使用paddle2onnx将量化后的ppyoloe_plus_s导出为onnx后，onnxruntime推理结果与paddle静态图推理不符

[EvW1998]

问题描述

根据PaddleDetection提供的模型压缩文档，我将训练好的ppyoloe plus s模型进行量化训练。量化后的模型，以及导出为paddle静态图模型均推理无误。

但使用paddle2onnx将量化后的模型导出为onnx后，使用onnxruntime推理，无法得到任何检测框（使用paddle静态图推理可得到多个）

使用环境

使用的环境如下：
paddle2onnx == 1.2.11
paddledet == develop (>= 2.8.0)
paddlepaddle-gpu == 2.6.1.post116
paddleslim == 2.6.0
onnxruntime == 1.19.2
onnxruntime-gpu == 1.19.2

paddle静态模型及onnx文件

ppyoloe_s_qat_abnormal_object.zip

错误详情

原图：

paddle静态图推理效果：

在使用以下命令，导出为onnx文件

paddle2onnx --model_dir /mnt/data01/wzj/code/PaddleDetection-develop/export/ppyoloe_s_data1118_quant_lr0000125/ppyoloe_s_qat_abnormal_object \
            --model_filename model.pdmodel \
            --params_filename model.pdiparams \
            --save_file /mnt/data01/wzj/code/PaddleDetection-develop/export/ppyoloe_s_data1118_quant_lr0000125/ppyoloe_s_qat_abnormal_object/ppyoloe_s_qat_241118.onnx

输出结果：

[Paddle2ONNX] Start to parse PaddlePaddle model...
[Paddle2ONNX] Model file path: /mnt/data01/wzj/code/PaddleDetection-develop/export/ppyoloe_s_data1118_quant_lr0000125/ppyoloe_s_qat_abnormal_object/model.pdmodel
[Paddle2ONNX] Parameters file path: /mnt/data01/wzj/code/PaddleDetection-develop/export/ppyoloe_s_data1118_quant_lr0000125/ppyoloe_s_qat_abnormal_object/model.pdiparams
[Paddle2ONNX] Start to parsing Paddle model...
[Paddle2ONNX] [Info] The Paddle model is a quantized model. 
[Paddle2ONNX] [reduce_mean: mean_0.tmp_0] Requires the minimal opset version of 11.
[Paddle2ONNX] [reduce_mean: mean_1.tmp_0] Requires the minimal opset version of 11.
[Paddle2ONNX] [reduce_mean: mean_2.tmp_0] Requires the minimal opset version of 11.
[Paddle2ONNX] [reduce_mean: mean_3.tmp_0] Requires the minimal opset version of 11.
[Paddle2ONNX] [multiclass_nms3: multiclass_nms3_0.tmp_1] Requires the minimal opset version of 10.
[Paddle2ONNX] Due to the operator: dequantize_linear, requires opset_version >= 13.
[Paddle2ONNX] Opset version will change to 13 from 9
[Paddle2ONNX] Use opset_version = 13 for ONNX export.
[WARN][Paddle2ONNX] [multiclass_nms3: multiclass_nms3_0.tmp_1] [WARNING] Due to the operator multiclass_nms3, the exported ONNX model will only supports inference with input batch_size == 1.
[Paddle2ONNX] Deploy backend is: onnxruntime
[Paddle2ONNX] PaddlePaddle model is exported as ONNX format now.

然而使用onnxruntime对导出的onnx文件推理，结果中没有任何检测框。请问是将量化后的模型导出时，哪里有精度丢失问题吗？非量化的模型按照此流程导出没有任何问题

[Zheng-Bicheng]

输入对齐过了嘛？

[EvW1998]

输入对齐了。图片尺寸均缩放为640，640，图片的预处理也一致

ppdet中的config为NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}

onnx的输入也以此处理了

frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB).astype(np.float32)
input_onnx = np.transpose(np.array([(frame-0)/255]), (0,3,1,2)) # n,c,h,w

[Zheng-Bicheng]

我的意思是，您应该把相同的数据同时用 PaddleInference 和 ONNXRuntime 推理一下，对比一下结果，抛开后处理

[EvW1998]

谢谢回复，我明白您的意思了。但是ppyoloe plus模型，本身带有nms后处理，输出为直接的检测框结果和检测框个数。

您说抛开后处理，是指对比nms之前一个node的结果吗？

[Zheng-Bicheng]

是的，ONNXRuntime的NMS和Paddle静态图的NMS在实际效果上不一定等价的，有区别

[EvW1998]

好的，我找到了nms的位置，会尝试对比这两个节点

想请教一下paddle与onnx有逐层比较精度的工具吗？或者说paddle静态图的裁剪工具

[Zheng-Bicheng]

这个没有，所以让您对比NMS前的两个节点

[Zheng-Bicheng]

一般来说量化后的精度差距1%内，量化前的差距0.1的-5次方内

[EvW1998]

您好，通过实验对比nms前的两个节点，量化后的精度差距比较大，对比结果：

Input shape: (1, 3, 640, 640) mean: 0.45289892 max: 1.0 min:  0.08627451
Paddle output0 shape: (1, 8400, 4) mean: 320.46814 max: 831.07166 min:  -280.5739
Paddle output1 shape: (1, 3, 8400) mean: 0.0025582889 max: 0.95241463 min:  1.0192251e-05
Onnx output0 shape: (1, 8400, 4) mean: 320.3321 max: 876.1374 min:  -249.33652
Onnx output1 shape: (1, 3, 8400) mean: 0.001652396 max: 0.007899523 min:  0.00082200766
Output0 mean cosine distance:  0.011558139 mean euclidean distance 75.20233
Output1 mean cosine distance:  0.6552932 mean euclidean distance 1.5684191

量化前的精度差距基本没有：

Input shape: (1, 3, 640, 640) mean: 0.45289892 max: 1.0 min:  0.08627451
Paddle output0 shape: (1, 8400, 4) mean: 320.18158 max: 815.7729 min:  -231.14545
Paddle output1 shape: (1, 3, 8400) mean: 0.0025046652 max: 0.94339174 min:  4.4206395e-06
Onnx output0 shape: (1, 8400, 4) mean: 320.18158 max: 815.7726 min:  -231.14482
Onnx output1 shape: (1, 3, 8400) mean: 0.002504666 max: 0.9433913 min:  4.440546e-06
Output0 mean cosine distance:  -7.947286e-09 mean euclidean distance 0.00020155673
Output1 mean cosine distance:  -7.947286e-08 mean euclidean distance 5.4948055e-06

还是可以确定是paddle2onnx过程中有精度损失。

我是通过在PaddleDetection动转静时加入exclude_nms命令去除nms节点的，静态图和onnx文件地址链接: https://pan.baidu.com/s/1wu8_b3gGsMCmA4rlZ73XVw 提取码: emb4

对比代码：

import paddle
import numpy as np
import cv2
import onnxruntime as rt

def cosine_distance(a, b):
    dot_product = np.sum(a * b, axis=-1)

    norm_a = np.linalg.norm(a, axis=-1)
    norm_b = np.linalg.norm(b, axis=-1)
    
    cosine_similarity = dot_product / (norm_a * norm_b)
    
    cosine_distance = 1 - cosine_similarity
    
    return np.mean(cosine_distance)


def euclidean_distance(a, b):
    diff = a - b
    
    distances = np.linalg.norm(diff, axis=-1)
    
    mean_distance = np.mean(distances)
    
    return mean_distance


#-----------------paddle推理---------------------

paddle.enable_static()

# 加载模型
# [inference_program, feed_target_names, fetch_targets] = paddle.static.load_inference_model(
#     path_prefix='./compare/ppyoloes_exclude_nms/ppyoloe_plus_crn_s_80e_abnormal_object',
#     model_filename='model.pdmodel',
#     params_filename='model.pdiparams',
#     executor=paddle.static.Executor()
# )

[inference_program, feed_target_names, fetch_targets] = paddle.static.load_inference_model(
    path_prefix='./compare/ppyoloes_qat_exclude_nms/ppyoloe_s_qat_abnormal_object',
    model_filename='model.pdmodel',
    params_filename='model.pdiparams',
    executor=paddle.static.Executor()
)

# 准备输入数据
frame = cv2.imread("./yiwu.png")
original_im = frame.copy()

frame = cv2.resize(frame, (640, 640))
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB).astype(np.float32)
input_image = np.transpose(np.array([(frame-0)/255]), (0,3,1,2)) # n,c,h,w

print("Input shape:", input_image.shape, "mean:", input_image.mean(), "max:", input_image.max(), "min: ", input_image.min())

input_scale = np.array([[1, 1]], dtype="float32")

# 创建执行器
place = paddle.CUDAPlace(0)
exe = paddle.static.Executor(place)

# 执行推理
results = exe.run(program=inference_program,
                  feed={feed_target_names[0]: input_image, feed_target_names[1]: input_scale},
                  fetch_list=fetch_targets)

# cv2.imwrite("./compare/yiwu.png", vis_detection(original_im, results[0]))
 
print("Paddle output0 shape:", results[0].shape, "mean:", results[0].mean(), "max:", results[0].max(), "min: ", results[0].min())
print("Paddle output1 shape:", results[1].shape, "mean:", results[1].mean(), "max:", results[1].max(), "min: ", results[1].min())

#-----------------onnx推理---------------------

# sess = rt.InferenceSession("./compare/ppyoloes_exclude_nms.onnx", providers=['CPUExecutionProvider'])
sess = rt.InferenceSession("./compare/ppyoloes_qat_exclude_nms.onnx", providers=['CPUExecutionProvider'])
output_names = [sess.get_outputs()[0].name, sess.get_outputs()[1].name]
onnx_pred = sess.run(output_names, {sess.get_inputs()[0].name: input_image, sess.get_inputs()[1].name: input_scale})

print("Onnx output0 shape:", onnx_pred[0].shape, "mean:", onnx_pred[0].mean(), "max:", onnx_pred[0].max(), "min: ", onnx_pred[0].min())
print("Onnx output1 shape:", onnx_pred[1].shape, "mean:", onnx_pred[1].mean(), "max:", onnx_pred[1].max(), "min: ", onnx_pred[1].min())

print("Output0 mean cosine distance: ", cosine_distance(results[0], onnx_pred[0]), "mean euclidean distance", euclidean_distance(results[0], onnx_pred[0]))
print("Output1 mean cosine distance: ", cosine_distance(results[1], onnx_pred[1]), "mean euclidean distance", euclidean_distance(results[1], onnx_pred[1]))