Would you mind to provide the CIFAR-10 model?

Question

Would you mind to provide the CIFAR-10 model?

D-X-Y opened this issue 6 years ago · 4 comments

Would you mind to provide the CIFAR-10 model?
I use the provided architecture https://github.com/chenxi116/PNASNet.pytorch/blob/master/genotypes.py#L5
and the code from https://github.com/quark0/darts/blob/master/cnn/model.py#L111, but obtain the model parameter size of 4.2 MB (I use initial channel of 48 and layers of 11). Is there anything wrong?

Answer 1 · 2018-09-25T17:01:26.000Z

Thanks for the question. The number of layers should be 9, and their code used a different stem structure (between image and first cell) than mine. These are probably the cause of the parameter size mismatch.

Answer 2 · 2018-09-25T18:02:23.000Z

@chenxi116 Thanks for your information. Would you mind to let me know the stem structure that you use for CIFAR?

Answer 3 · 2018-09-25T18:04:56.000Z

Please see the TensorFlow implementation for reference: https://github.com/tensorflow/models/blob/master/research/slim/nets/nasnet/nasnet.py#L275

Answer 4 · 2018-09-25T18:13:56.000Z

Thanks a lot. It seems the stem structures for CIFAR are similar (maybe one use the bias term and one not).