Lesson 9 in-class

Part 2 (2017)
142

I haven’t tried it with DeVISE, but I’d expect it to work.

143

@jeremy

Any advice for those of us without .edu addresses to download ImageNet? Tried my .org address but no response, and my private domains have struck out too.

144

@davecg best not asking questions outside of class in these in-class threads, since I don’t normally look at them after class! Could you ask in the Lesson 9 Discussion thread? I just found a solution for you :slight_smile:

145

Sure, thanks!

146

Hi Dave,
Do you know if this Mendeley group has been started? I’m happy to contribute.

147

Don’t Do this (Better way Below)
stride = 1/2 means that we are doing deconvolution which is accomplished in pytorch using nn.ConvTranspose2d so when a paper says stride = 1/2, it is the same as this:

nn.ConvTranspose2d(64, 64, 3, stride=2)

Here is the better way:
This is what Jeremy uses on the paper http://arxiv.org/abs/1603.08155 when he sees the deconvolution layers:

PixelShuffle_ICNR(64,64)

This is because deconvolutions cause the checkerboard artifact that can be seen in the paper and this aims to remove that issue.