search

amasky / ram

Recurrent Models of Visual Attention (RAM) with Chainer
MIT License
44 stars 8 forks source link

rho dim is correct? #7

Closed jeong-tae closed 6 years ago

jeong-tae commented 6 years ago

https://github.com/amasky/ram/blob/master/ram.py#L128

maybe the rho has [B, C n_scales, size, size] dim. But the network applies fully connected layer: [C n_scales size, size], not conv2d. although the number of parameters are same, the operation can't be applied

How can it works properly for the fc operation?

amasky commented 6 years ago

Hi jeong-tae! In Chainer, as far as I remember, 4-axis data are automatically converted into 2 axis when we feed conv2d data into FC layers.

jeong-tae commented 6 years ago

oh, that's very useful! thx for the reply!