So, don’t expect it to magically turn massive models into a smaller model. Of course, the reduction can only be so much. To reduce these problems, PyTorch has a very useful feature called TorchScript which optimizes the model for deployment. Models can take a while to make a prediction.The model is then saved to a file named ‘model.pkl’ which we will use as the model for the Android app.īefore working on the app, you can imagine a normal model has several problems. Torch.save(model.state_dict(), 'model.pkl') # Save the model to a file named model.pkl # Send the noise vectors through the model The function below is used to show how we can get a sample of N essentially random numbers from the model by inputting N noise vectors. Nonetheless, we can just take the index of the max output value (argmax) in the vector to get the predicted value. Of course, the model isn’t trained, so the probability output doesn’t have any context. The model will take in a vector with 2 elements (which is just some arbitrary number I chose) and output a vector with 10 elements where each element in the output vector is the probability the model thinks that number should be chosen. # the softmax probabilities of a number 0 to 9 # Basic MLP with 2 inputs, 4 hidden layers To show how PyTorch can be used with a model, I’m going to be creating a simple model that takes N noise vectors as input and outputs N numbers between 0 and 9. Now the setup is complete and we can start creating our app! Model Creation Click “Sync Now” in the upper right-hand corner and the project should update.

Pt nn models