search

crestonbunch / tbcnn

Efficient tree-based convolutional neural networks in TensorFlow
MIT License
150 stars 44 forks source link

Is the implement of tree-based convolutional layer same as the paper ? #4

Open Zero0zz opened 7 years ago

Zero0zz commented 7 years ago

I confused with the output_size of conv_layer function.Then I compared the code with the paper. I found that, the output of the feature detectors in the code is not same as the formula in the paper described. Is there something wrong?

crestonbunch commented 7 years ago

Hi, thanks for your comment. My implementation is not precisely the same. I followed the techniques described in their paper rather than trying to follow their original source code.

The output of the convolution layers has shape (batch_size, num_nodes, output_size * num_conv). The two differences from the author's source code is the batch_size and num_conv dimensions. The batch_size dimension is simply the size of the batches.

I believe the author's code uses batch sizes of 1. Additionally, I enable the possibility of learning multiple separate convolutions, inspired by the way CNN layers typically work on image datasets. By setting num_conv to something larger than 1 (the default) you can train multiple, separate convolutions (conv_nodes) which are concatenated together as inputs to the next layer. I have observed no improvements by setting num_conv > 1, but it may be useful for more complex problems.

BGMpool2014 commented 4 years ago

求一下联系方式，有偿提问。thanks.