english-only-app.py

import os
import re
from datetime import datetime

import gradio
import sign_language_translator as slt

DESCRIPTION = """Enter your English text and click Submit to generate a sign language video. [`Library Repository`](https://github.com/sign-language-translator/sign-language-translator)
The text is preprocessed, tokenized and rearranged and then each token is mapped to a prerecorded video which are concatenated and returned. [`Model Code`](https://github.com/sign-language-translator/sign-language-translator/blob/main/sign_language_translator/models/text_to_sign/concatenative_synthesis.py)
> **NOTE**
> - This model only supports a fixed vocabulary. See the [`*-dictionary-mapping.json`](https://github.com/sign-language-translator/sign-language-datasets/tree/main/parallel_texts) files for supported words.
> - This version needs to re-encode the generated video so that will take some extra time after translation.
> - Since this is a rule-based model, you will have to add **context** to ambiguous words (e.g. glass(material) vs glass(container)).
""".strip()

TITLE = "English to Sign Language Translator"

CUSTOM_CSS = """
#auto-complete-button {
    border-color: var(--button-primary-border-color-hover);
}
"""

HF_TOKEN = os.getenv("HF_TOKEN")
request_logger = (
    gradio.HuggingFaceDatasetSaver(
        HF_TOKEN,
        "sltAI/crowdsourced-text-to-sign-language-rule-based-translation-corpus",
    )
    if HF_TOKEN
    else gradio.CSVLogger()
)

translation_model = slt.models.ConcatenativeSynthesis("en", "pk-sl", "video")
language_models = {}

def auto_complete_text(model_code: str, text: str):
    if model_code not in language_models:
        lm = slt.get_model(model_code)
        language_models[model_code] = slt.models.BeamSampling(
            lm,  # type: ignore
            start_of_sequence_token=getattr(lm, "start_of_sequence_token", "<"),  # type: ignore
            end_of_sequence_token=getattr(lm, "end_of_sequence_token", ">"),  # type: ignore
        )

    tokens = [w for w in re.split(r"\b", text) if w]
    lm = language_models[model_code]
    lm.max_length = len(tokens) + 10
    completion, _ = lm.complete(tokens or None)
    if completion[0] == lm.start_of_sequence_token:  # type: ignore
        completion = completion[1:]  # type: ignore
    if completion[-1] == lm.end_of_sequence_token:  # type: ignore
        completion = completion[:-1]  # type: ignore
    new_text = "".join(completion)

    return new_text


def text_to_video(
    text: str,
    sign_language: str = "pakistan-sign-language",
    sign_format: str = "video",
    output_path: str = "output.mp4",
    codec="h264",  # ToDo: install h264 codec for opencv
):
    translation_model.text_language = "en"  # Hardcoded to English
    translation_model.sign_language = sign_language
    translation_model.sign_format = sign_format
    if sign_format == "landmarks":
        translation_model.sign_embedding_model = "mediapipe-world"

    # Convert first letter to lowercase as per original code
    text = text[:1].lower() + text[1:]
    
    sign = translation_model.translate(text)
    if isinstance(sign, slt.Landmarks):
        # hands moved to pose wrists
        sign.data[:, 33:54, :3] += -sign.data[:, 33:34, :3] + sign.data[:, 15:16, :3]
        sign.data[:, 54:  , :3] += -sign.data[:, 54:55, :3] + sign.data[:, 16:17, :3]

        sign.save_animation(output_path, overwrite=True)
    else:
        sign.save(output_path, overwrite=True, codec=codec)


def translate(text: str, sign_lang: str, sign_format: str):
    log = [
        text,
        "en",
        sign_lang,
        None,
        datetime.now().strftime("%Y-%m-%d %H:%M:%S.%f"),
    ]
    try:
        path = "output.mp4"
        text_to_video(
            text,
            sign_language=sign_lang,
            sign_format=sign_format,
            output_path=path,
            codec="mp4v",
        )
        request_logger.flag(log)
        return path

    except Exception as exc:
        log[3] = str(exc)
        request_logger.flag(log)
        raise gradio.Error(f"Error during translation: {exc}")


with gradio.Blocks(title=TITLE, css=CUSTOM_CSS) as gradio_app:
    gradio.Markdown(f"# {TITLE}")
    gradio.Markdown(DESCRIPTION)
    with gradio.Row():
        with gradio.Column():  # Inputs
            gradio.Markdown("## Select Output Options")
            with gradio.Row():
                sign_lang_dropdown = gradio.Dropdown(
                    choices=[code.value for code in slt.SignLanguageCodes],
                    value=slt.SignLanguageCodes.PAKISTAN_SIGN_LANGUAGE.value,
                    label="Sign Language",
                )
                output_format_dropdown = gradio.Dropdown(
                    choices=[
                        slt.SignFormatCodes.VIDEO.value,
                        slt.SignFormatCodes.LANDMARKS.value,
                    ],
                    value=slt.SignFormatCodes.VIDEO.value,
                    label="Output Format",
                )

            gradio.Markdown("## Input English Text")
            with gradio.Row():  # Source TextArea
                source_textbox = gradio.Textbox(
                    lines=4,
                    placeholder="Enter English Text Here...",
                    label="English Sentence",
                    show_copy_button=True,
                    elem_id="source-textbox",
                )
            with gradio.Row():  # clear/auto-complete/Language Model
                # We'll keep the language model for English text completion
                language_model_dropdown = gradio.Dropdown(
                    choices=[
                        slt.ModelCodes.TRANSFORMER_LM_EN_SUPPORTED.value,
                    ],
                    value=slt.ModelCodes.TRANSFORMER_LM_EN_SUPPORTED.value,
                    label="Select language model to Generate sample text",
                )

                auto_complete_button = gradio.Button(
                    "Auto-Complete", elem_id="auto-complete-button"
                )
                auto_complete_button.click(
                    auto_complete_text,
                    inputs=[language_model_dropdown, source_textbox],
                    outputs=[source_textbox],
                    api_name=False,
                )
                clear_button = gradio.ClearButton(source_textbox, api_name=False)

            with gradio.Row():  # Translate Button
                translate_button = gradio.Button("Translate", variant="primary")
                translate_button.click(
                    translate,
                    inputs=[
                        source_textbox,
                        sign_lang_dropdown,
                        output_format_dropdown,
                    ],
                    outputs=[output_video],
                    api_name="translate",
                )
                
        with gradio.Column():  # Outputs
            gradio.Markdown("## Output Sign Language")
            output_video = gradio.Video(
                format="mp4",
                label="Synthesized Sign Language Video",
                autoplay=True,
                show_download_button=True,
                include_audio=False,
            )

    gradio.Examples(
        [
            ["We are here to use this.", "pakistan-sign-language", "video"],
            ["I admire art.", "pakistan-sign-language", "landmarks"],
            ["This is very good.", "pakistan-sign-language", "video"],
            ["That work was easy.", "pakistan-sign-language", "landmarks"],
            ["How are you?", "pakistan-sign-language", "video"],
            ["Five hours.", "pakistan-sign-language", "landmarks"],
        ],
        inputs=[
            source_textbox,
            sign_lang_dropdown,
            output_format_dropdown,
        ],
        outputs=output_video,
    )
    request_logger.setup(
        [
            source_textbox,
            gradio.Markdown(label="Language", value="en"),
            sign_lang_dropdown,
            gradio.Markdown(label="Exception"),
            gradio.Markdown(label="Timestamp"),
        ],
        "flagged",
    )

if __name__ == "__main__":
    gradio_app.launch()