Training my model on a new classes of data

[Krieker (Tantrum)]

I’ve already trained my model to classify images on 2 classes. I exported this model and saved the weight though. Now I want to train this entire model to recognise 2 additional classes. I’ve been trying to

learn = cnn_learner(data, models.resnet50, metrics, pretrained=True)
learn.load("stage1-rn50")
learn.fit_one_cycle(5)

But this not worked. So now I have a question, how can I train my classifier to recognise new classes??

[tango (Lasse Karlsson)]

You would have to replace the final layer to output a different number of classes. Your last model will have 2 output nodes but your new model should have 4 so you should use .split() and replace your final layer with a create_head()

[Krieker (Tantrum)]

In which lesson can I find how to do that?

[tango (Lasse Karlsson)]

I am unsure if there is an exact lesson on this. Refer to the learner And basic_train Part of the documentation to see how you can use the methods I mentioned to replace the head of your network.

Otherwise, you could train your model with 2 classes together with the dataset with the 2 additional classes. In this way you just have no labels of the two additional classes in the first part of your dataset. Try both and see what works!

[muellerzr (Zachary Mueller)]

See here:

Transfer learning... twice?

youre right. i think this might fix it. once youre done training on the planet dataset: learn.model[-1][-1]=nn.Linear(in_features=512,out_features=5, bias=True) i think this should set the last linear layer (outputs) to the number of outputs for your data. i just plugged in the same number of in_features that learn.model[-1][-1] printed out and switched out_features with the number of classes your dataset has. then you can do learn.save(‘double-pretrain’), and you should be able to make a ne…

And here:

Transfer learning from fastai learner to another leaner

i think this is doing the same thing you’re talking about but a little differently. you can index into the last linear layer, set it to what you want, then train on the new output classes

Basically change the outputs after you load the weights