load JSON files, get the errors

[PatricYan]

Hi, does this bug be fixed? when I load JSON files, I get the same errors by the command
!python3 run.py --do_train --task qa --dataset squad-retrain-data/train-v2.0.json --output_dir ./re_trained_model/

change the dateset to load json by refering to https://huggingface.co/docs/datasets/loading.html
dataset = datasets.load_dataset('json', data_files=args.dataset)

Errors:
Downloading and preparing dataset json/default (download: Unknown size, generated: Unknown size, post-processed: Unknown size, total: Unknown size) to /root/.cache/huggingface/datasets/json/default-c1e124ad488911b8/0.0.0/45636811569ec4a6630521c18235dfbbab83b7ab572e3393c5ba68ccabe98264...

Originally posted by @yanllearnn in #730 (comment)

[lhoestq]

Hi ! The message you're getting is not an error. It simply says that your JSON dataset is being prepared to a location in /root/.cache/huggingface/datasets

[PatricYan]

but I want to load local JSON file by command
python3 run.py --do_train --task qa --dataset squad-retrain-data/train-v2.0.json --output_dir ./re_trained_model/

squad-retrain-data/train-v2.0.json is the local JSON file, how to load it and map it to a special structure?

[lhoestq]

You can load it with dataset = datasets.load_dataset('json', data_files=args.dataset) as you said.
Then if you need to apply additional processing to map it to a special structure, you can use rename columns or use dataset.map. For more information, you can check the documentation here: https://huggingface.co/docs/datasets/process.html

Also feel free to share your run.py code so we can take a look

[PatricYan]

# Dataset selection
    if args.dataset.endswith('.json') or args.dataset.endswith('.jsonl'):
        dataset_id = None
        # Load from local json/jsonl file
        dataset = datasets.load_dataset('json', data_files=args.dataset)
        # By default, the "json" dataset loader places all examples in the train split,
        # so if we want to use a jsonl file for evaluation we need to get the "train" split
        # from the loaded dataset
        eval_split = 'train'
    else:
        default_datasets = {'qa': ('squad',), 'nli': ('snli',)}
        dataset_id = tuple(args.dataset.split(':')) if args.dataset is not None else \
            default_datasets[args.task]
        # MNLI has two validation splits (one with matched domains and one with mismatched domains). Most datasets just have one "validation" split
        eval_split = 'validation_matched' if dataset_id == ('glue', 'mnli') else 'validation'
        # Load the raw data
        dataset = datasets.load_dataset(*dataset_id)

I want to load JSON squad dataset instead dataset = datasets.load_dataset('squad') to retrain the model.

[lhoestq]

If your JSON has the same format as the SQuAD dataset, then you need to pass field="data" to load_dataset, since the SQuAD format is one big JSON object in which the "data" field contains the list of questions and answers.

dataset = datasets.load_dataset('json', data_files=args.dataset, field="data")

Let me know if that helps :)

[PatricYan]

Yes, code works. but the format is not as expected.

dataset = datasets.load_dataset('json', data_files=args.dataset, field="data")

python3 run.py --do_train --task qa --dataset squad --output_dir ./re_trained_model/

************ train_dataset: Dataset({
features: ['id', 'title', 'context', 'question', 'answers'],
num_rows: 87599
})

python3 run.py --do_train --task qa --dataset squad-retrain-data/train-v2.0.json --output_dir ./re_trained_model/

************ train_dataset: Dataset({
features: ['title', 'paragraphs'],
num_rows: 442
})

I want the JSON to have the same format as before features. https://github.com/huggingface/datasets/blob/master/datasets/squad_v2/squad_v2.py is the script dealing with squad but how can I apply it by using JSON?

[lhoestq]

Ok I see, you have the paragraphs so you just need to process them to extract the questions and answers. I think you can process the SQuAD-like data this way:

def process_squad(articles):
    out = {
        "title": [],
        "context": [],
        "question": [],
        "id": [],
        "answers": [],
    }
    for title, paragraphs in zip(articles["title"], articles["paragraphs"]):
        for paragraph in paragraphs:
            for qa in paragraph["qas"]:
                out["title"].append(title)
                out["context"].append(paragraph["context"])
                out["question"].append(qa["question"])
                out["id"].append(qa["id"])
                out["answers"].append({
                    "answer_start": [answer["answer_start"] for answer in qa["answers"]],
                    "text": [answer["text"] for answer in qa["answers"]],
                })
    return out

dataset = dataset.map(process_squad, batched=True, remove_columns=["paragraphs"])

I adapted the code from squad.py. The code takes as input a batch of articles (title + paragraphs) and gets all the questions and answers from the JSON structure.

The output is a dataset with features: ['answers', 'context', 'id', 'question', 'title']

Let me know if that helps !

[PatricYan]

Yes, this works. But how to get the training output during training the squad by Trainer
for example https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/trainer_qa.py
I want the training inputs, labels, outputs for every epoch and step to produce the training dynamic graph

[lhoestq]

I think you may need to implement your own Trainer, from the QuestionAnsweringTrainer for example.
This way you can have the flexibility of saving all the inputs/output used at each step

[PatricYan]

does there have any function to be overwritten to do this?

[PatricYan]

does there have any function to be overwritten to do this?

ok, I overwrote the compute_loss, thank you.

[PatricYan]

Hi, I add one field example_id, but I can't see it in the comput_loss function, how can I do this? below is the information of inputs

*********************** inputs: {'attention_mask': tensor([[1, 1, 1,  ..., 0, 0, 0],
        [1, 1, 1,  ..., 0, 0, 0],
        [1, 1, 1,  ..., 0, 0, 0],
        ...,
        [1, 1, 1,  ..., 0, 0, 0],
        [1, 1, 1,  ..., 0, 0, 0],
        [1, 1, 1,  ..., 0, 0, 0]], device='cuda:0'), 'end_positions': tensor([ 25,  97,  93,  44,  25, 112, 109, 134], device='cuda:0'), 'input_ids': tensor([[ 101, 2054, 2390,  ...,    0,    0,    0],
        [ 101, 2054, 2515,  ...,    0,    0,    0],
        [ 101, 2054, 2106,  ...,    0,    0,    0],
        ...,
        [ 101, 2339, 2001,  ...,    0,    0,    0],
        [ 101, 2054, 2515,  ...,    0,    0,    0],
        [ 101, 2054, 2003,  ...,    0,    0,    0]], device='cuda:0'), 'start_positions': tensor([ 20,  90,  89,  41,  25,  96, 106, 132], device='cuda:0'), 'token_type_ids': tensor([[0, 0, 0,  ..., 0, 0, 0],
        [0, 0, 0,  ..., 0, 0, 0],
        [0, 0, 0,  ..., 0, 0, 0],
        ...,
        [0, 0, 0,  ..., 0, 0, 0],
        [0, 0, 0,  ..., 0, 0, 0],
        [0, 0, 0,  ..., 0, 0, 0]], device='cuda:0')} 

# This function preprocesses a question answering dataset, tokenizing the question and context text
# and finding the right offsets for the answer spans in the tokenized context (to use as labels).
# Adapted from https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_qa.py
def prepare_train_dataset_qa(examples, tokenizer, max_seq_length=None):
    questions = [q.lstrip() for q in examples["question"]]
    max_seq_length = tokenizer.model_max_length
    # tokenize both questions and the corresponding context
    # if the context length is longer than max_length, we split it to several
    # chunks of max_length
    tokenized_examples = tokenizer(
        questions,
        examples["context"],
        truncation="only_second",
        max_length=max_seq_length,
        stride=min(max_seq_length // 2, 128),
        return_overflowing_tokens=True,
        return_offsets_mapping=True,
        padding="max_length"
    )

    # Since one example might give us several features if it has a long context,
    # we need a map from a feature to its corresponding example.
    sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping")
    # The offset mappings will give us a map from token to character position
    # in the original context. This will help us compute the start_positions
    # and end_positions to get the final answer string.
    offset_mapping = tokenized_examples.pop("offset_mapping")

    tokenized_examples["start_positions"] = []
    tokenized_examples["end_positions"] = []

    tokenized_examples["example_id"] = []

    for i, offsets in enumerate(offset_mapping):
        input_ids = tokenized_examples["input_ids"][i]
        # We will label features not containing the answer the index of the CLS token.
        cls_index = input_ids.index(tokenizer.cls_token_id)
        sequence_ids = tokenized_examples.sequence_ids(i)
        # from the feature idx to sample idx
        sample_index = sample_mapping[i]
        # get the answer for a feature
        answers = examples["answers"][sample_index]

        tokenized_examples["example_id"].append(examples["id"][sample_index])

        if len(answers["answer_start"]) == 0:
            tokenized_examples["start_positions"].append(cls_index)
            tokenized_examples["end_positions"].append(cls_index)
        else:
            # Start/end character index of the answer in the text.
            start_char = answers["answer_start"][0]
            end_char = start_char + len(answers["text"][0])

            # Start token index of the current span in the text.
            token_start_index = 0
            while sequence_ids[token_start_index] != 1:
                token_start_index += 1

            # End token index of the current span in the text.
            token_end_index = len(input_ids) - 1
            while sequence_ids[token_end_index] != 1:
                token_end_index -= 1

            # Detect if the answer is out of the span (in which case this feature is labeled with the CLS index).
            if not (offsets[token_start_index][0] <= start_char and
                    offsets[token_end_index][1] >= end_char):
                tokenized_examples["start_positions"].append(cls_index)
                tokenized_examples["end_positions"].append(cls_index)
            else:
                # Otherwise move the token_start_index and token_end_index to the two ends of the answer.
                # Note: we could go after the last offset if the answer is the last word (edge case).
                while token_start_index < len(offsets) and \
                        offsets[token_start_index][0] <= start_char:
                    token_start_index += 1
                tokenized_examples["start_positions"].append(
                    token_start_index - 1)
                while offsets[token_end_index][1] >= end_char:
                    token_end_index -= 1
                tokenized_examples["end_positions"].append(token_end_index + 1)

    return tokenized_examples