How to get intermediate output not only the last features

[liutianling]

#coding:utf8
import torch
from efficientnet_pytorch import EfficientNet
import torch.nn as nn

class myModel():
def init(self):
base_model = EfficientNet.from_pretrained(model_name="efficientnet-b5", weights_path="efficientnet-b5-b6417697.pth")
#self.model = nn.Sequential(*list(base_model.children())[:2], *list(base_model.children())[2][:38])
self.model = nn.Sequential(*list(base_model.children())[:6])
self.model.eval()
#print(self.model)

def forword(self, x):
	x = self.model.forward(x)
	print(x.shape)

if name=="main":
model = myModel()
x = torch.randn(1,3,32,32)
model.forword(x)

I get the error: TypeError: forward() takes 1 positional argument but 2 were given

when I set the self.model = nn.Sequential(*list(base_model.children())[:2], *list(base_model.children())[2][:38])
It works !
Why and how to fix it...
Thank you very much ...

[Emmunaf]

If you need to prune just the last layers to use it as features extractor this seems a duplicate of this issue that was solved with this PR by nwschurink

If you need a more customised model look at:

• This SO answer of mine for a short case explanation
• This (basically explains how you can override some layers with identity and so how to "delete" some layers)
• This if you want to extend the model

[veeara282]

Use EfficientNet.extract_endpoints(). It returns the intermediate values of the neural network before each reduction in image size (224 × 224 → 112 × 112, etc.).