import gradio as gr
import datasets
import huggingface_hub
import sys
import os
from pathlib import Path


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_submit(model_id, dataset_id, dataset_config, dataset_split, local):
    # 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
    if isinstance(ppl, Exception):
        gr.Warning(f'Failed to load "{model_id} model": {ppl}')
        return dataset_config, dataset_split

    # 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

    # TODO: Validate column mapping by running once

    del ppl

    if local:
        if "cicd" not in sys.path:
            sys.path.append("cicd")
        from giskard_cicd.loaders import HuggingFaceLoader
        from giskard_cicd.pipeline.runner import PipelineRunner

        from automation import create_discussion_detailed
        supported_loaders = {
            "huggingface": HuggingFaceLoader(),
        }

        runner = PipelineRunner(loaders=supported_loaders)

        runner_kwargs = {
            "loader_id": "huggingface",
            "model": m_id,
            "dataset": d_id,
            "scan_config": None,
            "dataset_split": split,
            "dataset_config": config,
        }

        report = runner.run(**runner_kwargs)

        # TODO: Publish it with given repo id/model id
        if os.environ.get(HF_REPO_ID) or os.environ.get(HF_SPACE_ID) and os.environ.get(HF_WRITE_TOKEN):
            rendered_report = report.to_markdown(template="github")
            repo = os.environ.get(HF_REPO_ID) or os.environ.get(HF_SPACE_ID)
            create_discussion_detailed(repo, m_id, d_id, config, split, os.environ.get(HF_WRITE_TOKEN), rendered_report)

        # Cache locally
        rendered_report = report.to_html()
        output_dir = Path(f"output/{m_id}/{d_id}/{config}/{split}/")
        output_dir.mkdir(parents=True, exist_ok=True)
        with open(output_dir / "report.html", "w") as f:
            print(f'Writing to {output_dir / "report.html"}')
            f.write(rendered_report)

    return config, split

with gr.Blocks(theme=theme) as iface:
    with gr.Row():
        with gr.Column():
            model_id_input = gr.Textbox(
                label="Hugging Face model id",
                placeholder="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,
            )
            run_local = gr.Checkbox(value=True, label="Run in this Space")

        with gr.Column():
            dataset_id_input = gr.Textbox(
                label="Hugging Face dataset id",
                placeholder="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.Row():
        run_btn = gr.Button("Validate and submit evaluation task", variant="primary")
        run_btn.click(
            try_submit,
            inputs=[
                model_id_input,
                dataset_id_input,
                dataset_config_input,
                dataset_split_input,
                run_local,
            ],
            outputs=[
                dataset_config_input,
                dataset_split_input
            ],
        )

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