tflite only python package

[PhilipXue]

Hi, I appreciate a lot for your effort! There are ways to build tflite only python package which will reduce the package size tremendously: https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/tools/pip_package
I tied it before and I don't think this runtime can utilize multiprocessing. Is there any chance that you can check this and build a tflite package that has multiprocessing ability?

[PINTO0309]

@PhilipXue
I tried to create it, so please try it. I have not verified the behavior yet. Please let me know if it worked.
https://github.com/PINTO0309/TensorflowLite-bin.git

[PhilipXue]

@PINTO0309 Wow that's amazing! I'll try it immediately!

[PhilipXue]

@PINTO0309
Hi,
I tied you the whl you provide today and here are some findings:

• The Interpretor class doesn't have set_num_threads attribute.
• but the runtime seems to be using 4 threads
  
• The speed of running mobilenet ssdlite model is slower than the full-sized tensorflow runtime with 4 threads (650 ms/img vs 600 ms/img), and adding thread with full-sized tensorflow runtime can improve the speed to 520 ms/img.

Does the tflite-runtime package include the improvment you did to tensorflow build?

[PINTO0309]

@PhilipXue
I'm sorry. I noticed a mistake in correcting the program. I will fix it today. please wait a moment. I have identified the cause.

[PINTO0309]

@PhilipXue
Fixed and recommitted.
tflite_runtime-1.14.0-cp35-cp35m-linux_armv7l.whl
tflite_runtime-1.14.0-cp37-cp37m-linux_armv7l.whl

https://github.com/PINTO0309/TensorflowLite-bin.git

[PhilipXue]

@PINTO0309
Thanks for your swift response!
I tried out the new whl file and it works like a charm (multithread performance improved over full-sized tf runtime)! Thanks a lot for your effort!
BTW, Do you have any plan to build tflr whls for 64-bit debian buster (python3.7 aarch64)?

[PINTO0309]

@PhilipXue
It is complete. The working time is about 5 minutes. However, I don't have a device that can be tested.
tflite_runtime-1.14.0-cp35-cp35m-linux_aarch64.whl
tflite_runtime-1.14.0-cp37-cp37m-linux_aarch64.whl

https://github.com/PINTO0309/TensorflowLite-bin/tree/master/1.14.0

[PhilipXue]

NP!
I'll test and give you feedbacks!

[PhilipXue]

Hi,
I tested the aarch64 whl on 64-bit debian buster preview image of rpi3, and this time it behaves similarly to the 32-bit version before the fixing:

@PINTO0309
Hi,
I tied you the whl you provide today and here are some findings:

• The Interpretor class doesn't have set_num_threads attribute.
• but the runtime seems to be using 4 threads
  
• The speed of running mobilenet ssdlite model is slower than the full-sized tensorflow runtime with 4 threads (650 ms/img vs 600 ms/img), and adding thread with full-sized tensorflow runtime can improve the speed to 520 ms/img.

Does the tflite-runtime package include the improvment you did to tensorflow build?

• No set_num_threads attribute:AttributeError: type object 'Interpreter' has no attribute 'set_num_threads'
• (4 threads) performance is pretty poor (1.3x slower compare to 32-bit version before fix). I'm not sure if this is OS problem since the OS is a preview version.

Where do you think the problem is?

[PINTO0309]

@PhilipXue
Ah ... sorry. Perhaps I have repeated the same mistake. I can't work until I get home, so please be patient.

[PhilipXue]

@PINTO0309
No need to be sorry.
I heard that 64-bit system is better at floating-point calculation and just curious to verify it. Please take your time.

[PINTO0309]

@PhilipXue
I checked the contents of the wheel file. After all, I made the same mistake, so I modified the wheel file.
https://github.com/PINTO0309/TensorflowLite-bin.git
tflite_runtime-1.14.0-cp35-cp35m-linux_aarch64.whl
tflite_runtime-1.14.0-cp37-cp37m-linux_aarch64.whl

By the way, there is also a 64-bit kernel OS image that I created independently.
https://github.com/PINTO0309/RaspberryPi-bin.git

The benchmark tool you are using is cool. Is it possible for me to share it?

[PhilipXue]

@PINTO0309
Hi, I will check when I go back to work on Monday!

The benchmark is very simple, it measures the processing time for each image. Through simple, the code is created for internal projects. You can use code provided in this article to do the exact same thing. (If you are referring the image that I put in the comment, it's htop)
I can share the results measuring some publicly available models (mobilenet ssd/ssdlite), which should be enough to see the improvement made by your whl files.

[PINTO0309]

@PhilipXue
Thank you! What I wanted to know was htop.
It's so cool. 😃

[PhilipXue]

I'm glad you enjoyed it! If you are interested, there is another similar system monitoring tool called glances that you may want to try out.

Have a nice day!