Are there any plans to release TensorFlow v1.x checkpoints (.ckpt, .meta, .index)?
Closed this issue · 4 comments
Hello, Author.
Thank you for publishing this wonderful model. I am amazed at how well this model performs.
test3_338x640_.mp4
Also, using the frozen_model.pb you have made available, I have generated an optimized ONNX and made it available in my model zoo, citing the GPLv2 license.
https://github.com/PINTO0309/PINTO_model_zoo/tree/main/333_E2Pose
It is my understanding that this model is optimized and trained for [N,512,512,3] input resolutions, but I would like to generate optimized models for other resolutions if possible. I would be very happy to make use of this model, especially in the field of HCI research.
Therefore, I would be very happy if you plan to publish the checkpoints generated by TensorFlow v1.x. Because I have the skills to transform the checkpoints (.ckpt, .meta, .index) on my own to optimize the ONNX. I would like to generate my own Protocl Buffers file with input resolution redefined as [N,H,W,3] and use it in my research.
If you are not very interested, please ignore this issue.
Sorry to bother you.
I understood that variable resolution checkpoint's are very difficult. The accuracy and speed at close range with a resolution of 512x512 is very good, so I am withdrawing this issue.
I will try to train the model on my own to generate checkpoints of different resolutions when I need to do so.
Thank you very much.
Thank you for your interest in E2Pose.
As you understand, the current model cannot support flexible resolution input.
However, we have trained models that include a variety of fixed resolution inputs.
We will publish these in soon, so please be patient.
Thank you! I am looking forward to it.
I have optimized all the models you additionally committed with post-processed ONNX and committed them to my repository along with the demo code. 😸
https://github.com/PINTO0309/PINTO_model_zoo/tree/main/333_E2Pose