dlaptev / ti-pooling Goto Github PK

hi @dlaptev, Thanks. But it does't have the code using the keras.
Do you give Some advice to do it using keras? or rewritten by keras.
I know that keras support the TensorFollow, but i use the theano as the backend. Thanks.

That‘s great. I can try this with caffe.

Hello, I'd like to replicate the results of experiments in Sec.4.2 and cite TI-Pooling in my paper. But I found that the ratio of white samples (label = 255) to black samples (label = 0) significantly affects the final classification accuracy. Could you describe the details on how you did data balance of ISBI 2012 dataset? I'd appreciate it if you kindly provide the script for data balance or the dataset of balanced version used in the paper.

Thank you for publishing the code. This paper is very interesting. I have one question about the architecture and the Lemma 1 in the paper.

I see that the first step of the network is to transform an training image to various versions and then feed them to the rest of the network. Experiments provided by the paper considered the rotation case and the code corresponding to experiment 4.1.1 uses 24 directions to train the network on the mnist-rot-12k dataset.

My question is that can I transform the training image using more than one transformation at the first step ? Lemma 1 says that the set Φ of all possible transformations forms a group. Does the set Φ mean transformations of different forms (e.g., rotation and scale) or mean a particular transformation with different parameters (e.g., 30° rotation and 60° rotation)?

If, in practice, I set the number of transformation to be 24 in the source code, where 12 of them are used to rotate the input image and the rest 12 are used to scale the input image, the outputs of the 24 paths are TI-pooled to form the transformation-invariant features. Is this case covered by Lemma 1?

I have several questions in respect of applying your model for other applications:
I have a dataset with big images. Do I have to change architecture of your default network for using in other datasets?
I want to know what happens if reduce the number of layers in you model (mentioned in paper)? In other words, it is a big opaque spot for me whether I can take your model's convolutional layers and use them as the first layers of my convolutional network to make it rotation invariant?
Moreover, whether reduce_max among extracted features from parallel model has bad effect on result of my new network or not since data are a little different than normal situation in previous works?

Thanks you in advance for giving your time to me.