Possible GPU Memory Leak?

[ProfFan]

The notebook cannot be run with GPU S4TF because of constant OOM when calculating the loss. It appears that GPU memory is filled after

let data = BeeVideo(videoName: "bee_video_1")!
print(data.frames.count)
print(data.tracks.count)

> nvidia-smi
Wed Sep 30 11:00:26 2020
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 455.23.04    Driver Version: 455.23.04    CUDA Version: 11.1     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  GeForce GTX 1080    Off  | 00000000:01:00.0  On |                  N/A |
| 28%   54C    P2    53W / 180W |   7540MiB /  8119MiB |      1%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|    0   N/A  N/A      1121      G   /usr/lib/Xorg                     102MiB |
|    0   N/A  N/A      2609      G   /usr/lib/Xorg                     545MiB |
|    0   N/A  N/A      3261      G   ...akonadi_archivemail_agent        2MiB |
|    0   N/A  N/A      3264      G   .../bin/akonadi_ews_resource       14MiB |
|    0   N/A  N/A      3268      G   ...bin/akonadi_imap_resource        2MiB |
|    0   N/A  N/A      3279      G   .../akonadi_mailfilter_agent        2MiB |
|    0   N/A  N/A      3288      G   ...n/akonadi_sendlater_agent        2MiB |
|    0   N/A  N/A      3291      G   ...nadi_unifiedmailbox_agent        2MiB |
|    0   N/A  N/A      3879      G   ...AAAAAAAAA= --shared-files      309MiB |
|    0   N/A  N/A      4179      G   /usr/bin/krunner                   11MiB |
|    0   N/A  N/A      4285      G   /usr/bin/alacritty                 11MiB |
|    0   N/A  N/A    349438      G   /usr/bin/plasmashell              160MiB |
|    0   N/A  N/A   3385747      G   ...AAAAAAAAA= --shared-files       39MiB |
|    0   N/A  N/A   4011142      C   ...lchain/usr/bin/repl_swift     6313MiB |
+-----------------------------------------------------------------------------+

However when I look at the device of the Tensors, they are all on the CPU:

print(beeBatch.device)
print(data.frames[0].device)

For now I'll just use the CPU-only version of S4TF instead as a workaround.

[ProfFan]

Closing as this is upstream issue

tensorflow/swift-apis#1089
tensorflow/swift-apis#1094