Can you run the Opacus privacy engine with pytorch SequenceTaggingDataset?

Question

I am trying to adapt a pytorch Named Entity Recognition model to incorporate differential privacy with the Opacus library. My model uses torchtext to build the dataset and feed sentences through a word embedding layer and char embedding layer, concatenate the representation and feed it in an LSTM which tags words.

In building the dataset I use the SequenceTagginDataset from torchtext 0.6.0 with fields word, char and tag.

self.train_dataset, self.val_dataset, self.test_dataset = SequenceTaggingDataset(path=input_folder + '/DRUG-AE.tsv', fields=(
            (("word", "char"), (self.word_field, self.char_field)),
            ("tag", self.tag_field)
        )).split(split_ratio=[0.3, 0.1, 0.1])

Then I create DataLoader from the created datasets:

self.train_iter = DataLoader(self.train_dataset, batch_size=batch_size)
self.val_iter = DataLoader(self.val_dataset, batch_size=batch_size)
self.test_iter = DataLoader(self.test_dataset, batch_size=batch_size)

When defining the privacy engine of opacus it expects the model, optimizer and train dataloader. However, when doing so I receive an error message: Uniform sampling is not supported for IterableDataset. Is there a way to transform this IterableDataset to a regular one to use with the privacy engine? The data is relatively small on a local file. Thank you for any time you spend looking into this. In case anything isn't clear or you want more code snippets let me know.