Problems in deployment when I configure my own labels

Amazon SageMaker
1

Hi.

I am training a binary classification model based on this model checkpoint: dccuchile/bert-base-spanish-wwm-cased

When deployment, it uses by default these labels: LABEL_0, LABEL_1.
My goal is to deploy it with my own labels. Browsing the forum I came across this thread, and it seems that resolve a similiar problem.

After read this thread, my code looks like:

label2id = {
        "0": "goodUser",
        "1": "badUserFraud"
    }
    
    id2label = {
        "goodUser": 0,
        "badUserFraud": 1
    }

    # download model from model hub
    config = AutoConfig.from_pretrained(args.model_name, label2id=label2id, id2label=id2label)
    model = AutoModelForSequenceClassification.from_pretrained(args.model_name, config=config)
    tokenizer = AutoTokenizer.from_pretrained(args.model_name)

The training step going well, but when deployment step I find this error:

image
image1632Ă—552 101 KB

Thanks in advance

2

I would like to know this aswell

3

Hey @Oigres,

Your configuration is vice-versa.
If you take a look at config.json · distilbert-base-uncased-finetuned-sst-2-english at main
So the id2label needs to have as key the id and as value the string.
And the label2id needs to have as key the string and as value the id .

id2label = {
        "0": "goodUser",
        "1": "badUserFraud"
    }

label2id = {
        "goodUser": 0,
        "badUserFraud": 1
    }

Easy to remember is the name is always key2Value.

3 Likes
4

Ohh! Seems true that the best debugging technique is taking a rest :sweat_smile:

Thank you very much

1 Like
5

thanks for the awesome information.

1 Like
6

thanks my issue has been fixed.

1 Like
7

Hi, I have a doubt here.

If I have already converted my labels to numbers and want to use the same, should I still create label2id and id2label?

I am trying to perform sequence classification in binary and multi-class settings.

Any discussion to help me understand this is greatly appreciated.

Thanks