How to process the dataset for layoutLM3

[talhaanwarch]

Hi. I want to apply LayoutLMv3Processor on huggingface dataset.
This is the snapshot of how my huggingface dataset look like

Each feature is a list such as words is a list, bbox is a list of list, ner tag is a list of integers representing the labels and image is a string showing path of image.

This is how i am applying processor

# we need to define custom features
features = datasets.Features(
    {
        'pixel_values': datasets.Array3D(dtype="float32", shape=(3, 224, 224)),
        "input_ids": datasets.Sequence(datasets.Value(dtype="int64")),
        "attention_mask": datasets.Sequence(datasets.Value(dtype="int64")),
        #"token_type_ids": datasets.Sequence(datasets.Value(dtype="int64")),
        "bbox": datasets.Array2D(dtype="int64", shape=(512, 4)),
        "labels": datasets.Sequence(datasets.ClassLabel(names=labels)),
    }
)

# preprocess function to perpare into the correct format for the model
def process(sample, processor=None):
    
#     print('sample',sample)
    
    encoding = processor(
        Image.open(sample["image_path"]).convert("RGB"),
        sample["words"],
        boxes=sample["bboxes"],
        word_labels=sample["ner_tags"],
        padding="max_length",
        truncation=True,
    )
    
    #the encodings are set, i converted them to dictionary
    #https://github.com/huggingface/datasets/issues/5025
    new_encoding={}
    for k in encoding.items():
        new_encoding[k[0]]=k[1][0]
    
    
#     del encoding["image"]
#     print('encoding',encoding['pixel_values'][0].max(),encoding['pixel_values'][0].min())
    print(new_encoding)
    return new_encoding


# process the dataset and format it to pytorch
proc_dataset = raw_datasets.map(
    partial(process, processor=processor),
    remove_columns=["image_path", "words", "ner_tags",  "bboxes"],
#     features=features,
    batched=False,
).with_format("torch")

But it is throwing this error

TypeError: Couldn't cast array of type
int64
to
Sequence(feature=Value(dtype='int64', id=None), length=-1, id=None)

Second I did not understand how Features work. Like Array3D knew that on which column it has to act, or on whiich column Sequence should be applied. In short how these features are mapped on original data columns