vascolopes/epe-nas

epe-nas's Issues

Use EPE-NAS for custom classification

Hi there, I really like your approach and would like to use your code for my thesis. Would you please give me a very short clue about how can I apply EPE-NAS with a random search strategy (the Random Search and Reproducibility NAS that you used in the paper) for an custom dataset for two classes? I like to find the best classification architecture with NAS via your solution.
should I just change dataset path and dataloader? or it is in contradict with NAS-Bench-201.pth file?
I really appreciate your help.
Best

sample networks with higher input resolutions

Hi there,
I would appreciate if you could guide me how can sample netwrok configurations from NAS-Bench while increasing the input size of networks.
For instance all sampled netwrok configurations by network = get_cell_based_tiny_net(config) are having input size of 16 for con2d in their first few blocks:

TinyNetwork(
TinyNetwork(C=16, N=5, L=17)
(stem): Sequential(
(0): Conv2d(3, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(1): BatchNorm2d(16, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(cells): ModuleList(
(0): InferCell(
info :: nodes=4, inC=16, outC=16, [1<-(I0-L0) | 2<-(I0-L1,I1-L2) | 3<-(I0-L3,I1-L4,I2-L5)], |nor_conv_1x10|+|avg_pool_3x30|none1|+|skip_connect0|nor_conv_1x11|nor_conv_3x32|
(layers): ModuleList(
(0): ReLUConvBN(
(op): Sequential(
(0): ReLU()
(1): Conv2d(16, 16, kernel_size=(1, 1), stride=(1, 1), bias=False)
(2): BatchNorm2d(16, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
)

The issue is, would it cause unreliable learning when a custom dataset has small objects in it?
The input size of my train_loader and test loader for a custom datasets are having bigger size like torch.Size([32, 3, 64, 64]) while getting passed to the networks with Conv2d(3, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)

• Do you consider changing config['C'] = 32 or any other values is effective way for increasing accuracy for a custom datasets with resolution like 128x128 or 256x256?