fixed)
#  - N = Number of tags (sensor inputs/features)

Example Input Layout

Coming Soon: Variable-Length Inputs

Future versions of Koios will allow each tag to have a different historical depth. This feature is inspired by flexible input handling approaches like gym.spaces.Dict or stable_baselines3's spaces.Tuple.

This will enable models to use:

• Long-term history for slow-changing variables

• Short-term slices for fast-changing variables

Stay tuned for updates in future releases.

Best Practices

• Ensure your model was trained with [1, T, N] data, in the same format as Koios expects.

• Preprocess and align tag data (handle nulls, normalize) before using it for inference.

• Validate .tflite models offline using test arrays that mimic expected Koios input format.

• Use Koios’ built-in historian to avoid relying on external systems for inference.

Troubleshooting