app.py

import gradio as gr
import datasets
import huggingface_hub
import os
import time
import subprocess
import logging

import json

from transformers.pipelines import TextClassificationPipeline

from text_classification import text_classification_fix_column_mapping


HF_REPO_ID = 'HF_REPO_ID'
HF_SPACE_ID = 'SPACE_ID'
HF_WRITE_TOKEN = 'HF_WRITE_TOKEN'


theme = gr.themes.Soft(
    primary_hue="green",
)

def check_model(model_id):
    try:
        task = huggingface_hub.model_info(model_id).pipeline_tag
    except Exception:
        return None, None

    try:
        from transformers import pipeline
        ppl = pipeline(task=task, model=model_id)
        
        return model_id, ppl
    except Exception as e:
        return model_id, e


def check_dataset(dataset_id, dataset_config="default", dataset_split="test"):
    try:
        configs = datasets.get_dataset_config_names(dataset_id)
    except Exception:
        # Dataset may not exist
        return None, dataset_config, dataset_split

    if dataset_config not in configs:
        # Need to choose dataset subset (config)
        return dataset_id, configs, dataset_split

    ds = datasets.load_dataset(dataset_id, dataset_config)

    if isinstance(ds, datasets.DatasetDict):
        # Need to choose dataset split
        if dataset_split not in ds.keys():
            return dataset_id, None, list(ds.keys())
    elif not isinstance(ds, datasets.Dataset):
        # Unknown type
        return dataset_id, None, None
    return dataset_id, dataset_config, dataset_split


def try_validate(model_id, dataset_id, dataset_config, dataset_split, column_mapping):
    # Validate model
    m_id, ppl = check_model(model_id=model_id)
    if m_id is None:
        gr.Warning(f'Model "{model_id}" is not accessible. Please set your HF_TOKEN if it is a private model.')
        return (
            dataset_config, dataset_split,
            gr.update(interactive=False),   # Submit button
            gr.update(visible=False),       # Model prediction preview
            gr.update(visible=False),       # Label mapping preview
            gr.update(visible=True),        # Column mapping
        )
    if isinstance(ppl, Exception):
        gr.Warning(f'Failed to load "{model_id} model": {ppl}')
        return (
            dataset_config, dataset_split,
            gr.update(interactive=False),   # Submit button
            gr.update(visible=False),       # Model prediction preview
            gr.update(visible=False),       # Label mapping preview
            gr.update(visible=True),        # Column mapping
        )

    # Validate dataset
    d_id, config, split = check_dataset(dataset_id=dataset_id, dataset_config=dataset_config, dataset_split=dataset_split)

    dataset_ok = False
    if d_id is None:
        gr.Warning(f'Dataset "{dataset_id}" is not accessible. Please set your HF_TOKEN if it is a private dataset.')
    elif isinstance(config, list):
        gr.Warning(f'Dataset "{dataset_id}" does not have "{dataset_config}" config. Please choose a valid config.')
        config = gr.update(choices=config, value=config[0])
    elif isinstance(split, list):
        gr.Warning(f'Dataset "{dataset_id}" does not have "{dataset_split}" split. Please choose a valid split.')
        split = gr.update(choices=split, value=split[0])
    else:
        dataset_ok = True

    if not dataset_ok:
        return (
            config, split,
            gr.update(interactive=False),   # Submit button
            gr.update(visible=False),       # Model prediction preview
            gr.update(visible=False),       # Label mapping preview
            gr.update(visible=True),        # Column mapping
        )

    # TODO: Validate column mapping by running once
    prediction_result = None
    id2label_df = None
    if isinstance(ppl, TextClassificationPipeline):
        try:
            column_mapping = json.loads(column_mapping)
        except Exception:
            column_mapping = {}

        column_mapping, prediction_result, id2label_df = \
            text_classification_fix_column_mapping(column_mapping, ppl, d_id, config, split)

        column_mapping = json.dumps(column_mapping, indent=2)

    del ppl

    if prediction_result is None:
        gr.Warning('The model failed to predict with the first row in the dataset. Please provide column mappings in "Advance" settings.')
        return (
            config, split,
            gr.update(interactive=False),   # Submit button
            gr.update(visible=False),   # Model prediction preview
            gr.update(visible=False),   # Label mapping preview
            gr.update(value=column_mapping, visible=True, interactive=True),    # Column mapping
        )
    elif id2label_df is None:
        gr.Warning('The prediction result does not conform the labels in the dataset. Please provide label mappings in "Advance" settings.')
        return (
            config, split,
            gr.update(interactive=False),   # Submit button
            gr.update(value=prediction_result, visible=True),   # Model prediction preview
            gr.update(visible=False),   # Label mapping preview
            gr.update(value=column_mapping, visible=True, interactive=True),    # Column mapping
        )

    gr.Info("Model and dataset validations passed. Your can submit the evaluation task.")

    return (
        config, split,
        gr.update(interactive=True),    # Submit button
        gr.update(value=prediction_result, visible=True),   # Model prediction preview
        gr.update(value=id2label_df, visible=True), # Label mapping preview
        gr.update(value=column_mapping, visible=True, interactive=True),    # Column mapping
    )


def try_submit(m_id, d_id, config, split, column_mappings, local):
    label_mapping = {}
    try:
        column_mapping = json.loads(column_mappings)
        if "label" in column_mapping:
            label_mapping = column_mapping.pop("label", {})
    except Exception:
        column_mapping = {}

    if local:
        command = [
            "python",
            "cli.py",
            "--loader", "huggingface",
            "--model", m_id,
            "--dataset", d_id,
            "--dataset_config", config,
            "--dataset_split", split,
            "--hf_token", os.environ.get(HF_WRITE_TOKEN),
            "--discussion_repo", os.environ.get(HF_REPO_ID) or os.environ.get(HF_SPACE_ID),
            "--output_format", "markdown",
            "--output_portal", "huggingface",
            "--feature_mapping", json.dumps(column_mapping),
            "--label_mapping", json.dumps(label_mapping),
        ]

        eval_str = f"[{m_id}]<{d_id}({config}, {split} set)>"
        start = time.time()
        logging.info(f"Start local evaluation on {eval_str}")

        evaluator = subprocess.Popen(
            command,
            cwd=os.path.join(os.path.dirname(os.path.realpath(__file__)), "cicd"),
            stderr=subprocess.STDOUT,
        )
        result = evaluator.wait()

        logging.info(f"Finished local evaluation exit code {result} on {eval_str}: {time.time() - start:.2f}s")


with gr.Blocks(theme=theme) as iface:
    with gr.Row():
        with gr.Column():
            model_id_input = gr.Textbox(
                label="Hugging Face model id",
                placeholder="e.g.: cardiffnlp/twitter-roberta-base-sentiment-latest",
            )

            # TODO: Add supported model pairs: Text Classification - text-classification
            model_type = gr.Dropdown(
                label="Hugging Face model type",
                choices=[
                    ("Auto-detect", 0),
                    ("Text Classification", 1),
                ],
                value=0,
            )
            example_labels = gr.Label(label='Model prediction result', visible=False)

            id2label_mapping_dataframe = gr.DataFrame(
                label="Model-dataset label mapping",
                visible=False,
            )

        with gr.Column():
            dataset_id_input = gr.Textbox(
                label="Hugging Face dataset id",
                placeholder="e.g.: tweet_eval",
            )

            dataset_config_input = gr.Dropdown(
                label="Hugging Face dataset subset",
                choices=[
                    "default",
                ],
                allow_custom_value=True,
                value="default",
            )

            dataset_split_input = gr.Dropdown(
                label="Hugging Face dataset split",
                choices=[
                    "test",
                ],
                allow_custom_value=True,
                value="test",
            )

            with gr.Accordion("Advance", open=False):
                run_local = gr.Checkbox(value=True, label="Run in this Space")
                column_mapping_input = gr.Textbox(
                    value="",
                    lines=6,
                    label="Column mapping",
                    placeholder="Description of mapping of columns in model to dataset, in json format, e.g.:\n"
                                '{\n'
                                '   "text": "context",\n'
                                '   "label": {0: "Positive", 1: "Negative"}\n'
                                '}',
                )

    with gr.Row():
        validate_btn = gr.Button("Validate model and dataset", variant="primary")
        run_btn = gr.Button(
            "Submit evaluation task",
            variant="primary",
            interactive=False,
        )
        validate_btn.click(
            try_validate,
            inputs=[
                model_id_input,
                dataset_id_input,
                dataset_config_input,
                dataset_split_input,
                column_mapping_input,
            ],
            outputs=[
                dataset_config_input,
                dataset_split_input,
                run_btn,
                example_labels,
                id2label_mapping_dataframe,
                column_mapping_input,
            ],
        )
        run_btn.click(
            try_submit,
            inputs=[
                model_id_input,
                dataset_id_input,
                dataset_config_input,
                dataset_split_input,
                column_mapping_input,
                run_local,
            ],
        )

iface.queue(max_size=20)
iface.launch()