kernel size in class FBnetblock

[666zz666]

You didn't use kernel size in class FBnetblock.
What's wrong with Bn?

[JunrQ]

kernel size is specified in fbnet-pytorch/blocks.py by parameter _kernel
and BN will help traing deep neural network

[666zz666]

NAS/fbnet-pytorch/blocks.py

Line 43 in 2cdf88b

nn.Conv2d(C_in*expansion, C_in*expansion, 3, stride=stride,

NAS/fbnet-pytorch/blocks.py

Line 55 in 2cdf88b

nn.Conv2d(C_in*expansion, C_in*expansion, 3, stride=stride,

It's three,not kernel-size.

[JunrQ]

You are right,that’s a bug.

[666zz666]

So,Why don't you use bn in your code?

[JunrQ]

The paper doesn't use bn

[666zz666]

Maybe this is the default

[JunrQ]

maybe, BN is the default for deep neural network.

The insteresting thing is that the search part converges very well without BN.
But training the sampled model without BN will fail to converge.

Any idea why this happens?

[GuntherZhong]

You are right,that’s a bug.

Hi, it seems that the padding size should be modified too :-)

[666zz666]

Maybe padding = (kernel_size - 1) // 2 is better.