A few questions about the implementation

[jewes]

Hello abhaydoke09,
Thanks for sharing your nice work. I have a few questions about the implementation:

1. The default input image size is 448x488, and this results in the shape of last convolution layer (conv5_3) is (-1, 28, 28, 512). In the following line, is the 784 the result of 28x28? If the input image size is reduced to 224x224, then the shape of conv5_3 becomes (-1, 14, 14, 512), should the 784 in the following be changed to 256 (which is 14 x 14)?

self.phi_I = tf.divide(self.phi_I,784.0)  

2. It looks like the number of parameter in last fully connected layer is always 512512num_classes no matter what the input image size is. Is this expected?

Agains, thanks for the nice work!

[reckdk]

From these lines, '784' is the result of 28x28, which corresponds to 'feature dimension c' in Bilinear-CNN paper.
Please refer to #7 and these lines for more info.