h3r | Uber 's h3 bindings to the R Programming Language | Map library

You are trying to pass nested list/array objects as a feature to convert to tensor. That's the reason for the error. You can handle it at pandas level by simply converting the n length lists/arrays to n columns (check solution 2). However, usually, when working with such columns, you ideally want to process them differently in the network (such as pass this column into an LSTM for example). Therefore an ideal way is to have a multi-input model, which is usually how we work with these features in the industry (check solution 1).

This is a fairly common problem especially when we are working with multiple sequences of data or multiple encodings.

One straightforward method of solving this is by creating separate inputs for each encoding.

1. (Assuming X_train has 9 columns) Pass 8 columns out of the 9 to the first input, and the encoding (column with list/array) as a separate input.
2. Concatenate these to create an 8+398 length tensor which now passes through the computation graph.
3. The single series with lists can be converted to a tensor/np.array by np.array(df.column.tolist()). This will convert the (398,) length series with lists to a (398, 398) shaped NumPy array.
4. Now, you can handle the features and the encodings separately as well before concatenating them and passing them through Dense layers. E.g passing the second input via LSTM layers.