tdevries / enas_pytorch Goto Github PK

RT, is this a complete implementation? Do you get the reported performance on CIFAR and Imagenet?

Hi,
Thanks for the code sharing!

I tried to read it and run it. 10 epochs in the main function take one Nividia Tesla P100 12G 15h to run. Am I wrong or it just takes this amount of time?

There seems to be some problems in the file ‘models/shared_cnn.py’

class ENASLayer(nn.Module):
    def forward(self, x, prev_layers, sample_arc):
        layer_type = sample_arc[0]
        if self.layer_id > 0:
            skip_indices = sample_arc[1]
        else:
            skip_indices = []
        .......
        for i, skip in enumerate(skip_indices):
            if skip == 1:
                out += prev_layers[i]
        out = self.bn(out)
        return out

In the above code, when layer_id = 0, an error occurs when running to enumerate (skip_indices). Is this part of the code not complete enough?

Looking forward to and thank you for your answer !