export_eai

https://github.com/quic/aimet?tab=readme-ov-file#supported-features import os import shlex import subprocess import numpy as np

------------------------------

根据图2提供的工具路径 (必须配置)

------------------------------

eai_builder="/local/mnt/workspace/Qualcomm/Hexagon_SDK/6.2.0.1/addons/LPAI/eNPU6/5.7.0/eai_builder/eai_builder" eai_quantize="/local/mnt/workspace/Qualcomm/Hexagon_SDK/6.2.0.1/addons/LPAI/eNPU6/5.7.0/utilities/tensor_quantization/eai_quantize" eai_sample_app="/local/mnt/workspace/Qualcomm/Hexagon_SDK/6.2.0.1/addons/LPAI/eNPU6/5.7.0/binaries/linux_x86/eai_sample_app/fixed32/eai_sample_app"

------------------------------

核心处理函数 (基于图1流程)

------------------------------

def run_eai_pipeline( eai_model: str, # EAI模型路径 (.onnx等) prep_data_root: str, # 输入数据目录 eai_model_dir: str, # 工作目录 float_input_name: str, # 浮点输入文件名 float_cached_len_name: str, # 缓存长度文件名 float_cached_avg_name: str # 缓存平均值文件名 ) -> dict: """执行完整的EAI模型量化->推理->反量化流程"""

# 阶段1: 复制输入数据到工作目录
for file_name in [float_input_name, float_cached_len_name, float_cached_avg_name]:
    src = os.path.join(prep_data_root, file_name)
    dst = os.path.join(eai_model_dir, file_name)
    cmd = f"cp {src} {dst}"
    result = subprocess.run(shlex.split(cmd), shell=False)
    assert result.returncode == 0, f"复制失败: {cmd}"

# 阶段2: 量化模型输入
quant_cmd = (
    f"{eai_quantize} --eai {eai_model} --quantize 1 --input "
    f"{os.path.join(eai_model_dir, float_input_name)} "
    f"{os.path.join(eai_model_dir, float_cached_len_name)} "
    f"{os.path.join(eai_model_dir, float_cached_avg_name)}"
)
result = subprocess.run(shlex.split(quant_cmd), shell=False)
assert result.returncode == 0, f"量化失败: {quant_cmd}"

# 阶段3: 执行EAI模型推理
infer_cmd = (
    f"{eai_sample_app} -m {eai_model} -o {eai_model_dir} -i "
    f"{os.path.join(eai_model_dir, 'quantized_' + float_input_name)} "
    f"{os.path.join(eai_model_dir, 'quantized_' + float_cached_len_name)} "
    f"{os.path.join(eai_model_dir, 'quantized_' + float_cached_avg_name)}"
)
result = subprocess.run(shlex.split(infer_cmd), shell=False)
assert result.returncode == 0, f"推理失败: {infer_cmd}"

# 阶段4: 反量化模型输出
dequant_cmd = (
    f"{eai_quantize} --eai {eai_model} --quantize 0 --output "
    f"{os.path.join(eai_model_dir, 'output_0.raw')} "
    f"{os.path.join(eai_model_dir, 'output_1.raw')} "
    f"{os.path.join(eai_model_dir, 'output_2.raw')}"  # 修正原始代码的重复行
)
result = subprocess.run(shlex.split(dequant_cmd), shell=False)
assert result.returncode == 0, f"反量化失败: {dequant_cmd}"

# 阶段5: 读取处理结果
results = {}
for i in range(3):
    raw_path = os.path.join(eai_model_dir, f'dequantized_output_{i}.raw')
    results[f'output_{i}'] = np.fromfile(raw_path, dtype=np.float32)

return results

------------------------------

使用示例

------------------------------

if name == "main": # 用户需要配置以下参数 output_data = run_eai_pipeline( eai_model="/path/to/your_model.onnx", prep_data_root="./input_data", eai_model_dir="./workspace", float_input_name="input.bin", float_cached_len_name="seq_len.bin", float_cached_avg_name="avg_values.bin" )

print(f"获取输出张量: {list(output_data.keys())}")
print(f"Output0 shape: {output_data['output_0'].shape}")

# 1.
# encoding的被操作，什么差距，如何生成准确值的encoding，解决encoding，
#
# 2.
# 真实音频量化（不太完整）然后比较相似度，量化的这个需要
#
# 3.
# eai模型推理比较
#
# 4.
# 输入改到40