search

jfzhang95 / pytorch-video-recognition

PyTorch implemented C3D, R3D, R2Plus1D models for video activity recognition.
MIT License
1.18k stars 250 forks source link

Decreasing Test Accuracy in README .png #13

Open Robert-Browning opened 5 years ago

Robert-Browning commented 5 years ago

In your README, the tensorboardx output for Test Acc is steadily decreasing over all 100 epochs. Is that just a 'typo' or were those your actual results?

image

PS Thanks a ton for the code. It has been very helpful!

wave-transmitter commented 5 years ago

Hi, pay attention to the y-axis of the first figure, values are between 0.962 and 0.963. Moreover, evaluation on test set is carried out every 20 epochs.

So test accuracy is just, let's say, oscillating between those two values. It's totally normal. Probably one would notice the same effect if he/she could zoom in on the other two accuracy curves.

cantonioupao commented 5 years ago

How did you get the graph results? i run the "inference.py" file but didn't get the results

wave-transmitter commented 5 years ago

These graph results were provided by @jfzhang95 in README file after training the model. I also managed to get similar result by running train.py.

As mentioned in issues #8 and #10 there is problem with inference.py and pretrained model loading.

cantonioupao commented 5 years ago

I am running it in Linux and i didn't get this frame windows after running inference.py . I only got a window displaying the video under testing

wave-transmitter commented 5 years ago

As I said previously, you can only get these tensorboard graph results by running "train.py". By running "inference.py" you are just using a pretrained model to make predictions for a single input video. Epochs, accuracy, loss and etc. are parameters of the training procedure.

cantonioupao commented 5 years ago

I already run the train.py in a Linux system , however i don't get the results , only some statistics in every epoch it executes. pytorch_breakfast_testing

wave-transmitter commented 5 years ago

Hey, in my opinion you should open a new issue since your questions are no longer relevant with that issue. We should help the future readers of that topic.

To answer to your question, as specified in @Robert-Browning initial question but also through README, in order to get these graphs you gonna need tensorboardX installed and a tensorboard console launched while running "train.py". You can find instructions on how to do that here. You only have to start the tensorboard server, the setup is already done.