Most of the items in show_batch is xxpad strings

fastai
1

I am facing an issue while loading the data from a custom csv file. Almost all are xxpad strings. I have created a vocab list from a file and i was trying to use the vocab on another csv file for multi category classification.
The multi category classification csv file is in the following format -

text, label_1, label_2, label_3, label_4, label_5, label_6.
“hello this is cool”, 1, 0, 0, 1, 0, 0

dls_lm is the language model previously trained and loaded.

dls_labels = DataBlock(
    blocks=(TextBlock.from_df('text',vocab=dls_lm.vocab), MultiCategoryBlock),
    get_x=ColReader("text"),
    get_y=ColReader([1,2,3,4,5,6]),
    splitter=RandomSplitter(0.2)
).dataloaders(train_label, bs=128, seq_len=72)

when I run dls_labels.show_batch() in first row’s first column is a proper string (After tokenizing) from the csv and in the Y column it’s the semi colon seperated values from label_x columns (1 or 0). After that all the rows have repeated xxpad strings and y column value same as first one.

What is I am doing wrong here ? Please help me understand ?

2

Hey Ted!

The short version is that it’s a known issue with show_batch - your dataloaders are probably fine, but since batches are sorted by length, the longest item might be batched together with shorter items thus resulting in show_batch showing all this padding. Try training and if it works then most likely everything is fine :slight_smile:

1 Like
3

Hey orendar,
thanks for the reply. my training was not working as expcted so I was trying to find any errors in thre previous steps. And the dataloader part looked dubious. Now I am back to square one. I am doubting the loading of label data is wrong here. In my data set a “1” in the column label_x means the text has classified as a particular label and 0 means it’s not.

I think the way I am loading the Dataloader is it wrong ? Because in the show batch the y column only has values (0;1). What is the proper way to read this type of csv for multi class classification, I have the list of column names (label_x) ? So based on the value in column choose the correct label from list.

4

Hey Ted,

If you are doing multilabel text classification, try something like this (I am not sure about the parentheses for the block and you can add your own metrics):
dls_clas = TextDataLoaders.from_df(df=df, valid_pct=0.2, seed=42, text_col='text', text_vocab=vocab, bs=bs, y_names=labels, y_block=MultiCategoryBlock(), label_col=labels)
learn = text_classifier_learner(dls_clas, AWD_LSTM, drop_mult=0.5, n_out=len(labels), metrics=[])

1 Like
5

Hey thanks, But still it’s taking the values inside the columns instead of the labels from list. Producing a dls like following

data text2, 0;1
data text1, 0:1

Is it mandatory to put the category name in the columns values instead of a flag 1/0 ? I have tried with another dummy csv file which has category name in the column itself instead of 1/0 flag which worked.

6

Hey Ted,

I don’t understand your last comment. Could you please try my method (with “labels” being a list of your label columns) and paste the full code together with the problem you’re trying to solve? I work a lot with multilabel text so hopefully I can provide some assistance there :slight_smile:

7

Hey Orendar, I had control over the data source. So I made my csv data in the format :
"text", "labels"
"text_body", "semi column seperated labels, ie label1;label2;label3".
I was able to do the create dataloader and do the classificaiton with out any hiccups with some parameter changes in the code you have provided.

dls_labels = TextDataLoaders.from_df(
    df=labeled_data, 
    valid_pct=0.2, 
    text_vocab=dls_lm.vocab,
    seed=42, 
    text_col='text', 
    label_col='labels', 
    label_delim=";", 
    y_block=MultiCategoryBlock, 
 )

Earlier my data was in following format, 1 in a label column represent whether that label is associated with that text. In this format I was not able to get it working with the code you have have provided. -

"text", "label1", "label2", "label3"
"text body", 1, 0, 0

8

Hey Ted,

Glad you managed to get it to work - feel to mark this issue as resolved if you’re happy with the outcome :slight_smile:

Otherwise if you ever want to debug the previous version of your data, feel free to tag me - pretty sure I managed to train models on dataframes in that format before.

9

Just hit the same issue today, and here is how I was able to look into the entire sample (link to the post I commented on):