How to define my models? Any documentations?

[BostonLobster]

First of all, nice project!!!

In your cifar demo, your models are hard-coded into your lib. So, how do I define my own models and make them work well with the config system?

Hope you can give some detailed documentations on this!

[marsggbo]

Thanks for your interest. Designing a new model in torchline is very easy. You may refer to projects/fake_demo.

In main.py, I build a FakeNet, which is decorated by META_ARCH_REGISTRY. The code looks like

from torchline.models import META_ARCH_REGISTRY

@META_ARCH_REGISTRY.register()
class FakeNet(torch.nn.Module):
    def __init__(self, cfg):
        super(FakeNet, self).__init__()
        self.cfg = cfg
        h, w = cfg.input.size
        self.feat = torch.nn.Conv2d(3,16,3,1,1)
        self.clf = torch.nn.Linear(16, cfg.model.classes)

    def forward(self, x):
        b = x.shape[0]
        out = self.feat(x)
        out = torch.nn.AdaptiveAvgPool2d(1)(out).view(b, -1)
        out = self.clf(out)
        out = torch.softmax(out, dim=1)
        return out

You can just simply assign model.name=FakeNet in the configuration file, then torchline will automatically load your specified model.

By the way, if you want to build a model in another py file (e.g., mymodel.py), instead of main.py, you should import mymodel in main.py; otherwise, torchline will fail to load your model.

# mymodel.py
from torchline.models import META_ARCH_REGISTRY

@META_ARCH_REGISTRY.register()
class NewModel(torch.nn.Module):
    ....

# main.py
import mymodel
...

Building a new loss function or dataset loader is similar to designing a new model. Please contact me if you have any questions.