Training results of `train_search.py`? about darts HOT 7 CLOSED

OK that works -- I can verify that running train_search.py and training the resulting genotype gives comparable results to the DARTS model in cnn/genotypes.py.

from darts.

Yes, just copy & paste it into genotypes.py as you described : )

Then, run python train.py --arch $NAME_OF_THE_ARCH --auxiliary --cutout for evaluation.

from darts.

Cool thanks -- and how would you recommend sampling a random architecture? Something like this?

model = Network(args.init_channels, CIFAR_CLASSES, args.layers, criterion)
model.alphas_normal = Variable(torch.randn(k, num_ops))
model.alphas_reduce = Variable(torch.randn(k, num_ops))
random_genotype = model.genotype()

from darts.

Yep, that should do.

from darts.

I am also getting around 89% validation accuracy when running train_search.py. Are these expected results for this step. I see no expected results for this in the documentation or paper.

from darts.

It looks fine. The validation acc during arch search does not tell too much because the weights are under-trained. You'll need to train the architecture from scratch in order to evaluate it and achieve ~2.83%.

from darts.

@quark0 hi, I'm wondering how many epochs do you train the searched architecture, I used the searched arch on epoch 49, and it it only gets 95.0 valid acc after retrained for 350 epochs

from darts.