OE版本:1.74
目前检测框位置没有问题,上次发帖发现单张推理没有问题是因为对比的calibration的模型和origin模型的效果。这次改正回来quanti模型发现精度下降非常大,并且检测数量也变少。已经改为int16的量化模式。
图中只剩了两个框,并且精度也很低。原始检测应该是10多个框。
您好!具体是哪个模型呢?能贴一下量化的log文件和yaml文件吗?意思是 calibration 的结果较好, quanti 的结果比较差?
model_parameters: onnx_model: './original_model/yolo11x.onnx' march: "bayes" layer_out_dump: False working_dir: 'model_output' output_model_file_prefix: 'yolo11_640x640_nv12_a1' remove_node_type: "Dequantize" input_parameters: input_name: "" input_type_rt: 'rgb' input_layout_rt: 'NCHW' input_type_train: 'rgb' input_layout_train: 'NCHW' input_shape: '' norm_type: 'data_scale' mean_value: '' scale_value: 0.003921568627451 calibration_parameters: cal_data_dir: './calibration_data_rgb_f32' cal_data_type: 'float32' calibration_type: 'default' optimization: 'set_all_nodes_int16' run_on_bpu: /model.10/m/m.0/attn/Softmax;/model.10/m/m.1/attn/Softmax;/model.23/dfl/Softmax compiler_parameters: compile_mode: 'latency' debug: False optimize_level: 'O2'
2026-01-07 17:32:55,851 INFO The quantify model output: ===================================================================================== Node Cosine Similarity L1 Distance L2 Distance Chebyshev Distance ------------------------------------------------------------------------------------- /model.23/Concat_5 0.999625 0.172588 0.001960 326.932922 2026-01-07 17:32:55,860 INFO End to Horizon NN Model Convert. 2026-01-07 17:32:56,355 WARNING node: /model.10/m/m.0/attn/Softmax does not exist, please double check your input 2026-01-07 17:32:56,355 WARNING node: /model.10/m/m.1/attn/Softmax does not exist, please double check your input 2026-01-07 17:32:56,355 WARNING node: /model.23/dfl/Softmax does not exist, please double check your input
我参考了你发布的j5 yolo11n部署实战,用你的方式这个帖子遇到的问题没有了,这是quanti模型得到的精度(前三个):
2026-01-09 17:25:07,436 INFO person is in the picture with confidence: 0.7754
2026-01-09 17:25:07,437 INFO kite is in the picture with confidence: 0.7624
2026-01-09 17:25:07,437 INFO person is in the picture with confidence: 0.7578
这是原始的yolo11n的精度(前三个):
2026-01-09 17:22:57,351 INFO kite is in the picture with confidence: 0.8586
2026-01-09 17:22:57,351 INFO kite is in the picture with confidence: 0.8417
2026-01-09 17:22:57,351 INFO person is in the picture with confidence: 0.8197
请问这个精度下降是不是合理范围内的。