aihacker111/Efficient-Live-Portrait

My evaluation using mac m2 pro

Opened this issue · 7 comments

x4080 commented

Hi, this is not an issue just an appreciation, after some modifications :

  • @requirements onnxruntime==1.18.1 instead of onnxruntime-gpu==1.18.0
  • @fast_live_portait_pipeline self.providers = ['CoreMLExecutionProvider', 'CPUExecutionProvider']
  • @config.py dir_path = os.path.join(current_dir, './live_portrait_onnx_weights', main_key)

I got [00:31<00:00, 1.96s/it] vs [00:30<00:00, 1.91s/it] using kaggle with P100, not bad

inference : python run_live_portrait.py -v 'experiment_examples/examples/driving/d1.mp4' -i 'experiment_examples/examples/source/s1.jpg'

The memory usage is high though, since I only got 16gb of RAM, but great speed and kaggle onnx GPU seems to work

keep up the good job !

edit : can we optimize the memory usage more ? there's pull request in the main repo about lazy loading (not read it thoroughly yet)

x4080 commented

I found this message 'Context leak detected, msgtracer returned -1' while animating but it then finish the job, what is this message mean ?

aihacker111 commented

@x4080 That’s mean some node im onnx is skip and fall back to CPU , that’s reason why it slower than official

aihacker111 commented

@x4080 Can you create pull request ? I spend time to do TensorRT

x4080 commented

@aihacker111 Sorry, I'm not experienced with pull request, you can add my code if you like

aihacker111 commented

@x4080 never mind, I’ll update later

aihacker111 commented

@x4080 Please accept a invitation, you can manage git and push into repo, don't need pull request

x4080 commented

@aihacker111, I'll try to add the changes