How integrate custom pytorch (not pre-trained) model into Fastai v1

fastai
1

Hi Fastai fellows,

I am trying to implement Darknet code from lesson 12 into Fastai v1.0. But looks like learn.create_body is for the pre-trained model, not for custom Pytorch model. Does learn.create_body still works for custom Pytorch model and how to do that? How do I put my custom model integrate with into fastaiv1? Hope some directions and small demo.

Really thanks for all developers creating the amazing library and all classes on Fastai.

Best,
Nelson

2

I found the way to implement Darknet model in doc. :grinning:

m = models.Darknet([1, 2, 4, 6, 3], num_classes=10, nf=32)
learn = Learner(data, m, metrics=[accuracy])
learn.fit_one_cycle(10,1e-2)

:slight_smile:

2 Likes
3

Ah so the Darknet is already available?

4

The question remains. If you want to use a custom model, how do you do?

5

It looks creat_body may not work for all the PyTorch models as it is, as it expects the model to have certain characteristics, like the model creation class to have pretrained as its first argument.
Let’s take an example of densenet from pretrainedmodels package models and see how to use create_body on it.

Let’s take a look at create_body code.

def create_body(arch:Callable, pretrained:bool=True, cut:Optional[Union[int, Callable]]=None):
    "Cut off the body of a typically pretrained `model` at `cut` (int) or cut the model as specified by `cut(model)` (function)."
    model = arch(pretrained)
    cut = ifnone(cut, cnn_config(arch)['cut'])
    if cut is None:
        ll = list(enumerate(model.children()))
        cut = next(i for i,o in reversed(ll) if has_pool_type(o))
    if   isinstance(cut, int):      return nn.Sequential(*list(model.children())[:cut])
    elif isinstance(cut, Callable): return cut(model)
    else:                           raise NamedError("cut must be either integer or a function")

The function passes pretrained to the arch argument, where we would be passing typically our model.

Let’s also look at how we can create a densenet model from pretrainedmodels.

densenet = pretrainedmodels.densenet121(num_classes=1000, pretrained='imagenet')

If you pass directly pretrainedmodels.densenet121 directly to create_body it fails as it passes pretrained value to denset function for num_classes.

Let’s create a custom function which accepts arguments in a way that satisfies both create_body and densenet.

def densenet(pretrained):
    return pretrainedmodels.densenet121(pretrained='imagenet') if pretrained else pretrainedmodels.densenet121(pretrained=None)

Lets try passing our custom model to create_body. It successfully runs and returns empty Sequential object.

create_body(densenet)

It returns an empty object as fastai cnn_config method does not know anything about the densenet architecture. Taking a close look at the densent model we will realize we want create_body function to cut the densenet model just before the last layer to create the model body. So passing -1 to cut should get our body of densenet.

create_body(densenet,cut=-1)

For other architectures cut value may not be simple -1, in such cases you can pass a function which knows how to split the model.

2 Likes
6

Do you mean like this:

layers = [some layers here]
model = nn.Sequential(layers)

learn = Learner(data, model)
learn.fit_one_cycle(10, 3e-4)
1 Like
7

This will give you an error. TypeError: list is not a Module subclass
code correction, it should be

model = nn.Sequential(*layers)