InceptionResNetv2 and Xception low performace

Part 1 (2019) Advanced (Part 1 v3)
1

Hi,
I’m working on a semantic segmentation problem. I’m using U-Net architecture. For the encoder, I tested pre-trained ResNet50, ResNext50, Xception, and InceptionResNet v2 and I always get that ResNet 50 is most accurately (F1: 0.78, 0.72, 0.5, 0.51 respectively). For ResNext, Xception, InceptionResNetv2 I’m using the Cadene pre-trained models. This is my code for Xception:

def get_model(model_name:str, pretrained:bool, seq:bool=False, pname:str='imagenet', **kwargs):
    pretrained = pname if pretrained else None
    model = getattr(pretrainedmodels, model_name)(pretrained=pretrained, **kwargs)
    return nn.Sequential(*model.children()) if seq else model
def xception(pretrained:bool=False):   return get_model('xception', pretrained)
def _xception_split(m): return (m[0][11], m[1])
learn = unet_learner(data, xception, metrics=metrics, path = path, split_on = _xception_split)

In literature, usually xception and InceptionResNetv2 outperform ResNet. Am I doing something wong?

2

Well, did you try cadene resnet? Fastai usually works better than others, and resnets are not always implemented the same way. Also fastai resnet50 has more parameters than one I used from another library, maybe try checking how many the ones from cadene have

3

I tested ResNet50 architecture from Cadene and I got F1:0.69. So, the value is lower but still not so bad as Xception and InceptionResNetv2.