Deployment: Model ke Production

import torch # Load trained model model = MyModel() model.load_state_dict(torch.load("model_weights.pt")) model.eval() # Method 1: Tracing (recommended untuk model tanpa control flow) example_input = torch.randn(1, 3, 224, 224) traced = torch.jit.trace(model, example_input) traced.save("model_traced.pt") # Method 2: Scripting (untuk model dengan if/for) scripted = torch.jit.script(model) scripted.save("model_scripted.pt") # Load di production (tanpa definisi class!) loaded = torch.jit.load("model_traced.pt") output = loaded(example_input) # ✅ Bisa run di C++, mobile, tanpa Python!

import torch import onnxruntime as ort # ===== EXPORT ke ONNX ===== model.eval() dummy = torch.randn(1, 3, 224, 224) torch.onnx.export( model, dummy, "model.onnx", input_names=["input"], output_names=["output"], dynamic_axes={"input": {0: "batch"}}, # Dynamic batch size dynamo=True # PyTorch 2.5+ recommended exporter ) # ===== INFERENCE dengan ONNX Runtime ===== session = ort.InferenceSession("model.onnx") result = session.run( None, {"input": dummy.numpy()} ) # ✅ 3× faster than PyTorch eager mode! # ✅ Bisa deploy di: CPU, GPU, TensorRT, OpenVINO, mobile

# pip install fastapi uvicorn onnxruntime pillow from fastapi import FastAPI, UploadFile import onnxruntime as ort import numpy as np from PIL import Image import io app = FastAPI(title="Image Classifier API") # Load ONNX model SEKALI saat startup session = ort.InferenceSession("model.onnx") CLASSES = ["cat", "dog", "bird", "fish", "horse"] def preprocess(image_bytes): img = Image.open(io.BytesIO(image_bytes)).resize((224, 224)) arr = np.array(img).transpose(2,0,1).astype(np.float32) / 255.0 arr = (arr - [.485,.456,.406]) / [.229,.224,.225] return arr[np.newaxis] @app.post("/predict") async def predict(file: UploadFile): data = await file.read() tensor = preprocess(data) result = session.run(None, {"input": tensor}) probs = np.exp(result[0]) / np.exp(result[0]).sum() idx = probs.argmax() return {"class": CLASSES[idx], "confidence": float(probs[0][idx])} # Run: uvicorn 23_fastapi_server:app --host 0.0.0.0 --port 8000 # Test: curl -X POST -F "file=@cat.jpg" http://localhost:8000/predict # → {"class": "cat", "confidence": 0.9847}

FROM python:3.11-slim WORKDIR /app COPY requirements.txt . RUN pip install --no-cache-dir -r requirements.txt COPY model.onnx . COPY 23_fastapi_server.py main.py EXPOSE 8000 CMD ["uvicorn", "main:app", "--host", "0.0.0.0", "--port", "8000"] # Build & Run: # docker build -t ml-api . # docker run -p 8000:8000 ml-api # Image size: ~350MB (tanpa GPU)

import torch.quantization # Dynamic Quantization (paling mudah, CPU only) quantized = torch.quantization.quantize_dynamic( model, {torch.nn.Linear, torch.nn.LSTM}, # Layer yang di-quantize dtype=torch.qint8 # Float32 → Int8 ) # Bandingkan ukuran torch.save(model.state_dict(), "original.pt") # 44.7 MB torch.save(quantized.state_dict(), "quantized.pt") # 11.2 MB ← 4× lebih kecil! # Benchmark # Original: 45ms/inference, 44.7 MB # Quantized: 18ms/inference, 11.2 MB ← 2.5× faster, 4× smaller! # Akurasi drop: ~0.5-1% (hampir tidak terasa)