Accuracy at first few Step

[nguyenminhduc9988]

Hi,

The accuracy at the first 300 step is 0.1000 or 0.0500. Is this normal or is there something wrong?

Thank you

[Tream733]

have you solved the problem？i got a result just like your‘s.it's normal?

[nguyenminhduc9988]

@Tream733 nope, did you leave it to train for a few more iter. Loading the pretrained model seems to work. But I dont understand the logic. It should improve gradually.

[Tream733]

@ducminhnguyen i have finished all the 5000 step.but it's result not very good.i will send the tensorboard picture to your email。i will try to use some different parameter in my follow work 。will you continue do this work？if you continue，can i talk more to you about this work in the following time？

[nguyenminhduc9988]

@Tream733 Yeah, I will continue working on C3D for a while. Thanks in advance for the tensorboard pics. I think the init function of the training process have problem. With the right param (pretrained) this works very well, but if we train from scratch it does not work?

[geroge-gao]

@ducminhnguyen have you ever solved the problem?I met the same problem,and do you know how to use the pre-trained model?

[nguyenminhduc9988]

yeah, I left t over night and the result come out good the next day.

[geroge-gao]

@ducminhnguyen did you train the model from the scratch or by using the pre-trained model provided by hx173149?

[nguyenminhduc9988]

I trained from scratch

[geroge-gao]

I have trained for 7200 steps, but the result is still not very good. the output is as following:

Training Data Eval: accuracy: 1.00000 Validation Data Eval: accuracy: 1.00000

[FENGShuanglang]

@geroge-gao your result is 1.000,Why do you think he is not very good，

[geroge-gao]

@FENGShuanglang because the validation accuracy is only 0.023231257

[FENGShuanglang]

@geroge-gao I know your problem,i got a same issue,you should replace you trained model in predict_c3d_ucf101.py before you start predicting it

[geroge-gao]

@FENGShuanglang I have changed it, the model in predict_c3d_ucf101.py is what I trained myself. but it doesn't work.

[FENGShuanglang]

@geroge-gao mybe you should try reducing the num of training iterations

[geroge-gao]

@FENGShuanglang Actually I train the C3D from scratch，but the accuracy is always 1.0000 .you can see #59

[FENGShuanglang]

you should trian it use pre-trained model ,it is very hard and no-rubustness if you train the C3D from scratch .why do you train it from scratch?by the way ,i have to go to chifan! see you later

[geroge-gao]

@FENGShuanglang so how do you use the pre-trained model? as I know ,the tensorflow model contains four files: .meta .index .data and the checkpoint files. The author only provided the .model file. i guess it corresponds to the .data file. so I use the .model file to cover the .data file simply. But it still doesn't work.

[FENGShuanglang]

I don't understand what you said about "use the .model file to cover the .data file simply",I can tell you what i did ,first you should download the pre-trained model(.model file).And you should replace the .model file in train-c3d-ucf101.py in line 117(Pay attention to the file path) before you start training you network ,and then you can start training .

[geroge-gao]

I did as what you said,but there occurred an error

ValueError: The specified path: conv3d_deepnetA_sport1m_iter_1900000_TF.model is a file. Please specify only the path prefix to the checkpoint files. you can see #10 , as for "use the .model file to cover the .data file simply", actually I first train the C3D from the scratch, when there will be four files in the models foder. then I replace the .data-00000-of-00001 with .model file.

[FENGShuanglang]

You can try to change a pre-training network，for example “sports1m_finetuning_ucf101.model”

[geroge-gao]

I have tried all the three .model files ,but it's still the same, no any change.

[FENGShuanglang]

I don't know why ,but you can reference https://zhuanlan.zhihu.com/p/35388036 and Ask the blogger about your question,he is my classmate

[geroge-gao]

It seems a normal issue on windows, if you train it on Linux, it won't happen. hope it will help you.