pip.md

try modelfeast via pip

1. install

pip3 install modelfeast

2. get a model

from modelfeast import *
model = squeezenet(n_class=10, img_size=(224, 224), pretrained=True)
print(model)

The interface to create a 2D CNN model can be used in the manner:

model = modelname(n_class=10, img_size=256, pretrained=True, pretrained_path="./pretrained/")

Check ./models/__init__.py to see avaliable modelname.

3. train a model using modelfeast

from modelfeast import *
if __name__ == '__main__':
    clf = classifier('xception', 17, (30, 30), 'E:/Oxford_Flowers17/train')
    clf.train()

The class classifier is very flexible.

You can define a model on your own, and train it using classifier.

from modelfeast import *
from torch import nn

#define your own model
class FuckerNet(nn.Module):

    def __init__(self):
        super(dal_BN, self).__init__()
        self.sq1 = nn.Sequential(
            nn.Conv2d(1, 6, 3, padding = 1),
            nn.BatchNorm2d(6),
            nn.ReLU(),
            nn.MaxPool2d(2, 2),

            nn.Conv2d(6, 16, 5), #padding = 0 , stride=1,默认
            nn.BatchNorm2d(16),
            nn.ReLU(),
            nn.MaxPool2d(2, 2)
            )
        self.linear = nn.Linear(400 ,10) 

    def forward(self, x):
        x = self.sq1(x)
        y = self.linear(x.view(x.shape[0], -1))
        return y

if __name__ == '__main__':

    clf = classifier(model=FuckerNet(), 17, (30, 30), 'E:/Oxford_Flowers17/train')
    clf.train()

You can define your own dataloader, optimizer, lr_schedule, loss, metric and use classifier to do the rest ! To learn more, please read classifier.py.

Life is short, there's no reason to spend time on meaningless things. So, enjoy modelfeast !