MXNet NDArray with dtype string?

Question

I want to train a CV network with MXNet and I created a custom Dataset class (works well) but I'd now like this class to return the name of the image file (a string).

Problem is MXNet refuses to batch strings.

So I tried to create a NDArray from my string but NDArray won't take strings as dtype. What should I do?

>>> import numpy as np
>>> import mxnet.ndarray as nd
>>> nd.array(["blabla"])
ValueError: could not convert string to float: blabla

Answer 1

@Manon Rmn, the solution for that is to have a custom Batchify function.

import mxnet as mx
import numpy as np
from gluonnlp.data.batchify import Tuple, Stack

class MyDataset(mx.gluon.data.Dataset):

    def __init__(self, size=10):
        self.size = size
        self.data = [np.array((1,2,3,4))]*size
        self.data_text = ["this is a test"]*size

    def __getitem__(self, idx):
        return self.data[idx], self.data_text[idx]

    def __len__(self):
        return self.size

dataset = MyDataset()
print(dataset[0])


class List:
    def __call__(self, x):
        return x

data = mx.gluon.data.DataLoader(dataset, batchify_fn=Tuple(Stack(), List()), batch_size=2)

for matrix, text in data:
    break
print(matrix)
print(text)

(array([1, 2, 3, 4]), 'this is a test')

[[1 2 3 4]
 [1 2 3 4]]
<NDArray 2x4 @cpu_shared(0)>
['this is a test', 'this is a test']

I have issued a PR to get that into GluonNLP https://github.com/dmlc/gluon-nlp/pull/812