I have developed a tabular fastai deep learning model. Since lambda does not support fastai, I need to import the model to pyTorch and do the prediction. I can get the model state using:
state = torch.load(‘models/mymodel.pth’, map_location=torch.device(‘cpu’))
But I think this is not the full model. How can I import the model and make the prediction? I think I need to import it differently, or maybe design the structure again on pyTorch.
I have gone through the forum but the answers on similar topics are not clear enough.
Why do you believe it’s not the full model? All the layers point back to PyTorch so it should be usable. Also: the model (unsure if you looked) accepts the cat and cont variables as two separate tensors for your input. Also how are you preparing your prediction tensor?
Thanks Zachary for your reply.
I am not fully familiar with pyTorch (unlike fastai). I loaded the model, but the data type is dict which has all the fastai model layers and info (is this accurate? [Solved] Using a fastai-trained model with plain Pytorch).
How can I make the model from this dictionary? and how can I make predictions using it?
In fastai everything was simple:
fastai_prediction = learn_nn.predict(data.iloc)
where learn_nn is the NN model. What is the equivalent process in pyTorch?
I think what you refer to as prediction tensor is equivalent of data.iloc in this example, but I am stuck in one step before that.