The evaluation curve on training is not gradually rize to the values in the paper.

[ff521]

My GPUs are poor, and I set the batchsize to 2, num_workers to 4. The other configs are same to default.
When training the J&F rize to top with 65 epochs, and then it towards to lower. Is it normal？
Training only with DAVIS17, the top J&F only 72. Is it caused by the half batchsize?
Thank you!!

[ff521]

65 epoch is with half LR, and with default config, it's top at 40 epochs.