textbook 0.0.4

Text classification datasets

Textbook: Universal NLP Datasets

Dependency

• av==6.2.0
• jsonnet==0.14.0
• opencv_python==4.1.1.26
• torch==1.3.1
• torchvision==0.4.2
• numpy==1.17.4

Download raw datasets

bash fetch.sh

It downloads alphanli, hellaswag, physicaliqa and socialiqa from AWS.

Usage

Load multichoice dataset

from textbook.config import *
from textbook.transforms_video import *

from transformers import GPT2Tokenizer

text_renderer = TextRenderer(
    tokenizer=GPT2Tokenizer.from_pretrained('distilgpt2'),
    special_tokens={'cls_token': '[CLS]', 'pad_token': '[PAD]', 'mask_token': '[MASK]'},
)
config_alphanli = Configuration(alphanli_config)
alphanli_dataset = ClassificationDataset("data_cache/alphanli/eval.jsonl", config_alphanli, [text_renderer])

alphanli_dev_dataloader: DataLoader = iter(
    DataLoader(
        alphanli_dataset, batch_sampler=DynamicBatchSampler(alphanli_dataset),
        collate_fn=collate_fn))

Load multimodal dataset

upscale_size = int(84 * 1.1)
transform_pre = ComposeMix([
    [Scale(upscale_size), "img"],
    [RandomCropVideo(84), "vid"],
])

transform_post = ComposeMix([
    [torchvision.transforms.ToTensor(), "img"],
])

video_renderer = VisionRenderer(
    nframe=72,
    nclip=1,
    nstep=2,
    transform_pre=transform_pre,
    transform_post=transform_post,
    data_dir="data_cache/smthsmth/20bn-something-something-v2"
)

config_smthsmth = Configuration(smthsmth_config)

smthsmth_dataset = ClassificationVisionDataset(
    "data_cache/smthsmth/something-something-v2-validation.json", config_smthsmth, [text_renderer, video_renderer])

smthsmth_dev_dataloader: DataLoader = iter(DataLoader(
    smthsmth_dataset, batch_size=16, collate_fn=lambda x: collate_fn(
        x, mlm=True, mlm_probability=0.15, tokenizer=text_renderer.tokenizer)))

Let's multitask

multitask_dataloader = MultiTaskDataset([alphanli_dev_dataloader, smthsmth_dev_dataloader])

# alternate through different dataloaders
for batch in multitask_dataloader:
    print(batch["input_ids"].shape)
    print(batch["images"].shape)