Can i use this method with a [3x320x320] img shape on a Nvidia 1080ti 11GB ? about pc-darts HOT 5 CLOSED

hi， @cristyioan2000 , I think it would be ok if you use a smaller batch-size, while the time cost will be increased.

from pc-darts.

The only batch-size i can run is 2 for this image size. I guess that's not so good because the batchnorm does nothing.

from pc-darts.

For imagesize as big is as 320， I recommend you use the supernet we adopt to search on imagenet in model_search_imagenet.py which is different from that we use in cifar(three 2x downsampling layers in the beginning) . I think if you use this architecture， batch-size can be larger.

from pc-darts.

If I want to search an architecture for a custom dataset is it enough to use train_search.py using the architecture from model_search_imagenet.py ?

# file:  train_search.py
# from model_search import Network
from model_search_imagenet import Network

from pc-darts.

Hi，yes. But the data processing is also necessary to be replaced. Learning rate, batch-size should be suitable for your data.

from pc-darts.