nncase模型转换后结果出错

重现步骤

设计了一个图像增强模型，onnx模型推理正确，通过onnx转换为kmodel，在模拟器上推理出现错误色块
（可以提供原始onnx模型）

期待结果和实际结果

kmodel应当能够正确输出图像数据

软硬件版本信息

nncase v2.9.0

错误日志

尝试解决过程

补充材料

测试代码：

import os
import numpy as np
from PIL import Image
import matplotlib.pyplot as plt
from nncase_base_func import *
import onnxruntime  # 导入onnxruntime库用于ONNX模型推理

dump_path = "./tmp_onnx"
kmodel_path = "./tmp_onnx/test.kmodel"
onnx_path = "./model_dynamic_sim.onnx"  # ONNX模型路径
npy_path = "input_data.npy"  # 新增的 npy 文件路径

# 添加运行ONNX模型的函数
def run_onnx_model(onnx_path, input_data):
    print(f"Running ONNX model: {onnx_path}")
    # 创建ONNX运行时会话
    session = onnxruntime.InferenceSession(onnx_path)
    
    # 获取输入名称
    input_name = session.get_inputs()[0].name
    
    # 准备输入数据
    input_data = input_data.astype(np.float32) / 255.0  # 归一化到[0, 1]范围
    
    # 运行模型推理
    result = session.run(None, {input_name: input_data})
    
    # 将输出转换回[0, 255]范围
    result_scaled = [r * 255.0 if r.max() <= 1.0 else r for r in result]
    
    return result_scaled

# 加载输入数据
input_data = np.load(npy_path)  # 形状 (1,3, 400, 600)
print(f"Input data shape: {input_data.shape}")

# 确保dump_path目录存在
os.makedirs(dump_path, exist_ok=True)

# 运行ONNX模型
onnx_result = run_onnx_model(onnx_path, input_data)
print("ONNX model inference completed")

# 保存ONNX模型输出
for idx, output in enumerate(onnx_result):
    print(f"ONNX output {idx} shape: {output.shape}")
    output.tofile(os.path.join(dump_path, f"onnx_result_{idx}.bin"))
    
    # 保存为图像
    output_image = output.squeeze().transpose(1, 2, 0).astype(np.uint8)
    output_image = Image.fromarray(output_image)
    output_image.save(os.path.join(dump_path, f"onnx_result_{idx}.png"))
    print(f"Saved ONNX result to {os.path.join(dump_path, f'onnx_result_{idx}.png')}")

# 运行kmodel
print(f"Running kmodel: {kmodel_path}")
result = run_kmodel(kmodel_path, input_data)
print("kmodel inference completed")

# 保存kmodel输出和比较结果
for idx, output in enumerate(result):
    print(f"kmodel output {idx} shape: {output.shape}")
    output.tofile(os.path.join(dump_path, f"nncase_result_{idx}.bin"))
    
    # 保存为图像
    output_image = output.squeeze().transpose(1, 2, 0).astype(np.uint8)
    output_image = Image.fromarray(output_image)
    output_image.save(os.path.join(dump_path, f"nncase_result_{idx}.png"))
    print(f"Saved kmodel result to {os.path.join(dump_path, f'nncase_result_{idx}.png')}")
    
    # 计算与ONNX模型输出的差异
    if idx < len(onnx_result):
        diff = np.abs(output - onnx_result[idx])
        mean_diff = np.mean(diff)
        max_diff = np.max(diff)
        print(f"Output {idx} - Mean difference: {mean_diff:.4f}, Max difference: {max_diff:.4f}")
        
        # 可视化差异
        plt.figure(figsize=(15, 5))
        plt.subplot(1, 3, 1)
        plt.title("ONNX Output")
        plt.imshow(onnx_result[idx].squeeze().transpose(1, 2, 0).astype(np.uint8))
        
        plt.subplot(1, 3, 2)
        plt.title("kmodel Output")
        plt.imshow(output.squeeze().transpose(1, 2, 0).astype(np.uint8))
        
        plt.subplot(1, 3, 3)
        plt.title("Difference (x10)")
        plt.imshow(np.clip(diff.squeeze().transpose(1, 2, 0) * 10, 0, 255).astype(np.uint8))
        
        plt.savefig(os.path.join(dump_path, f"comparison_{idx}.png"))
        print(f"Saved comparison visualization to {os.path.join(dump_path, f'comparison_{idx}.png')}")

模型转换配置：

compile_options = nncase.CompileOptions()
    compile_options.target = "k230" #"cpu"
    compile_options.dump_ir = True  # if False, will not dump the compile-time result.
    compile_options.dump_asm = True
    compile_options.dump_dir = dump_path
    compile_options.input_file = ""
    
    # preprocess args
    compile_options.preprocess = True
    if compile_options.preprocess:
        compile_options.input_type = "uint8" # "uint8" "float32"
        compile_options.input_shape = [1,3,400,600]
        #compile_options.input_shape = [1,400,600,3]
        compile_options.input_range = [0,255]
        compile_options.input_layout = "NCHW" # "NHWC"
        compile_options.swapRB = False
        compile_options.mean = [0,0,0]
        compile_options.std = [255,255,255]
        compile_options.letterbox_value = 0
        compile_options.output_layout = "NCHW" # "NHWC"

    # quantize options
    ptq_options = nncase.PTQTensorOptions()
    ptq_options.quant_type = "uint8" # datatype : "float32", "int8", "int16"
    ptq_options.w_quant_type = "uint8"  # datatype : "float32", "int8", "int16"
    ptq_options.calibrate_method = "NoClip" # "Kld"
    ptq_options.finetune_weights_method = "NoFineTuneWeights"
    ptq_options.dump_quant_error = False
    ptq_options.dump_quant_error_symmetric_for_signed = False

    # mix quantize options
    # more details in docs/MixQuant.md
    ptq_options.quant_scheme = ""
    ptq_options.quant_scheme_strict_mode = False
    ptq_options.export_quant_scheme = False
    ptq_options.export_weight_range_by_channel = False
    ############################################
    
    ptq_options.samples_count = len(calib_data[0])
    ptq_options.set_tensor_data(calib_data)