We were unable to obtain the tflite benchmark of 137 ms on Adreno 530 GPU (One Plus 3 phone ), using the same model in the app( also using our own modified version of official tensorflow gpu classification app ).It comes around 450-500 ms on this device.How did you measure the timing? Is it the time required for 'tflite.run' function, as shown in the app?Or does it have something to do with the phone hardware itself (although specs seem to be same)?
Also, is there any other way to improve the performance, say if we want to directly segment camera frames instead of segmenting local images?The tensorflow site suggests us to use opengl SSBO direclty to avoid io buffers related to CPU-GPU copying.Another tip was to use RGBA (4 channel) frames from camera as input to model.Any idea about how to implement these techniques?
We were unable to obtain the tflite benchmark of 137 ms on Adreno 530 GPU (One Plus 3 phone ), using the same model in the app( also using our own modified version of official tensorflow gpu classification app ).It comes around 450-500 ms on this device.How did you measure the timing? Is it the time required for 'tflite.run' function, as shown in the app?Or does it have something to do with the phone hardware itself (although specs seem to be same)?
Also, is there any other way to improve the performance, say if we want to directly segment camera frames instead of segmenting local images?The tensorflow site suggests us to use opengl SSBO direclty to avoid io buffers related to CPU-GPU copying.Another tip was to use RGBA (4 channel) frames from camera as input to model.Any idea about how to implement these techniques?