As I said in the title I want to know which weights in the first layer of my mlp model are connected to a certain freature.
So let's say that i have an ordered list of feature names
feature_names = data.feat_names
and a matrix with a shape (number of samples, number of features) where each feature is represented by a column of a matrix.
X = data.predictors
Since the order is the same between feature_names and the X columns, the first feature contained in feature_names is represented by the first column in X and so on.
After that i use dataloader to load the training and test set (both obtained from X)
trainloader = torch.utils.data.DataLoader(list(zip(X_train, y_train)), batch_size=7, shuffle=True)
valloader = torch.utils.data.DataLoader(list(zip(X_val, y_val)), batch_size=1, shuffle=True)
At this point I train my mlp model defined by two hidden layers and that's it. The problem is that I'd like to know which are the values of the weights connected to a certain input variable "i".
For example, in picture above I may want to know which are the values of the weights that connect the feature x_1 to the network.
I proceeded in this way:
mlp.fc1.weight.numpy()[:,i]
where mlp is my model, fc1 is the first layer and "i" represents the feature for which I want to retrieve the weights. Is it the right solution? Basically with this last line of code I'm taking the i-th column of matrix of weights of the first mlp layer.
