yjxiong
commented
8 years ago Your understanding is correct.
The accuracy should evaluated by using the provided scripts on trained models. The number you see during training is single frame single crop accuracy, which is just for keeping track of the training process.
On Tuesday, August 30, 2016, praneetha111 notifications@github.com wrote:
Hi,
Thanks for sharing the code here. It is very helpful.
I am running training on very deep two stream convnets for action recognition on UCF-101 dataset. I am trying to get the performance of 91.4% as mentioned in the paper. From the earlier questions asked here, I understand that the performance number is average of the performance on all the classes measured using action_python.
Could you please tell me if my understanding is correct. Also, could you tell me what is the significance of accuracy that I see during the training. I am seeing an accuracy of around 0.69 at the end of 30,000 iterations, what does this exactly mean? Is 0.69 the expected number here?
Last few iteration log: I0829 19:50:45.809867 17938 solver.cpp:240] Iteration 29960, loss = 0.779676 I0829 19:50:45.809898 17938 solver.cpp:631] Iteration 29960, lr = 5e-05 I0829 19:51:05.940994 17938 solver.cpp:240] Iteration 29980, loss = 0.632111 I0829 19:51:05.941025 17938 solver.cpp:631] Iteration 29980, lr = 5e-05 I0829 19:51:46.307701 17938 solver.cpp:502] Snapshotting to cuhk_action_recognition_16_split1_flow_iter_30000.caffemodel I0829 19:52:09.980613 17938 solver.cpp:510] Snapshotting solver state to cuhk_action_recognition_16_split1_flow_iter_30000.solverstate I0829 19:52:10.869324 17938 solver.cpp:406] Iteration 30000, loss = 0.785312 I0829 19:52:10.869350 17938 solver.cpp:424] Iteration 30000, Testing net (#0) Could not create logging file: No such file or directory COULD NOT CREATE A LOGGINGFILE 20160829-195224.17938!I0829 19:52:24.443380 17938 solver.cpp:481] Test net output #0: accuracy = 0.692868 I0829 19:52:24.443433 17938 solver.cpp:481] Test net output #1 https://github.com/yjxiong/caffe/pull/1: loss = 1.18713 (* 1 = 1.18713 loss)
Thanks.
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/yjxiong/caffe/issues/102, or mute the thread https://github.com/notifications/unsubscribe-auth/AGg4d0fVGwaIkYLHCCmoUB6Momml8D-1ks5qk_qGgaJpZM4JwVYN .
Best regards,
Yuanjun
praneetha111
Hi,
Thanks for sharing the code here. It is very helpful.
I am running training on very deep two stream convnets for action recognition on UCF-101 dataset. I am trying to get the performance of 91.4% as mentioned in the paper. From the earlier questions asked here, I understand that the performance number is average of the performance on all the classes measured using action_python.
Could you please tell me if my understanding is correct. Also, could you tell me what is the significance of accuracy that I see during the training. I am seeing an accuracy of around 0.69 at the end of 30,000 iterations, what does this exactly mean? Is 0.69 the expected number here?
Last few iteration log: I0829 19:50:45.809867 17938 solver.cpp:240] Iteration 29960, loss = 0.779676 I0829 19:50:45.809898 17938 solver.cpp:631] Iteration 29960, lr = 5e-05 I0829 19:51:05.940994 17938 solver.cpp:240] Iteration 29980, loss = 0.632111 I0829 19:51:05.941025 17938 solver.cpp:631] Iteration 29980, lr = 5e-05 I0829 19:51:46.307701 17938 solver.cpp:502] Snapshotting to cuhk_action_recognition_16_split1_flow_iter_30000.caffemodel I0829 19:52:09.980613 17938 solver.cpp:510] Snapshotting solver state to cuhk_action_recognition_16_split1_flow_iter_30000.solverstate I0829 19:52:10.869324 17938 solver.cpp:406] Iteration 30000, loss = 0.785312 I0829 19:52:10.869350 17938 solver.cpp:424] Iteration 30000, Testing net (#0) Could not create logging file: No such file or directory COULD NOT CREATE A LOGGINGFILE 20160829-195224.17938!I0829 19:52:24.443380 17938 solver.cpp:481] Test net output #0: accuracy = 0.692868 I0829 19:52:24.443433 17938 solver.cpp:481] Test net output #1: loss = 1.18713 (* 1 = 1.18713 loss)
Thanks.