adding dropout to the googlenet extractor about ghiaseddin HOT 7 CLOSED

@yassersouri I'll work on these in two weeks. Is it ok?

from ghiaseddin.

Yes

On Sat, Apr 23, 2016 at 7:01 PM, Erfan Noury [email protected]
wrote:

@yassersouri https://github.com/yassersouri I'll work on these in two
weeks. Is it ok?

—
You are receiving this because you were mentioned.
Reply to this email directly or view it on GitHub
#1 (comment)

from ghiaseddin.

@erfannoury
I hope that everything went well. I am waiting on this.

from ghiaseddin.

@yassersouri Thank you very much. I'm on it.

from ghiaseddin.

@yassersouri There two places in which dropout is used in this network according to the paper.

1. One of them is used in before the final linear layer in the "extra networks on the side", with 70% ratio of dropped outputs.
2. Another dropout layer is used before the final linear layer in the main network (Table 1) with 40% ratio of dropped outputs.

However since the extra side networks are only used when training the network, we wouldn't use them in our architecture. Therefore only the second dropout layer before the final layer shall be used (after the final global average pooling layer). This wouldn't be needed if we only used the network, but since we are planning on pre-training the network, therefore we will be using a linear layer after the final average pooling layer, we need the dropout layer after the final average pooling layer (pool5/7x7_s1).

It is added now using commit 83e02ba

from ghiaseddin.

Albeit, since we are planning on pre-training the network, we might need the extra side networks after all. Have you decided on this @yassersouri?

from ghiaseddin.

I don't think we will be needing the auxiliary networks.

from ghiaseddin.