python - Fastai 文本分类器 : batch prediction on unseen data

Question

我一直在使用 fastai 的文本分类器 (https://docs.fast.ai/text.html)。我目前预测看不见的短语的情绪(正面或负面)如下:

def _unpack_prediction(self, text) -> Tuple[bool, float]:
    out = self._model.predict(text)
    return str(out[0]) == "positive", max(out[2][0].item(), out[2][1].item())

def example(self, messages: Sequence[str]):
    results = map(self._unpack_prediction, messages)
    for phrase, out in zip(messages, results):
        print(f"{phrase[:100]}...[{'pos' if out[0] else 'neg'}] - [{out[1]:.2f}]")

给定一个短语列表:

("I love this movie",
  "The actors are good, but this movie is definitely stupid",
  "There is no plot at all!!! Just special effects ")

结果是:

I love this movie...[pos] - [1.00]
The actors are good, but this movie is definitely stupid...[neg] - [0.96]
There is no plot at all!!! Just special effects ...[neg] - [0.95]

但是，按顺序对短语应用预测非常慢。

有没有一种方法可以在不创建测试数据集的情况下使用 fastai 库进行批量预测？

Answer 1

你当然可以。这是执行此操作的示例代码

test_df = pd.read_csv(path_to_test_csv_file)
learn.data.add_test(test_df[target_col_name])
prob_preds = learn.get_preds(ds_type=DatasetType.Test, ordered=True)