Deployment

KladML provides robust tools for deploying your trained models to various environments, with a primary focus on Edge and Embedded devices via TorchScript optimization.

Automatic Export

Starting from v0.3.0, every training run automatically generates a deployment-optimized artifact alongside the standard checkpoint.

When train() completes (or finds a best model), it creates: - checkpoints/best_model.pth: Python-dependent PyTorch checkpoint (contains weights + optimizer state). Use for resuming training or fine-tuning. - checkpoints/best_model_jit.pt: Deployment Artifact. A standalone TorchScript file containing the model architecture, weights, and embedded scaler statistics.

Manual Export

You can export any existing checkpoint using the CLI:

kladml export \
    --checkpoint ./data/projects/my_project/experiments/exp1/run_id/checkpoints/best_model.pth \
    --output ./deployment/model.pt \
    --config ./data/projects/my_project/experiments/exp1/run_id/config.yaml

Using the Deployed Model (Edge/Production)

The exported .pt file is self-contained. You do not need the kladml library installed on the target device, only torch.

Python (Edge/Server)

import torch

# 1. Load Model
model = torch.jit.load("model.pt")
model.eval()

# 2. Extract embedded scalar stats (Zero external dependencies!)
mean = float(model.metadata()["scaler_mean"])
scale = float(model.metadata()["scaler_scale"])

# 3. Inference
# Input: [Batch, SeqLen, 1] (Raw values)
input_data = torch.randn(1, 60, 1) 

# Preprocess (Standard Scaling)
input_scaled = (input_data - mean) / scale

# Forward Pass
# Output shape depends on the model.
# See specific architecture documentation for output details (e.g., gluformer/GLUFORMER_ARCHITECTURE.md).
outputs = model(input_scaled)

C++ (Embedded)

Since it is a TorchScript module, it can be loaded in C++ using libtorch:

#include <torch/script.h>

int main() {
  torch::jit::script::Module module = torch::jit::load("model.pt");
  // ... inference logic ...
}

Supported Architectures

Each architecture may implement custom wrapping logic for deployment. See the architecture-specific documentation for details on input shapes and output formats.

• Gluformer: See src/kladml/architectures/gluformer/GLUFORMER_ARCHITECTURE.md