CondeseNet-182* on Cifar100 validation 1 error rate is 19.73% where in paper is 18.47%

[lizhenstat]

Hi, I run CondenseNet-182* using command provided by issue 11 on Cfiar100

python main.py --model condensenet -b 64 -j 2 cifar100 --epochs 600 --stages 30-30-30 --growth 12-24-48

The result of the first run is 19.73%, the result in the second run is 19. 86% The result in the paper is 18.47% (Table) I just used all the default arguments in the code provide, do we need to make other changes?

Thanks in advance

[ShichenLiu]

Hi there,

You might need to set dropout rate to fully reproduce the results!

[lizhenstat]

Thanks for your replying, I will add the dropout rate and try again.

[lizhenstat]

@ShichenLiu Hi, I noticed that the dropout is placed before convolution layer, In the original densenet-torch implementation, the order in each block is BN-->relu-->conv-->dropout Is there a particular reason for doing so?

    def forward(self, x):
        self._check_drop()
        x = self.norm(x)
        x = self.relu(x)
        if self.dropout_rate > 0:
            x = self.drop(x)
        ### Masked output
        weight = self.conv.weight * self.mask
        return F.conv2d(x, weight, None, self.conv.stride,
                        self.conv.padding, self.conv.dilation, 1)

[lizhenstat]

After I add the dropout-rate-0.1, the final classification error rate is 19.2%, still higher than reported in the paper.

[ShichenLiu]

Hi,

You could try larger dropout rate, e.g. 0.2, 0.3 and even 0.5.

On Sat, Sep 14, 2019 at 8:58 PM lizhen notifications@github.com wrote:

After I add the dropout-rate-0.1, the final classification error rate is 19.2%, still higher than reported in the paper.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/ShichenLiu/CondenseNet/issues/28?email_source=notifications&email_token=ACF47YRMAAE64QYL35D3BNLQJWXG7A5CNFSM4IVQEYM2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD6XIVNA#issuecomment-531532468, or mute the thread https://github.com/notifications/unsubscribe-auth/ACF47YWQOFVOFX45DIBSPC3QJWXG7ANCNFSM4IVQEYMQ .

[lizhenstat]

Hi, I tried different dropout rate for condensenet-182 on cifar100, here is the result: condensenet-182-dropout 18.79%(dropout rate=0.1) condensenet-182-dropout 18.7% (dropout rate=0.2) I can't get the the proposed 18.47% So I was wondering how many GPUs do you use? (I am only use one GPU here)