Duration (sec/step) comparison for mimic2 training (ljspeech)

Hey there.

Since playing around with mimic2 for a few weeks and recording my german voice with mimic-recording-studio for cc0-contribution i played with two different hardware setups.
Since these may have difference in computing time per step i start this thread with my information on computing time.

1st setup (running virtual ubuntu 18.04.3 lts server on qnap nas virt. station):

hw: QNAP TS 451+ (Intel celeron J1900 2GHz 4 Core with totally 8gb ram)
vm provided: 2 virt. cpu, 6gb memory
Tacotron/mimic2 installed manually (not running as docker container)
Training took about 21,5sec/step. Nas not really busy by other processes.

2nd setup (Running on dedicated desktop computer):

hw: Workstation PC (Intel® Core™ i5-3550 CPU @ 3.30GHz × 4 with totally 16GB RAM, MSI 4GB D5 GTX 1050 Ti Gaming X 4G gpu, 500gb ssd disc)
os: Ubuntu 18.04.3 lts desktop
Tacotron/mimic2 installed manually (not running as docker container) using cpu not nvidia-docker (gpu usage)
Training took about 2,9sec/step.

3rd setup (todo: Using hw with nvidia-docker - gpu usage)

Does this work with a non nvidia gpu?

4th setup (todo: Running as docker container)


Hope this is helpful for others to get an idea of needed computing time.


#3: no, not without significant changes to code.

#4 will depend on the hardware it runs on.

Thanks for your feedback. So no need to test no. 3 for me.

No. 4 would be tested on identical hardware. Just running as container insted of tensorflow/mimic installed manually.