嵌入式落地實戰
模型壓縮、量化與邊緣部署

import torch
from torch.quantization import quantize_dynamic, prepare, convert

# 方法一：動態量化（Dynamic Quantization）
# 只量化線性層，activation 在運行時動態量化
model_int8 = quantize_dynamic(
    model,
    qconfig_spec={torch.nn.Linear},
    dtype=torch.qint8
)

# 方法二：靜態量化（Static Quantization）需要校準步驟
model.qconfig = torch.quantization.get_default_qconfig('fbgemm')
model_prepared = prepare(model)

# 跑一批校準數據
with torch.no_grad():
    for data in calibration_loader:
        model_prepared(data)

# 轉換為真正的 INT8 模型
model_int8 = convert(model_prepared)

# 驗證：模型大小縮小 ~4×
print(f"FP32 大小: {get_size_mb(model):.1f} MB")
print(f"INT8 大小: {get_size_mb(model_int8):.1f} MB")

import torch.nn.functional as F

def distillation_loss(student_logits, teacher_logits, labels, T=4.0, alpha=0.5):
    """
    T: Temperature（溫度）—— 越高，Soft Labels 越平滑，包含更多「暗知識」
    alpha: Hard Loss 的權重，(1-alpha) 給 Soft Loss
    """
    # Hard Loss：Student 輸出 vs 真實標籤
    hard_loss = F.cross_entropy(student_logits, labels)

    # Soft Loss：Student 與 Teacher 的輸出分布 KL 散度
    soft_student = F.log_softmax(student_logits / T, dim=-1)
    soft_teacher = F.softmax(teacher_logits / T, dim=-1)
    soft_loss = F.kl_div(soft_student, soft_teacher, reduction='batchmean') * (T ** 2)

    return alpha * hard_loss + (1 - alpha) * soft_loss

import tensorflow as tf

# 1. 將 PyTorch/ONNX 模型轉為 TFLite + INT8
converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir)
converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.target_spec.supported_types = [tf.int8]

# 2. 提供校準數據集（必須是真實數據，~100 筆即可）
def calibration_data_gen():
    for sample in calibration_dataset:
        yield [sample]

converter.representative_dataset = calibration_data_gen
tflite_model = converter.convert()

# 3. 儲存模型
with open('tts_model_int8.tflite', 'wb') as f:
    f.write(tflite_model)

# --- Android 端使用 NNAPI（自動路由到 NPU）---
# val options = Interpreter.Options().apply {
#     addDelegate(NnApiDelegate())  // 自動選擇 NPU/DSP/GPU
# }
# val interpreter = Interpreter(loadModel("tts_model_int8.tflite"), options)