Text data

Functions and transforms to help gather text data in a Datasets

Backwards

Reversing the text can provide higher accuracy with an ensemble with a forward model. All that is needed is a type_tfm that will reverse the text as it is brought in:


source

reverse_text

 reverse_text (x)
t = tensor([0,1,2])
r = reverse_text(t)
test_eq(r, tensor([2,1,0]))

Numericalizing

Numericalization is the step in which we convert tokens to integers. The first step is to build a correspondence token to index that is called a vocab.


source

make_vocab

 make_vocab (count, min_freq=3, max_vocab=60000, special_toks=None)

Create a vocab of max_vocab size from Counter count with items present more than min_freq

If there are more than max_vocab tokens, the ones kept are the most frequent.

Note

For performance when using mixed precision, the vocabulary is always made of size a multiple of 8, potentially by adding xxfake tokens.

count = Counter(['a', 'a', 'a', 'a', 'b', 'b', 'c', 'c', 'd'])
test_eq(set([x for x in make_vocab(count) if not x.startswith('xxfake')]), 
        set(defaults.text_spec_tok + 'a'.split()))
test_eq(len(make_vocab(count))%8, 0)
test_eq(set([x for x in make_vocab(count, min_freq=1) if not x.startswith('xxfake')]), 
        set(defaults.text_spec_tok + 'a b c d'.split()))
test_eq(set([x for x in make_vocab(count,max_vocab=12, min_freq=1) if not x.startswith('xxfake')]), 
        set(defaults.text_spec_tok + 'a b c'.split()))

source

LMTensorText

 LMTensorText (x, **kwargs)

Semantic type for a tensor representing text in language modeling


source

TensorText

 TensorText (x, **kwargs)

Semantic type for a tensor representing text


source

Numericalize

 Numericalize (vocab=None, min_freq=3, max_vocab=60000, special_toks=None)

Reversible transform of tokenized texts to numericalized ids

num = Numericalize(min_freq=2)
num.setup(L('This is an example of text'.split(), 'this is another text'.split()))
start = 'This is an example of text '

If no vocab is passed, one is created at setup from the data, using make_vocab with min_freq and max_vocab.

start = 'This is an example of text'
num = Numericalize(min_freq=1)
num.setup(L(start.split(), 'this is another text'.split()))
test_eq(set([x for x in num.vocab if not x.startswith('xxfake')]), 
        set(defaults.text_spec_tok + 'This is an example of text this another'.split()))
test_eq(len(num.vocab)%8, 0)
t = num(start.split())

test_eq(t, tensor([11, 9, 12, 13, 14, 10]))
test_eq(num.decode(t), start.split())
num = Numericalize(min_freq=2)
num.setup(L('This is an example of text'.split(), 'this is another text'.split()))
test_eq(set([x for x in num.vocab if not x.startswith('xxfake')]), 
        set(defaults.text_spec_tok + 'is text'.split()))
test_eq(len(num.vocab)%8, 0)
t = num(start.split())
test_eq(t, tensor([0, 9, 0, 0, 0, 10]))
test_eq(num.decode(t), f'{UNK} is {UNK} {UNK} {UNK} text'.split())

source

LMDataLoader

 LMDataLoader (dataset, lens=None, cache=2, bs=64, seq_len=72,
               num_workers=0, shuffle:bool=False, verbose:bool=False,
               do_setup:bool=True, pin_memory=False, timeout=0,
               batch_size=None, drop_last=False, indexed=None, n=None,
               device=None, persistent_workers=False,
               pin_memory_device='', wif=None, before_iter=None,
               after_item=None, before_batch=None, after_batch=None,
               after_iter=None, create_batches=None, create_item=None,
               create_batch=None, retain=None, get_idxs=None, sample=None,
               shuffle_fn=None, do_batch=None)

A DataLoader suitable for language modeling

dataset should be a collection of numericalized texts for this to work. lens can be passed for optimizing the creation, otherwise, the LMDataLoader will do a full pass of the dataset to compute them. cache is used to avoid reloading items unnecessarily.

The LMDataLoader will concatenate all texts (maybe shuffled) in one big stream, split it in bs contiguous sentences, then go through those seq_len at a time.

bs,sl = 4,3
ints = L([0,1,2,3,4],[5,6,7,8,9,10],[11,12,13,14,15,16,17,18],[19,20],[21,22,23],[24]).map(tensor)
dl = LMDataLoader(ints, bs=bs, seq_len=sl)
test_eq(list(dl),
    [[tensor([[0, 1, 2], [6, 7, 8], [12, 13, 14], [18, 19, 20]]),
      tensor([[1, 2, 3], [7, 8, 9], [13, 14, 15], [19, 20, 21]])],
     [tensor([[3, 4, 5], [ 9, 10, 11], [15, 16, 17], [21, 22, 23]]),
      tensor([[4, 5, 6], [10, 11, 12], [16, 17, 18], [22, 23, 24]])]])
dl = LMDataLoader(ints, bs=bs, seq_len=sl, shuffle=True)
for x,y in dl: test_eq(x[:,1:], y[:,:-1])
((x0,y0), (x1,y1)) = tuple(dl)
#Second batch begins where first batch ended
test_eq(y0[:,-1], x1[:,0]) 
test_eq(type(x0), LMTensorText)

Classification

For classification, we deal with the fact that texts don’t all have the same length by using padding.


source

Pad_Input

 Pad_Input (enc=None, dec=None, split_idx=None, order=None)

A transform that always take tuples as items

pad_idx is used for the padding, and the padding is applied to the pad_fields of the samples. The padding is applied at the beginning if pad_first is True, and if backwards is added, the tensors are flipped.

test_eq(pad_input([(tensor([1,2,3]),1), (tensor([4,5]), 2), (tensor([6]), 3)], pad_idx=0), 
        [(tensor([1,2,3]),1), (tensor([4,5,0]),2), (tensor([6,0,0]), 3)])
test_eq(pad_input([(tensor([1,2,3]), (tensor([6]))), (tensor([4,5]), tensor([4,5])), (tensor([6]), (tensor([1,2,3])))], pad_idx=0, pad_fields=1), 
        [(tensor([1,2,3]),(tensor([6,0,0]))), (tensor([4,5]),tensor([4,5,0])), ((tensor([6]),tensor([1, 2, 3])))])
test_eq(pad_input([(tensor([1,2,3]),1), (tensor([4,5]), 2), (tensor([6]), 3)], pad_idx=0, pad_first=True), 
        [(tensor([1,2,3]),1), (tensor([0,4,5]),2), (tensor([0,0,6]), 3)])
test_eq(pad_input([(tensor([1,2,3]),1), (tensor([4,5]), 2), (tensor([6]), 3)], pad_idx=0, backwards=True), 
        [(tensor([3,2,1]),1), (tensor([5,4,0]),2), (tensor([6,0,0]), 3)])
x = pad_input([(TensorText([1,2,3]),1), (TensorText([4,5]), 2), (TensorText([6]), 3)], pad_idx=0)
test_eq(x, [(tensor([1,2,3]),1), (tensor([4,5,0]), 2), (tensor([6,0,0]), 3)])
test_eq(pad_input.decode(x[1][0]), tensor([4,5]))

Pads x with pad_idx to length pad_len. If pad_first is false, all padding is appended to x, until x is len pad_len. Otherwise ff pad_first is true, then chunks of size seq_len are prepended to x, the remainder of the padding is appended to x.


source

pad_chunk

 pad_chunk (x, pad_idx=1, pad_first=True, seq_len=72, pad_len=10)

Pad x by adding padding by chunks of size seq_len

print('pad_first: ',pad_chunk(torch.tensor([1,2,3]),seq_len=3,pad_idx=0,pad_len=8))
print('pad_last:  ',pad_chunk(torch.tensor([1,2,3]),seq_len=3,pad_idx=0,pad_len=8,pad_first=False))
pad_first:  tensor([0, 0, 0, 1, 2, 3, 0, 0])
pad_last:   tensor([1, 2, 3, 0, 0, 0, 0, 0])

pad_input_chunk is the version of pad_chunk that works over a list of lists.


source

pad_input_chunk

 pad_input_chunk (samples, n_inp=1, pad_idx=1, pad_first=True, seq_len=72,
                  pad_len=10)

Pad samples by adding padding by chunks of size seq_len

The difference with the base pad_input is that most of the padding is applied first (if pad_first=True) or at the end (if pad_first=False) but only by a round multiple of seq_len. The rest of the padding is applied to the end (or the beginning if pad_first=False). This is to work with SequenceEncoder with recurrent models.

pad_input_chunk([(TensorText([1,2,3,4,5,6]),TensorText([1,2]),1)], pad_idx=0, seq_len=3,n_inp=2)
[(TensorText([1, 2, 3, 4, 5, 6]), TensorText([0, 0, 0, 1, 2, 0]), 1)]
test_eq(pad_input_chunk([(tensor([1,2,3,4,5,6]),1), (tensor([1,2,3]), 2), (tensor([1,2]), 3)], pad_idx=0, seq_len=2), 
        [(tensor([1,2,3,4,5,6]),1), (tensor([0,0,1,2,3,0]),2), (tensor([0,0,0,0,1,2]), 3)])
test_eq(pad_input_chunk([(tensor([1,2,3,4,5,6]),), (tensor([1,2,3]),), (tensor([1,2]),)], pad_idx=0, seq_len=2), 
        [(tensor([1,2,3,4,5,6]),), (tensor([0,0,1,2,3,0]),), (tensor([0,0,0,0,1,2]),)])
test_eq(pad_input_chunk([(tensor([1,2,3,4,5,6]),), (tensor([1,2,3]),), (tensor([1,2]),)], pad_idx=0, seq_len=2, pad_first=False), 
        [(tensor([1,2,3,4,5,6]),), (tensor([1,2,3,0,0,0]),), (tensor([1,2,0,0,0,0]),)])

test_eq(pad_input_chunk([(TensorText([1,2,3,4,5,6]),TensorText([1,2]),1)], pad_idx=0, seq_len=2,n_inp=2), 
        [(TensorText([1,2,3,4,5,6]),TensorText([0,0,0,0,1,2]),1)])

Transform version of pad_input_chunk. This version supports types, decoding, and the other functionality of Transform


source

Pad_Chunk

 Pad_Chunk (pad_idx=1, pad_first=True, seq_len=72, decode=True, **kwargs)

Pad samples by adding padding by chunks of size seq_len

Here is an example of Pad_Chunk

pc=Pad_Chunk(pad_idx=0,seq_len=3)
out=pc([(TensorText([1,2,3,4,5,6]),TensorText([1,2]),1)])
print('Inputs:  ',*[(TensorText([1,2,3,4,5,6]),TensorText([1,2]),1)])
print('Encoded: ',*out)
print('Decoded: ',*pc.decode(out))
Inputs:   (TensorText([1, 2, 3, 4, 5, 6]), TensorText([1, 2]), 1)
Encoded:  (TensorText([1, 2, 3, 4, 5, 6]), TensorText([0, 0, 0, 1, 2, 0]), 1)
Decoded:  (TensorText([1, 2, 3, 4, 5, 6]), TensorText([1, 2]), 1)
pc=Pad_Chunk(pad_idx=0, seq_len=2)
test_eq(pc([(TensorText([1,2,3,4,5,6]),1), (TensorText([1,2,3]), 2), (TensorText([1,2]), 3)]), 
        [(tensor([1,2,3,4,5,6]),1), (tensor([0,0,1,2,3,0]),2), (tensor([0,0,0,0,1,2]), 3)])

pc=Pad_Chunk(pad_idx=0, seq_len=2)
test_eq(pc([(TensorText([1,2,3,4,5,6]),), (TensorText([1,2,3]),), (TensorText([1,2]),)]), 
        [(tensor([1,2,3,4,5,6]),), (tensor([0,0,1,2,3,0]),), (tensor([0,0,0,0,1,2]),)])

pc=Pad_Chunk(pad_idx=0, seq_len=2, pad_first=False)
test_eq(pc([(TensorText([1,2,3,4,5,6]),), (TensorText([1,2,3]),), (TensorText([1,2]),)]), 
        [(tensor([1,2,3,4,5,6]),), (tensor([1,2,3,0,0,0]),), (tensor([1,2,0,0,0,0]),)])

pc=Pad_Chunk(pad_idx=0, seq_len=2)
test_eq(pc([(TensorText([1,2,3,4,5,6]),TensorText([1,2]),1)]), 
        [(TensorText([1,2,3,4,5,6]),TensorText([0,0,0,0,1,2]),1)])

source

SortedDL

 SortedDL (dataset, sort_func=None, res=None, bs:int=64,
           shuffle:bool=False, num_workers:int=None, verbose:bool=False,
           do_setup:bool=True, pin_memory=False, timeout=0,
           batch_size=None, drop_last=False, indexed=None, n=None,
           device=None, persistent_workers=False, pin_memory_device='',
           wif=None, before_iter=None, after_item=None, before_batch=None,
           after_batch=None, after_iter=None, create_batches=None,
           create_item=None, create_batch=None, retain=None,
           get_idxs=None, sample=None, shuffle_fn=None, do_batch=None)

A DataLoader that goes throught the item in the order given by sort_func

Type Default Details
dataset Map- or iterable-style dataset from which to load the data
sort_func NoneType None
res NoneType None
bs int 64 Size of batch
shuffle bool False Whether to shuffle data
num_workers int None Number of CPU cores to use in parallel (default: All available up to 16)
verbose bool False Whether to print verbose logs
do_setup bool True Whether to run setup() for batch transform(s)
pin_memory bool False
timeout int 0
batch_size NoneType None
drop_last bool False
indexed NoneType None
n NoneType None
device NoneType None
persistent_workers bool False
pin_memory_device str
wif NoneType None
before_iter NoneType None
after_item NoneType None
before_batch NoneType None
after_batch NoneType None
after_iter NoneType None
create_batches NoneType None
create_item NoneType None
create_batch NoneType None
retain NoneType None
get_idxs NoneType None
sample NoneType None
shuffle_fn NoneType None
do_batch NoneType None

res is the result of sort_func applied on all elements of the dataset. You can pass it if available to make the init much faster by avoiding an initial pass over the whole dataset. For example if sorting by text length (as in the default sort_func, called _default_sort) you should pass a list with the length of each element in dataset to res to take advantage of this speed-up.

To get the same init speed-up for the validation set, val_res (a list of text lengths for your validation set) can be passed to the kwargs argument of SortedDL. Below is an example to reduce the init time by passing a list of text lengths for both the training set and the validation set:

# Pass the training dataset text lengths to SortedDL
srtd_dl=partial(SortedDL, res = train_text_lens)

# Pass the validation dataset text lengths 
dl_kwargs = [{},{'val_res': val_text_lens}]

# init our Datasets 
dsets = Datasets(...)   

# init our Dataloaders
dls = dsets.dataloaders(...,dl_type = srtd_dl, dl_kwargs = dl_kwargs)

If shuffle is True, this will shuffle a bit the results of the sort to have items of roughly the same size in batches, but not in the exact sorted order.

ds = [(tensor([1,2]),1), (tensor([3,4,5,6]),2), (tensor([7]),3), (tensor([8,9,10]),4)]
dl = SortedDL(ds, bs=2, before_batch=partial(pad_input, pad_idx=0))
test_eq(list(dl), [(tensor([[ 3,  4,  5,  6], [ 8,  9, 10,  0]]), tensor([2, 4])), 
                   (tensor([[1, 2], [7, 0]]), tensor([1, 3]))])
ds = [(tensor(range(random.randint(1,10))),i) for i in range(101)]
dl = SortedDL(ds, bs=2, create_batch=partial(pad_input, pad_idx=-1), shuffle=True, num_workers=0)
batches = list(dl)
max_len = len(batches[0][0])
for b in batches: 
    assert(len(b[0])) <= max_len 
    test_ne(b[0][-1], -1)

TransformBlock for text

To use the data block API, you will need this build block for texts.


source

TextBlock

 TextBlock (tok_tfm, vocab=None, is_lm=False, seq_len=72, backwards=False,
            min_freq=3, max_vocab=60000, special_toks=None)

A TransformBlock for texts

For efficient tokenization, you probably want to use one of the factory methods. Otherwise, you can pass your custom tok_tfm that will deal with tokenization (if your texts are already tokenized, you can pass noop), a vocab, or leave it to be inferred on the texts using min_freq and max_vocab.

is_lm indicates if we want to use texts for language modeling or another task, seq_len is only necessary to tune if is_lm=False, and is passed along to pad_input_chunk.


source

TextBlock.from_df

 TextBlock.from_df (text_cols, vocab=None, is_lm=False, seq_len=72,
                    backwards=False, min_freq=3, max_vocab=60000,
                    tok=None, rules=None, sep=' ', n_workers=4,
                    mark_fields=None, tok_text_col='text', **kwargs)

Build a TextBlock from a dataframe using text_cols

Here is an example using a sample of IMDB stored as a CSV file:

path = untar_data(URLs.IMDB_SAMPLE)
df = pd.read_csv(path/'texts.csv')

imdb_clas = DataBlock(
    blocks=(TextBlock.from_df('text', seq_len=72), CategoryBlock),
    get_x=ColReader('text'), get_y=ColReader('label'), splitter=ColSplitter())

dls = imdb_clas.dataloaders(df, bs=64)
dls.show_batch(max_n=2)
text category
0 xxbos xxmaj raising xxmaj victor xxmaj vargas : a xxmaj review \n\n xxmaj you know , xxmaj raising xxmaj victor xxmaj vargas is like sticking your hands into a big , xxunk bowl of xxunk . xxmaj it 's warm and gooey , but you 're not sure if it feels right . xxmaj try as i might , no matter how warm and gooey xxmaj raising xxmaj victor xxmaj vargas became i was always aware that something did n't quite feel right . xxmaj victor xxmaj vargas suffers from a certain xxunk on the director 's part . xxmaj apparently , the director thought that the ethnic backdrop of a xxmaj latino family on the lower east side , and an xxunk storyline would make the film critic proof . xxmaj he was right , but it did n't fool me . xxmaj raising xxmaj victor xxmaj vargas is negative
1 xxbos xxup the xxup shop xxup around xxup the xxup corner is one of the xxunk and most feel - good romantic comedies ever made . xxmaj there 's just no getting around that , and it 's hard to actually put one 's feeling for this film into words . xxmaj it 's not one of those films that tries too hard , nor does it come up with the xxunk possible scenarios to get the two protagonists together in the end . xxmaj in fact , all its charm is xxunk , contained within the characters and the setting and the plot … which is highly believable to xxunk . xxmaj it 's easy to think that such a love story , as beautiful as any other ever told , * could * happen to you … a feeling you do n't often get from other romantic comedies positive

vocab, is_lm, seq_len, min_freq and max_vocab are passed to the main init, the other argument to Tokenizer.from_df.


source

TextBlock.from_folder

 TextBlock.from_folder (path, vocab=None, is_lm=False, seq_len=72,
                        backwards=False, min_freq=3, max_vocab=60000,
                        tok=None, rules=None, extensions=None,
                        folders=None, output_dir=None,
                        skip_if_exists=True, output_names=None,
                        n_workers=4, encoding='utf8', **kwargs)

Build a TextBlock from a path

vocab, is_lm, seq_len, min_freq and max_vocab are passed to the main init, the other argument to Tokenizer.from_folder.


source

TextDataLoaders

 TextDataLoaders (*loaders, path:str|pathlib.Path='.', device=None)

Basic wrapper around several DataLoaders with factory methods for NLP problems

You should not use the init directly but one of the following factory methods. All those factory methods accept as arguments:

  • text_vocab: the vocabulary used for numericalizing texts (if not passed, it’s inferred from the data)
  • tok_tfm: if passed, uses this tok_tfm instead of the default
  • seq_len: the sequence length used for batch
  • bs: the batch size
  • val_bs: the batch size for the validation DataLoader (defaults to bs)
  • shuffle_train: if we shuffle the training DataLoader or not
  • device: the PyTorch device to use (defaults to default_device())

source

TextDataLoaders.from_folder

 TextDataLoaders.from_folder (path, train='train', valid='valid',
                              valid_pct=None, seed=None, vocab=None,
                              text_vocab=None, is_lm=False, tok_tfm=None,
                              seq_len=72, splitter=None, backwards=False,
                              bs:int=64, val_bs:int=None,
                              shuffle:bool=True, device=None)

Create from imagenet style dataset in path with train and valid subfolders (or provide valid_pct)

Type Default Details
path str | Path . Path to put in DataLoaders
train str train
valid str valid
valid_pct NoneType None
seed NoneType None
vocab NoneType None
text_vocab NoneType None
is_lm bool False
tok_tfm NoneType None
seq_len int 72
splitter NoneType None
backwards bool False
bs int 64 Size of batch
val_bs int None Size of batch for validation DataLoader
shuffle bool True Whether to shuffle data
device NoneType None Device to put DataLoaders

If valid_pct is provided, a random split is performed (with an optional seed) by setting aside that percentage of the data for the validation set (instead of looking at the grandparents folder). If a vocab is passed, only the folders with names in vocab are kept.

Here is an example on a sample of the IMDB movie review dataset:

path = untar_data(URLs.IMDB)
dls = TextDataLoaders.from_folder(path)
dls.show_batch(max_n=3)
text category
0 xxbos xxmaj match 1 : xxmaj tag xxmaj team xxmaj table xxmaj match xxmaj bubba xxmaj ray and xxmaj spike xxmaj dudley vs xxmaj eddie xxmaj guerrero and xxmaj chris xxmaj benoit xxmaj bubba xxmaj ray and xxmaj spike xxmaj dudley started things off with a xxmaj tag xxmaj team xxmaj table xxmaj match against xxmaj eddie xxmaj guerrero and xxmaj chris xxmaj benoit . xxmaj according to the rules of the match , both opponents have to go through tables in order to get the win . xxmaj benoit and xxmaj guerrero heated up early on by taking turns hammering first xxmaj spike and then xxmaj bubba xxmaj ray . a xxmaj german xxunk by xxmaj benoit to xxmaj bubba took the wind out of the xxmaj dudley brother . xxmaj spike tried to help his brother , but the referee restrained him while xxmaj benoit and xxmaj guerrero pos
1 xxbos xxmaj okay , so xxmaj i 'm not a big video game buff , but was the game xxmaj house of the xxmaj dead really famous enough to make a movie from ? xxmaj sure , they went as far as to actually put in quick video game clips throughout the movie , as though justifying any particular scene of violence , but there are dozens and dozens of games that look exactly the same , with the hand in the bottom on the screen , supposedly your own , holding whatever weapon and goo - ing all kinds of aliens or walking dead or snipers or whatever the case may be . \n\n xxmaj it 's an interesting premise in xxmaj house of the xxmaj dead , with a lot of college kids ( loaded college kids , as it were , kids who are able to pay neg
2 xxbos xxup anchors xxup aweigh sees two eager young sailors , xxmaj joe xxmaj brady ( gene xxmaj kelly ) and xxmaj clarence xxmaj doolittle / xxmaj brooklyn ( frank xxmaj sinatra ) , get a special four - day shore leave . xxmaj eager to get to the girls , particularly xxmaj joe 's xxmaj lola , neither xxmaj joe nor xxmaj brooklyn figure on the interruption of little xxmaj navy - mad xxmaj donald ( dean xxmaj stockwell ) and his xxmaj aunt xxmaj susie ( kathryn xxmaj grayson ) . xxmaj unexperienced in the ways of females and courting , xxmaj brooklyn quickly enlists xxmaj joe to help him win xxmaj aunt xxmaj susie over . xxmaj along the way , however , xxmaj joe finds himself falling for the gal he thinks belongs to his best friend . xxmaj how is xxmaj brooklyn going to take pos

source

TextDataLoaders.from_df

 TextDataLoaders.from_df (df, path='.', valid_pct=0.2, seed=None,
                          text_col=0, label_col=1, label_delim=None,
                          y_block=None, text_vocab=None, is_lm=False,
                          valid_col=None, tok_tfm=None,
                          tok_text_col='text', seq_len=72,
                          backwards=False, bs:int=64, val_bs:int=None,
                          shuffle:bool=True, device=None)

Create from df in path with valid_pct

Type Default Details
df
path str | Path . Path to put in DataLoaders
valid_pct float 0.2
seed NoneType None
text_col int 0
label_col int 1
label_delim NoneType None
y_block NoneType None
text_vocab NoneType None
is_lm bool False
valid_col NoneType None
tok_tfm NoneType None
tok_text_col str text
seq_len int 72
backwards bool False
bs int 64 Size of batch
val_bs int None Size of batch for validation DataLoader
shuffle bool True Whether to shuffle data
device NoneType None Device to put DataLoaders

seed can optionally be passed for reproducibility. text_col, label_col and optionally valid_col are indices or names of columns for texts/labels and the validation flag. label_delim can be passed for a multi-label problem if your labels are in one column, separated by a particular char. y_block should be passed to indicate your type of targets, in case the library did no infer it properly.

Along with this, you can specify the specific column the tokenized text are sent to with tok_text_col. By default they are stored in a column named text after tokenizing.

Here are examples on subsets of IMDB:

path = untar_data(URLs.IMDB_SAMPLE)
df = pd.read_csv(path/"texts.csv"); df.head()
label text is_valid
0 negative Un-bleeping-believable! Meg Ryan doesn't even look her usual pert lovable self in this, which normally makes me forgive her shallow ticky acting schtick. Hard to believe she was the producer on this dog. Plus Kevin Kline: what kind of suicide trip has his career been on? Whoosh... Banzai!!! Finally this was directed by the guy who did Big Chill? Must be a replay of Jonestown - hollywood style. Wooofff! False
1 positive This is a extremely well-made film. The acting, script and camera-work are all first-rate. The music is good, too, though it is mostly early in the film, when things are still relatively cheery. There are no really superstars in the cast, though several faces will be familiar. The entire cast does an excellent job with the script.<br /><br />But it is hard to watch, because there is no good end to a situation like the one presented. It is now fashionable to blame the British for setting Hindus and Muslims against each other, and then cruelly separating them into two countries. There is som... False
2 negative Every once in a long while a movie will come along that will be so awful that I feel compelled to warn people. If I labor all my days and I can save but one soul from watching this movie, how great will be my joy.<br /><br />Where to begin my discussion of pain. For starters, there was a musical montage every five minutes. There was no character development. Every character was a stereotype. We had swearing guy, fat guy who eats donuts, goofy foreign guy, etc. The script felt as if it were being written as the movie was being shot. The production value was so incredibly low that it felt li... False
3 positive Name just says it all. I watched this movie with my dad when it came out and having served in Korea he had great admiration for the man. The disappointing thing about this film is that it only concentrate on a short period of the man's life - interestingly enough the man's entire life would have made such an epic bio-pic that it is staggering to imagine the cost for production.<br /><br />Some posters elude to the flawed characteristics about the man, which are cheap shots. The theme of the movie "Duty, Honor, Country" are not just mere words blathered from the lips of a high-brassed offic... False
4 negative This movie succeeds at being one of the most unique movies you've seen. However this comes from the fact that you can't make heads or tails of this mess. It almost seems as a series of challenges set up to determine whether or not you are willing to walk out of the movie and give up the money you just paid. If you don't want to feel slighted you'll sit through this horrible film and develop a real sense of pity for the actors involved, they've all seen better days, but then you realize they actually got paid quite a bit of money to do this and you'll lose pity for them just like you've alr... False
path = untar_data(URLs.IMDB_SAMPLE)
df = pd.read_csv(path/"texts.csv")
dls = TextDataLoaders.from_df(df, path=path, text_col='text', label_col='label', valid_col='is_valid')
dls.show_batch(max_n=3)
text category
0 xxbos xxmaj raising xxmaj victor xxmaj vargas : a xxmaj review \n\n xxmaj you know , xxmaj raising xxmaj victor xxmaj vargas is like sticking your hands into a big , xxunk bowl of xxunk . xxmaj it 's warm and gooey , but you 're not sure if it feels right . xxmaj try as i might , no matter how warm and gooey xxmaj raising xxmaj victor xxmaj vargas became i was always aware that something did n't quite feel right . xxmaj victor xxmaj vargas suffers from a certain xxunk on the director 's part . xxmaj apparently , the director thought that the ethnic backdrop of a xxmaj latino family on the lower east side , and an xxunk storyline would make the film critic proof . xxmaj he was right , but it did n't fool me . xxmaj raising xxmaj victor xxmaj vargas is negative
1 xxbos xxup the xxup shop xxup around xxup the xxup corner is one of the xxunk and most feel - good romantic comedies ever made . xxmaj there 's just no getting around that , and it 's hard to actually put one 's feeling for this film into words . xxmaj it 's not one of those films that tries too hard , nor does it come up with the xxunk possible scenarios to get the two protagonists together in the end . xxmaj in fact , all its charm is xxunk , contained within the characters and the setting and the plot … which is highly believable to xxunk . xxmaj it 's easy to think that such a love story , as beautiful as any other ever told , * could * happen to you … a feeling you do n't often get from other romantic comedies positive
2 xxbos xxmaj now that xxmaj che(2008 ) has finished its relatively short xxmaj australian cinema run ( extremely limited xxunk screen in xxmaj xxunk , after xxunk ) , i can xxunk join both xxunk of " at xxmaj the xxmaj movies " in taking xxmaj steven xxmaj soderbergh to task . \n\n xxmaj it 's usually satisfying to watch a film director change his style / subject , but xxmaj soderbergh 's most recent stinker , xxmaj the xxmaj girlfriend xxmaj xxunk ) , was also missing a story , so narrative ( and editing ? ) seem to suddenly be xxmaj soderbergh 's main challenge . xxmaj strange , after 20 - odd years in the business . xxmaj he was probably never much good at narrative , just xxunk it well inside " edgy " projects . \n\n xxmaj none of this excuses him this present , negative
dls = TextDataLoaders.from_df(df, path=path, text_col='text', is_lm=True, valid_col='is_valid')
dls.show_batch(max_n=3)
text text_
0 xxbos xxmaj critics need to review what they class as a quality movie . i think the critics have seen too many actions films and have xxunk to the xxmaj matrix style of films . xxmaj xxunk is a breath of fresh air , a film with so many layers that one viewing is not enough to understand or appreciate this outstanding film . xxmaj xxunk von xxmaj xxunk shows that old xxmaj critics need to review what they class as a quality movie . i think the critics have seen too many actions films and have xxunk to the xxmaj matrix style of films . xxmaj xxunk is a breath of fresh air , a film with so many layers that one viewing is not enough to understand or appreciate this outstanding film . xxmaj xxunk von xxmaj xxunk shows that old styles
1 xxmaj xxunk is something ) , but noticeable moments of xxunk as he still struggles to find his humanity . xxmaj this xxunk of his for a real life could get boring , and almost did in xxmaj supremacy , but just works better in xxmaj ultimatum ( better script ) . \n\n i am reminded of a scene in " xxunk " ( the only good xxmaj pierce xxmaj xxunk xxmaj xxunk is something ) , but noticeable moments of xxunk as he still struggles to find his humanity . xxmaj this xxunk of his for a real life could get boring , and almost did in xxmaj supremacy , but just works better in xxmaj ultimatum ( better script ) . \n\n i am reminded of a scene in " xxunk " ( the only good xxmaj pierce xxmaj xxunk xxmaj bond
2 xxmaj mr . xxmaj julia , played his role equally as perfect . xxmaj it was interesting to see how reluctant xxmaj richard xxmaj dreyfuss was in replacing the dictator against his will . xxmaj but he became more confident and comfortable with the role as time passed . xxmaj since everything happens for a reason in life , i believe he was forced to replace the dictator because he was meant mr . xxmaj julia , played his role equally as perfect . xxmaj it was interesting to see how reluctant xxmaj richard xxmaj dreyfuss was in replacing the dictator against his will . xxmaj but he became more confident and comfortable with the role as time passed . xxmaj since everything happens for a reason in life , i believe he was forced to replace the dictator because he was meant to

source

TextDataLoaders.from_csv

 TextDataLoaders.from_csv (path, csv_fname='labels.csv', header='infer',
                           delimiter=None, valid_pct=0.2, seed=None,
                           text_col=0, label_col=1, label_delim=None,
                           y_block=None, text_vocab=None, is_lm=False,
                           valid_col=None, tok_tfm=None,
                           tok_text_col='text', seq_len=72,
                           backwards=False, bs:int=64, val_bs:int=None,
                           shuffle:bool=True, device=None)

Create from csv file in path/csv_fname

Type Default Details
path str | Path . Path to put in DataLoaders
csv_fname str labels.csv
header str infer
delimiter NoneType None
valid_pct float 0.2
seed NoneType None
text_col int 0
label_col int 1
label_delim NoneType None
y_block NoneType None
text_vocab NoneType None
is_lm bool False
valid_col NoneType None
tok_tfm NoneType None
tok_text_col str text
seq_len int 72
backwards bool False
bs int 64 Size of batch
val_bs int None Size of batch for validation DataLoader
shuffle bool True Whether to shuffle data
device NoneType None Device to put DataLoaders

Opens the csv file with header and delimiter, then pass all the other arguments to TextDataLoaders.from_df.

dls = TextDataLoaders.from_csv(path=path, csv_fname='texts.csv', text_col='text', label_col='label', valid_col='is_valid')
dls.show_batch(max_n=3)
text category
0 xxbos xxmaj raising xxmaj victor xxmaj vargas : a xxmaj review \n\n xxmaj you know , xxmaj raising xxmaj victor xxmaj vargas is like sticking your hands into a big , xxunk bowl of xxunk . xxmaj it 's warm and gooey , but you 're not sure if it feels right . xxmaj try as i might , no matter how warm and gooey xxmaj raising xxmaj victor xxmaj vargas became i was always aware that something did n't quite feel right . xxmaj victor xxmaj vargas suffers from a certain xxunk on the director 's part . xxmaj apparently , the director thought that the ethnic backdrop of a xxmaj latino family on the lower east side , and an xxunk storyline would make the film critic proof . xxmaj he was right , but it did n't fool me . xxmaj raising xxmaj victor xxmaj vargas is negative
1 xxbos xxup the xxup shop xxup around xxup the xxup corner is one of the xxunk and most feel - good romantic comedies ever made . xxmaj there 's just no getting around that , and it 's hard to actually put one 's feeling for this film into words . xxmaj it 's not one of those films that tries too hard , nor does it come up with the xxunk possible scenarios to get the two protagonists together in the end . xxmaj in fact , all its charm is xxunk , contained within the characters and the setting and the plot … which is highly believable to xxunk . xxmaj it 's easy to think that such a love story , as beautiful as any other ever told , * could * happen to you … a feeling you do n't often get from other romantic comedies positive
2 xxbos xxmaj now that xxmaj che(2008 ) has finished its relatively short xxmaj australian cinema run ( extremely limited xxunk screen in xxmaj xxunk , after xxunk ) , i can xxunk join both xxunk of " at xxmaj the xxmaj movies " in taking xxmaj steven xxmaj soderbergh to task . \n\n xxmaj it 's usually satisfying to watch a film director change his style / subject , but xxmaj soderbergh 's most recent stinker , xxmaj the xxmaj girlfriend xxmaj xxunk ) , was also missing a story , so narrative ( and editing ? ) seem to suddenly be xxmaj soderbergh 's main challenge . xxmaj strange , after 20 - odd years in the business . xxmaj he was probably never much good at narrative , just xxunk it well inside " edgy " projects . \n\n xxmaj none of this excuses him this present , negative