New version of NLP

[muellerzr]

No description provided.

[marcglobality]

Hi @muellerzr , thanks for this repo. What happened with this?

[muellerzr]

Time and motivation :) Wound up moving to other things so NLP never came

[marcglobality]

I understand by your actions that it will not come with this PR then? :D

[muellerzr]

Correct. I may reopen and update these notebooks, as they're good NLP foundational tutorial notebooks for fastai, but nothing beyond that is planned at this moment

[marcglobality]

Fair enough. I was wondering how to predict on the test set. I know it's a bit abusing of you (and you don't need to answer fully, but maybe point me to the right place?)

What I manage to do (I simplified this for readability):

text_block = TextBlock.from_df(
        text_cols=INPUT_COLUMNS, 
        is_lm=False,
        seq_len=1_000,

        # add xxfld between fields
        mark_fields=True,

        # name for the output column
        tok_text_col='ulmfit_text',
)

data_cls = DataBlock(
        blocks=(text_block, CategoryBlock),
        get_x=ColReader("ulmfit_text"),
        get_y=ColReader("label"),  
    )

data_cls = data_cls.dataloaders(
        pd.concat([
            df_train.assign(is_dev=False), 
            df_dev.assign(is_dev=True),
        ]),
        shuffle_train=True,
        bs=32, 
        verbose=False,
        splitter=ColSplitter("is_dev"),
    )

learn = text_classifier_learner(
    data_cls, 
    AWD_LSTM, 
    drop_mult=0.5, 
    metrics=[Precision(), Recall()]
)
learn.load_encoder("1epoch_encoder")
learn.fit_one_cycle(1, 2e-2)

and it runs correctly in (I simplified this for readability)

text = "xxbos xxfld 1 ......... solicitors xxfld 2 xxunk xxunk xxunk xxfld 3 xxmaj gao xxmaj jia..."
learn.predict(text)

My problem now is, given a test_df, how can I predict on it (with the pipeline composing the fields)? Something like:

test_df = pd.concat( 
    (
        df_dev[df_dev.label.eq(1)].sample(10, random_state=88),
        df_dev[df_dev.label.eq(0)].sample(10, random_state=88),
    )
)
dl = learn.dls.test_dl(test_df)
_, __, preds = learn.get_preds(dl=dl)

[muellerzr]

IIRC you need to do proc_df or something along those lines. Have you searched in the forums? I know this was brought up in there (I think the v2 text thread)

[marctorsoc]

IIRC you need to do proc_df or something along those lines. Have you searched in the forums? I know this was brought up in there (I think the v2 text thread)

So I searched in the forums, and could arrive into a better version where at least it predicts something. But still seems like a hack, and I don't get the same metrics for dev when fitting than when using the dev set as a test set. Could you answer to me in there https://forums.fast.ai/t/predictions-for-the-test-set-are-they-correct/86994? Thanks