Thanks for your really good work! Is it possible to transfer form keras to pytorch? Pytorch was actually very popular, but Keras wasn't. Many advanced target detection systems are now pyTorch's domain.

Hi, I'm interested in this work very much and tried to reproduce it. But unfortunately, I met this problem. How can I fix it?

ps: I followed your steps carefully.

Hi, althogh I have run this project successfully, I'm still a little bit confused about the openmax threshold settings. Why you choose 10 as the openmax threshold? Could you please take minutes to explain about it?
Looking forward to your reply. Thanks!

Thank you for your work.
I am a little confused why it seems that the Openmax behaves worse than the Softmax.

Hi
Thanks for sharing your work , i want to use this because i think this might be worth trying, but i'm not sure how to go about this, in your jupyter notebook ,you use an already trained keras model, does it have to a keras model ? or can it be a tensorflow pb file?

Could you may be give a small description on how to use this?
I want to apply this is on a 4 class classification problem?Do i first train the model in keras?

Suggestions would be helpful.
Thanks in advance.

Hello, I tested your code on the MNIST data set and found that the effect is not so good. Can you achieve better open set recognition?

Hi, thank you for sharing your results!

I tried to replicate your example by training the model again, even if the model structure appears to be the same as the one I get from MNIST_CNN.h5 and test-loss=0.069 and test-accuracy=0.99. I get different results when I test it on MNIST dataset. It would be nice if you could share the code of the CNN keras model you used to generate MNIST_CNN.h5.

I hope you will help!

Thank you!

What's the mainly difference between MNIST_CNN.h5 and MNIST_CNN_tanh.h5 ? just finally activation layer using tanh or others ? It will be perfect as showing the source code of these two models.

Can you please copy the original code for training de MINIS_CNN.h5? thanks!!

I'm trying to figure out how to write it, but I can't figure it out. Can you tell me how to do it?

why the euclide-cosine distance is calculated as:
euclidean(mean_feature, feat)/200. + cosine(mean_feature, feat) ?
How did you decide on the factor 1/200?

Dear Author,

I found that in openmax.py, the compute_activation function:
img = np.array( Image.fromarray( (np.squeeze(img)).astype(np.uint8)).resize((28, 28)))

The input image of this function is in float32 since it has divided 255. While this line would cast the image into uint8 and make the output of the model changed (since the casted image is basically all 0s) (including the score and fc8 output). Should there be another way to process the input image?

Cheers!

Getting this error: ValueError: Input tensors to a Functional must come from tf.keras.Input. Received: 0 (missing previous layer metadata) inside the get_activations() function

Is there a way to implement openMax with pytorch?

Dear author! After reading the original paper and your work, I have a question: does open set recognition need to change the penultimate layer of deep learning network in the training phase?Or do you use the Softmax layer during the training phase and only change the penultimate layer of the deep learning network during the testing phase?
Looking forward to your kind reply!

Hello,

Please, the "channel_scores" is the feature vector of the "fc8" layer for a specific channel, then why did you loop over the "categoryid" to compute modified_fc8_score = channel_scores[categoryid] * ( 1 - wscore*ranked_alpha[categoryid] ) ?
Yet, the Len(channel_scores) is not equal to NCHANNELS.

Thank you.