Update utils/models.py

utils/models.py

@@ -1,58 +1,57 @@
 import torch
 import numpy as np
-from tsfm_public.toolkit.get_model import get_model
 from transformers import pipeline
+from sklearn.preprocessing import MinMaxScaler
 
-def predict_umkm(data, prediction_length=7, confidence=0.85):
+def predict_umkm(data, prediction_length=7, safety_stock=10):
     device = "cuda" if torch.cuda.is_available() else "cpu"
     
     try:
-        # ===== 1. GRANITE-TTM Forecasting =====
-        model = get_model(
-            model_path="ibm-granite/granite-timeseries-ttm-r2",
-            context_length=min(512, len(data)),
-            prediction_length=prediction_length,
-            device=device
-        )
-        
-        # Format input
-        inputs = torch.tensor(data['demand'].values, dtype=torch.float32)
-        inputs = inputs.unsqueeze(0).to(device)  # Shape: [1, seq_len]
+        # ===== 1. Persiapan Data =====
+        scaler = MinMaxScaler()
+        scaled_demand = scaler.fit_transform(data[['demand']]).flatten().tolist()
         
-        # Prediksi
-        with torch.no_grad():
-            preds = model.generate(inputs).cpu().numpy().flatten()
-        
-        # ===== 2. Chronos-T5 Decision =====
+        # ===== 2. Prediksi dengan Chronos-T5 =====
         chronos = pipeline(
             "text-generation",
             model="amazon/chronos-t5-small",
             device=device
         )
         
+        # Format prompt khusus time series
         prompt = f"""
         [INSTRUCTION]
-        Berikan rekomendasi untuk manajemen inventory dengan:
-        - Prediksi {prediction_length} hari: {preds.tolist()}
+        Berikan rekomendasi manajemen inventori untuk {prediction_length} hari ke depan:
+        - Data historis demand: {scaled_demand[-100:]}  # Ambil 100 data terakhir
         - Stok saat ini: {data['supply'].iloc[-1]}
-        - Tingkat kepercayaan: {confidence*100}%
+        - Safety stock: {safety_stock}
         
         [FORMAT]
-        1 kalimat dalam Bahasa Indonesia dengan angka spesifik.
-        Estimasi ROI dalam range persentase.
+        1 kalimat dalam Bahasa Indonesia dengan angka konkret
+        Estimasi ROI dalam persentase
         [/FORMAT]
         """
         
-        response = chronos(prompt, max_length=150)[0]['generated_text']
+        # Generate rekomendasi
+        response = chronos(
+            prompt,
+            max_new_tokens=100,
+            temperature=0.7
+        )[0]['generated_text']
         
         # Ekstrak teks rekomendasi
-        rec_text = response.split("[/FORMAT]")[-1].strip()
+        rekomendasi = response.split("[/FORMAT]")[-1].strip()
+        
+        # ===== 3. Simulasi ROI ===== (Contoh Sederhana)
+        avg_demand = np.mean(data['demand'][-30:])
+        roi = round((avg_demand * 0.2) / (data['supply'].iloc[-1] + safety_stock) * 100, 1)
         
         return {
-            "text": rec_text,
-            "predictions": preds.tolist(),
-            "roi": confidence * 0.8,  # Simulasi ROI
-            "confidence": confidence
+            "rekomendasi": rekomendasi,
+            "prediksi": [avg_demand] * prediction_length,  # Contoh prediksi
+            "roi": roi,
+            "stok_optimal": int(avg_demand * 1.2) + safety_stock,
+            "anomali": len(data) - len(clean_data(data))
         }
         
     except Exception as e: