samsungsailmontreal / ghn3 Goto Github PK

Can you explain the few-shot setting whose result are reported in
Table 7. Transfer learning from ImageNet to few-shot CIFAR-10
and CIFAR-100 with 1000 training labels with 3 networks: ResNet50 (R-50), ConvNext-B (C-B) and Swin-T (S-T)

How many shots in the training set as well as in the test set or query set.

This is the code i use to evaluate ghn3clm16 on cifar-10 but the top1 results was 0.41 and 1.68 for top5

did I do something wrong.
here is my test code

`import torch
import torchvision
from ppuda.config import init_config
from ghn3 import from_pretrained, norm_check, Graph, Logger
from ppuda.config import init_config
from ppuda.utils import infer, AvgrageMeter, adjust_net
from ppuda.vision.loader import image_loader

args = init_config(mode='eval', debug=0, arch='resnet50', split='torch') # load arguments from the command line
assert args.arch is not None, ('architecture must be specified using, e.g. --arch resnet50', args.arch)

def bn_set_train(module):
if isinstance(module, torch.nn.BatchNorm2d):
module.track_running_stats = False
module.training = True

ghn = from_pretrained(args.ckpt, debug_level=args.debug).to(args.device) # get a pretrained GHN

is_imagenet = args.dataset.startswith('imagenet')
is_torch = args.split == 'torch'
print('loading the %s dataset...' % args.dataset)
val_loader, num_classes = image_loader(args.dataset,
args.data_dir,
test=True,
test_batch_size=args.test_batch_size,
num_workers=args.num_workers,
noise=args.noise,
im_size=224,#args.imsize,
seed=args.seed)[1:]

model = eval(f'torchvision.models.{args.arch}()').to(args.device) # create a PyTorch model

if is_torch and not is_imagenet:
model = adjust_net(model, large_input=False) # adjust the model for small images such as 32x32 in CIFAR-10

with torch.no_grad(): # to improve efficiency
model = ghn(model) # predict parameters of the model
model.eval()
model.eval() # set to the eval mode to disable dropout, etc.

model.apply(bn_set_train)
top1, top5 = infer(model.to(args.device), val_loader, verbose=False)
print(f'top1: {top1} and top5:{top5} ')`

Thank you for your interesting research. I have some questions regarding the paper:

1. I'm curious about the adaptability of GHNs to other standard-sized datasets, particularly in different tasks such as image segmentation. The Penn-Fudan dataset discussed in your paper seems relatively small. Could you share your thoughts on this?
2. Do you remember the accuracy of the models in DeepNets-1M achieved during the training of GHNs model? How are they compared to the predicted parameter model accuracy with/without fine-tuning?

Thank you.

Hello thank you for your work.
Why random INIT performance is so high in Table 7 CIFAR-10 if the model has not been trained before.
does random init mean model initialize with Xavier initialization?
when using swint without any pretrained weights on CIFAR-10 as random initialize the performance is around 5 to 10. so I do not understand how you get that huge performance with no pretraining with the baseline random init.
thank you