Natural language processing

Feature request

We currently have ViLT in the library, which, among other tasks, is capable of performing visual question answering (VQA).

It would be great to have a pipeline for this task, with the following API:

from transformers import pipeline

pipe = pipeline("vqa")
pipe("cats.png", "how many cats are there?")
`

Describe the bug

Streaming Datasets can't be pickled, so any interaction between them and multiprocessing results in a crash.

Steps to reproduce the bug

import transformers
from transformers import Trainer, AutoModelForCausalLM, TrainingArguments
import datasets

ds = datasets.load_dataset('oscar', "unshuffled_deduplicated_en", split='train', streaming=True).with_format("

In gensim/models/fasttext.py:

    model = FastText(
        vector_size=m.dim,
        vector_size=m.dim,
        window=m.ws,
        window=m.ws,
        epochs=m.epoch,
        epochs=m.epoch,
        negative=m.neg,
        negative=m.neg,
        # FIXME: these next 2 lines read in unsupported FB FT modes (loss=3 softmax or loss=4 onevsall,
        # or model=3 supervi

Is your feature request related to a problem? Please describe.
I typically used compressed datasets (e.g. gzipped) to save disk space. This works fine with AllenNLP during training because I can write my dataset reader to load the compressed data. However, the predict command opens the file and reads lines for the Predictor. This fails when it tries to load data from my compressed files.

Checking the Python files in NLTK with "python -m doctest" reveals that many tests are failing. In many cases, the failures are just cosmetic discrepancies between the expected and the actual output, such as missing a blank line, or unescaped linebreaks. Other cases may be real bugs.

If these failures could be avoided, it would become possible to improve CI by running "python -m doctest" each t