Split model

I am trying to figure out the best way to use the split functionalities of fastai. I have a model merging image data, tabular data and variable length text data. I use a pre-trained cnn for the image, pre-trained language model for the text bit. I want to split my model to be able to use freeze and use discriminative learning rates…

Here is my model:

class ImageTabularModel(nn.Module):
    "Basic model for tabular data."
    def __init__(self, emb_szs:ListSizes, n_cont:int, layers:Collection[int], vocab_sz:int, encoder):
        self.cnn = create_body(models.resnet34)
        layers = [400 * 3] + [512]
        ps = [.4]
        self.lm_encoder = SequentialRNN(encoder[0], PoolingLinearClassifier(layers, ps))

        self.tab = TabularModel(emb_szs, n_cont, 512, layers)

        self.reduce = nn.Sequential(*([Flatten()] + bn_drop_lin((512*7*7), 512, bn=True, p=0.5, actn=nn.ReLU(inplace=True))))
        self.merge = nn.Sequential(*bn_drop_lin(512 + 512 + 512, 512, bn=True, p=0.5, actn=nn.ReLU(inplace=True)))
        self.final = nn.Sequential(*bn_drop_lin(512, 2, bn=True, p=0., actn=nn.ReLU(inplace=True)))

    def forward(self, img:Tensor, x:Tensor, text:Tensor) -> Tensor:
        imgLatent = self.reduce(self.cnn(img))
        tabLatent = self.tab(x[0], x[1])
        textLatent = self.lm_encoder(text)[0]

        cat = torch.cat([imgLatent, tabLatent, textLatent], dim=1)

        return self.final(self.merge(cat))
    def reset(self):
        for c in self.children():
            if hasattr(c, 'reset'): c.reset()

And then here is my code to split this into multiple layer_groups…

def split_layers(model:nn.Module) -> List[nn.Module]:
    groups = [[model.cnn, model.lm_encoder]]
    groups += [[model.tab, model.reduce, model.merge, model.final]]
    return groups

I pass this split function to my learner. I tried to copy the same kind of logic from the split examples in fastai code, but not sure I am doing it right.

From what I understand, Learner.freeze freeze all layer groups except last one. I just want to make sure I take full advantage of this.



I think it should split properly. You can check by inspecting learn.layer_groups after the split.

Yeah, I confirm it works I trained the model with freeze and unfreeze. When I inspect layer groups it seems fine but the layers are not flattened, so not sure if fastai is able to apply discriminative learning rate correctly here. Looking at other split examples in the code they seemed to use flatten_model but I had problems when I tried using flatten_model with my split function when trying to train my model (when fastai tried creating the optimizer).


I am trying to split a model in fastai2, in fastai1 we could do:


whereas in fastai2 there is no split method for Learner - and I just found that the difference now is that you can pass in a function to Learner directly ie:

learner = Learner(dls, model, loss_func=loff_func, splitter=my_model_splitter, metrics=[metric], model_dir=model_path)

No, you just have to pass your splitter at init. There is an example in this tutorial.

1 Like

Thanks sgugger, for all the help.
This link is broken, can you share it again.


The link points to the dev subdomain of fast.ai.
Changing dev to docs worked: Tutorial - Using fastai on a custom new task | fastai

Thanks, for help.