train on youtube-vos

[colorblank]

Grear work! Thanks for sharing your code!

I use the default training configure of ytvos to train the network.
But I only got best performance J&F=65.5 at epoch 490.
Is the default configure supposed to have this performance?
How can I get the best performance like your provided checkpoint?
I'd appreciate it if you could point out what I did wrong.

[arnike]

Hi @colorblank,
thanks! Are you referring to "0.655" in the snapshot name, or did you actually use infer_vos.sh script to evaluate on the whole validation set? Please note that we save the snapshot based on only 5 sequences from DAVIS-2017 during training, so the accuracy on the complete validation set will be different and tends to be higher.
Best,
Nikita

[colorblank]

My bad. I tried to the test code and got th results like:

J&F-Mean   J-Mean  J-Recall  J-Decay   F-Mean  F-Recall  F-Decay
 0.680324 0.658446  0.789883 0.182041 0.702202  0.835232 0.258544

Thanks for your clarification.
Besides, there is a small bug in this line :

dense-ulearn-vos/infer_vos.py

Line 327 in 9e39d35

model = nn.DataParallel(model).cuda()

. You should change to single gpu mode like "model = model.cuda()".

[arnike]

Thanks for the pointer about the inference code.

There may be some deviation in accuracy between the training runs, albeit your results are well below what I typically observed. You can try simply re-runing the training and/or decreasing the interval between the validation runs (check LOG.ITER_VAL in the YAML config).

I've now also experimented with decaying the learning rate for training on YouTube-VOS to make it more reproducible. My only change to the default setup was in config/ytvos.yaml:

MODEL.LR_SCHEDULER: "step"
MODEL.LR_GAMMA: 0.5
MODEL.LR_STEP: 100

This is the result I got on the complete DAVIS-2017 val:

--------------------------- Global results for val ---------------------------
 J&F-Mean    J-Mean  J-Recall   J-Decay    F-Mean  F-Recall   F-Decay
 0.694455  0.672305  0.818764  0.164835  0.716605  0.843103  0.239955

(The maximum on the validation subset we use in training for this snapshot was 64.5% J&F reached at epoch 330.)
Let me know if that works for you.
Best regards,
Nikita