search

deepjavalibrary / djl

An Engine-Agnostic Deep Learning Framework in Java
https://djl.ai
Apache License 2.0
4.14k stars 658 forks source link

about speed slowly for image convert to NDArray #2585

Closed guojian0614 closed 1 year ago

guojian0614 commented 1 year ago

Description

(A clear and concise description of what the feature is.)

Will this change the current api? How?

Who will benefit from this enhancement?

References

image

guojian0614 commented 1 year ago

image

could't make it faster ??? now python do same thing takes few.

frankfliu commented 1 year ago

@guojian0614

In theory, OpenCV should not be slower than PIL (https://towardsdatascience.com/image-processing-opencv-vs-pil-a26e9923cdf3) iff both doing identical operation. I guess your image processing might skipped some operation.

guojian0614 commented 1 year ago

@guojian0614

In theory, OpenCV should not be slower than PIL (https://towardsdatascience.com/image-processing-opencv-vs-pil-a26e9923cdf3) iff both doing identical operation. I guess your image processing might skipped some operation.

@frankfliu

ok，thanks，i will pay attention to inference step go on ，maybe the docker serving of tensorflow officail provide is too faster。but the whole process presstsure test result comparation gap indeed too huge。

python project tps 22-23 ， DJL tps 11-12.

DJL just inference step without pre/post process and other opertions， take time between 50-70ms。but i can't get takes time of python inference step .

I am still not sure my custom DJL project indeed used gpu to inference so far .