Open yangxuerong opened 4 years ago
Hi, Could you tell me about the config of training? My loss falls slowly. Almost one epoch went down by only 0.1. I think this is causing by lr,but I changed it and found no improving. Thank you.
I have the same problem as you. Have you solved it? I sincerely hope to receive your reply, thank you very much!
Any video classification result cannot be reproduced, which is a very bad problem. I suspect that the inability to train from 0 to a perfect result is false.
Hi , My own dataset is a similar dataset to Kinetics,When I train ten epoch,i get the result as follow:
[INFO: logging.py: 96]: json_stats: {"_type": "train_iter", "epoch": "10/196", "eta": "12 days, 12:50:42", "gpu_mem": "5.37 GB", "iter": "1870/6178", "loss": 0.056469, "lr": 0.032643, " time_diff": 0.938987, "top1_err": 3.125000, "top5_err": 0.000000}
So I stop to train ,use the "./checkpoints/checkpoint_epoch_00011.pyth" to test the accuracy, i get the result as follows:
[INFO: logging.py: 96]: json_stats: {"split": "test_final", "top1_acc": "42.50", "top5_acc": "76.50"} [06/17 11:34:28][INFO] slowfast.utils.logging: 96: json_stats: {"split": "test_final", "top1_acc": "42.50", "top5_acc": "76.50"}
why? Do you know the reason?