Stuck on an issue?

Lightrun Answers was designed to reduce the constant googling that comes with debugging 3rd party libraries. It collects links to all the places you might be looking at while hunting down a tough bug.

And, if you’re still stuck at the end, we’re happy to hop on a call to see how we can help out.

Do tfjs-posenet model and tflite-posenet model have the same weights?

See original GitHub issue

I am working on build an android app based on tflite-posenet model which comes from https://github.com/tensorflow/examples/tree/master/lite/examples/posenet/android.

However, after I compared the performance of tflite model with the performance of tfjs posenet demo (both of them use mobilenet v1_100 as backend), I found the tfjs demo is much more accurate.

Do they use same weights?

Issue Analytics

State:
Created 4 years ago
Comments:6 (1 by maintainers)

Top GitHub Comments

2reactions

tylerzhu-githubcommented, Jan 13, 2020

Hi @darcula1993, @emredog,

Thx for your inquery! Currently both the TFLite PoseNet models and the latest TF.js PoseNet models are converted from the same checkpoints (with batch norm variables folded into the preceding convolution layer). The differences that could cause the regression is that TF.js models supports different output stride e.g. 8, 16, 32 and configurable input resolutions, which can affect the accuracy. Let me know if this answers your questions and look forward to hearing more about your progress on building Keras PoseNet. @dsmilkov feel free to also include the main contacts of the TFLite Android PoseNet app and iOS PoseNet app, they will know more about the details.

Best, Tyler

0reactions

rthadurcommented, Jan 16, 2020

thank you closing this issue , feel free to reopen if you need more info.