Skip to content

Latest commit

 

History

History
48 lines (38 loc) · 2.17 KB

README.md

File metadata and controls

48 lines (38 loc) · 2.17 KB
Raw

EfficientNet Lateral Feature Extraction

This file extracts the lateral features of any efficient net model . THere are basically seven layers of features of various sizes and depth depending upon configurations of Effnet model as per its block_Args .Please check out for config of effnet models here https://github.com/lukemelas/EfficientNet-PyTorch The way it works is Every block of Effnet starts from some depth say 16 after series of convolution it would end with same depth depending upon number of repeats before new layer size would be starting (indicated by depth of last layers in MBblock) given by depth coefficient * num repeat - 1.2*3=3.6 take the largest integer=4 afte decimal ,3.2 is also 4.

One Limitation this has got is .This can work with image size multiple that is multiple of 64 or 128 any one of them.Use below script to check for the sizes . Upsample of uplayers should match with max pool of layers down #Script To validate he sizes

base_model = myEfficientNet.from_pretrained('efficientnet-b2') x=torch.randn(1,3,320,1280) #x_center = x[:, :, :, 1536 // 12: -1536 // 12] print(x_center.size()) x = base_model._swish(base_model._bn0(base_model._conv_stem(x_center))) size1=0 feature=[] for b in base_model._blocks:

tmp=b(x)
x=tmp
#print(tmp.size())
if size1!=tmp.size(1):
    feature.append(tmp.size(1))
    size1=tmp.size(1)
    print(tmp.size(),nn.MaxPool2d(2)(x).size())

feats= base_model._swish(base_model._bn1(base_model._conv_head(x)))
device='cpu' print(feats.size()) bg = torch.zeros([feats.shape[0], feats.shape[1], feats.shape[2], feats.shape[3] // 8]).to(device) feats = torch.cat([bg, feats, bg], 3) feats.size()

Loaded pretrained weights for efficientnet-b2 torch.Size([1, 3, 320, 896]) torch.Size([1, 16, 160, 448]) torch.Size([1, 16, 80, 224]) torch.Size([1, 24, 80, 224]) torch.Size([1, 24, 40, 112]) torch.Size([1, 48, 40, 112]) torch.Size([1, 48, 20, 56]) torch.Size([1, 88, 20, 56]) torch.Size([1, 88, 10, 28]) torch.Size([1, 120, 20, 56]) torch.Size([1, 120, 10, 28]) torch.Size([1, 208, 10, 28]) torch.Size([1, 208, 5, 14]) torch.Size([1, 352, 10, 28]) torch.Size([1, 352, 5, 14]) torch.Size([1, 1408, 10, 28])