Open wstang35 opened 5 years ago
@wstang35 , Just in case, You'd like to visualize your network using visualize_cam
or visualize_saliency
(not visualize_activation
) , wouldn't you?
I can help you. I think the reasonable way is that we make it based on the implement of PR/#128 . Could you share your model source code and some data by Gist, Slack or somewhere ?
Hi @keisen ! Thanks for reply. I want to know what kind of features of my network are capturing, so either visualize_saliency or visualize_activation can be of great help! As for my model, can I just send my model.h5 file to your email? Because I wrap up some conv layer in my source code a little bit, and It would be somehow cumbersome to sort out my model from my project.(And if .h5 model is not enough, I can clean up some model source code, but it might take some time!) Thank you again!
Yes, you can. There is my email-address in my profile. Thank you.
And I'd like to send me some input-data, because it will be used as variable argument value of visualize_*
function.
Hi,
So how should the TimeDistributed be used with the visualize_activation? My model consists of CNN layers and LSTM layers.
Thanks in advance!
Hi! I am facing exactly the same situation: CNN for features extraction + LSTM for classification (I label a sequence, not each frame). Did you find a way to visualize features activation within the image? Thanks for the help!
Hi, this is a good project! I am trying to visualize the class features captured by CNN+RNN model, which means using CNN as feature extractor to extract each timestep input's features, and then feed to RNN, to predict each timestep's label. And as for predictions, I am using TimeDistributed(Dense(5)), so I hope to find a solution to expand keras-vis to TimeDistributed layer. In addition, I am using multiple inputs, and It looks like that for now keras-vis doesn't support multiple input models.
So, how can keras-vis expand to these two situations? Really appreciate for any kinds of suggestions, or I can only try to modify my model into single input and separate TimeDistributed intomultiple output Dense layers.... That will not be a neat one...