[Bug] GPU utilization is low during training

## Description
Hi, I am running training by using `CUDA_VISIBLE_DEVISES='3' spleeter train -p ... -d ...`. Here I use spleeter-gpu. Now it has run for 1 day. I checked the nvidia-smi, and found the memory was all taken up, but the utilization was 0% (very occationally, it would go to some number). Could anyone tell me whether this is correct or wrong?

## Step to reproduce

`conda install -c conda-forge spleeter-gpu`
`pip install -r requirements.txt`
`pip install tensorflow-gpu-1.15.2`
`CUDA_VISIBLE_DEVISES='3' spleeter train -p ... -d ...`

## Output
+-------------------------------+----------------------+----------------------+
| 3 Tesla V100-SXM2... On | 00000000:88:00.0 Off | 0 |
| N/A 38C P0 57W / 300W | 31406MiB / 32510MiB | 0% Default |
+-------------------------------+----------------------+----------------------+

Start by reproducing the listed conda, pip, and `spleeter train` commands while monitoring the Tesla V100 with `nvidia-smi`. No source file or test is identified, so trace the training entry point and GPU execution path to determine whether the low utilization is expected; done means explaining the observed memory and utilization behavior or identifying a confirmed defect.

Written by the indexing model from the issue text.