V851se[求助] NPU使用问题

jaying LV 5

Q1：network_binary.nb文件太大：37.0 MB (38,898,688 )
参考生成方式：
https://v853.docs.aw-ol.com/npu/npu_yolov3/
xx\ovxilb\yolov3_nbg_unify\network_binary.nb 在851se上溢出了

awwwwa LV 8

需要裁剪优化yolo3模型，或者使用其他模型

jaying LV 5

@awwwwa 大佬请问裁剪优化有什么参考文档或者网址么？我现在换了yolov3-tiny版本实现，xx.nb文件5M左右，但是det_demo使用libawnn.a识别不出坐标值，不确定是为什么（没做过NPU开发，ai相关接触也不多）

煎Bingo子 LV 3

请问有NPU扩展软件包的网盘链接分析吗？非常感谢

jaying LV 5

@liujunjiang 没有压缩包资源我是买板子和店家要的SDK里面带了NPU相关的src

煎Bingo子 LV 3

@jaying 店家给的网盘资料方便分享一下吗？非常感谢！

jaying LV 5

@liujunjiang /openwrt/package/npu
通过网盘分享的文件：npu.7z
链接: https://pan.baidu.com/s/1iDByEzFMrZAXMVbUWSwKSQ?pwd=khxi 提取码: khxi

煎Bingo子 LV 3

@jaying 非常感谢！

jaying LV 5

这一步失败了，不明白为什么....
命令：

pegasus quantize --model yolov3.json --model-data yolov3.data --batch-size 2 --device CPU --with-input-meta yolov3_inputmeta.yml --rebuild --model-quantize yolov3.quantize --quantizer asymmetric_affine --qtype uint8

结果：

..............
D Process convolution_64_147 ...
D Acuity output shape(convolution): (2 76 76 18)
D Tensor @convolution_64_147:out0 type: float32
D Process output_65_148_acuity_mark_perm_152 ...
D Acuity output shape(permute): (2 18 76 76)
D Tensor @output_65_148_acuity_mark_perm_152:out0 type: float32
D Process output_65_148 ...
D Acuity output shape(output): (2 18 76 76)
D Tensor @output_65_148:out0 type: float32
I Build prune_16 complete.
D *********** Setup database (1) ***********
D Setup dataset "TEXT":
D Lids: ['input_0']
D Layouts: ['nchw']
D Shapes: [[2, 3, 608, 608]]
D Data types: ['float32']
D Sparse tensors: []
D Tensor names(H5FS only): []
W:tensorflow:AutoGraph could not transform <acuitylib.core.execution.BackendFunction object at 0x7f8c8c056d90> and will run it as-is.
Please report this to the TensorFlow team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output.
Cause: Unable to locate the source code of <function BackendFunction.__call__ at 0x7f8c8cec69d0>. Note that functions defined in certain environments, like the interactive Python shell do not expose their source code. If that is the case, you should to define them in a .py source file. If you are certain the code is graph-compatible, wrap the call using @tf.autograph.do_not_convert. Original error: could not get source code
To silence this warning, decorate the function with @tf.autograph.experimental.do_not_convert
I Running 1 iterations
D 0(100.00%)
D Calibrating ...
Traceback (most recent call last):
  File "pegasus.py", line 131, in <module>
  File "pegasus.py", line 108, in main
  File "acuitylib/app/medusa/commands.py", line 212, in execute
  File "acuitylib/vsi_nn.py", line 589, in quantize
  File "acuitylib/app/medusa/quantization.py", line 151, in run
  File "acuitylib/app/medusa/quantization.py", line 58, in _run_quantization
  File "acuitylib/quantization/quantize_manager.py", line 516, in calculate_params
  File "acuitylib/quantization/quantize_manager.py", line 272, in _cal_quantize_param_all
  File "acuitylib/quantization/types.py", line 200, in calibrate
  File "acuitylib/quantization/quantizers/asymmetric_quantizer.py", line 102, in convert_from_minmax
  File "acuitylib/quantization/quantizers/asymmetric_quantizer.py", line 161, in compute_from_minmax
  File "acuitylib/quantization/utils.py", line 230, in iscorrect_scalar_minmax_value
ValueError: NAN is detected in minmax (nan,nan)
[4296] Failed to execute script 'pegasus' due to unhandled exception!