|
import torch |
|
import gradio as gr |
|
from transformers import pipeline |
|
import os |
|
from mtranslate import translate |
|
|
|
device = torch.cuda.current_device() if torch.cuda.is_available() else "cpu" |
|
HF_AUTH_TOKEN = os.environ.get("HF_AUTH_TOKEN") |
|
text_generation_model = "cahya/indochat-tiny" |
|
text_generation = pipeline("text-generation", text_generation_model, use_auth_token=HF_AUTH_TOKEN, device=device) |
|
|
|
|
|
def get_answer(user_input, decoding_methods, num_beams, top_k, top_p, temperature, repetition_penalty, penalty_alpha): |
|
if decoding_methods == "Beam Search": |
|
do_sample = False |
|
penalty_alpha = 0 |
|
elif decoding_methods == "Sampling": |
|
do_sample = True |
|
penalty_alpha = 0 |
|
num_beams = 1 |
|
else: |
|
do_sample = False |
|
num_beams = 1 |
|
print(user_input, decoding_methods, do_sample, top_k, top_p, temperature, repetition_penalty, penalty_alpha) |
|
prompt = f"User: {user_input}\nAssistant: " |
|
generated_text = text_generation(f"{prompt}", min_length=50, max_length=200, num_return_sequences=1, |
|
num_beams=num_beams, do_sample=do_sample, top_k=top_k, top_p=top_p, |
|
temperature=temperature, repetition_penalty=repetition_penalty, |
|
penalty_alpha=penalty_alpha) |
|
answer = generated_text[0]["generated_text"] |
|
answer_without_prompt = answer[len(prompt)+1:] |
|
user_input_en = translate(user_input, "en", "id") |
|
answer_without_prompt_en = translate(answer_without_prompt, "en", "id") |
|
return [(f"{user_input}\n", None), (answer_without_prompt, "")], \ |
|
[(f"{user_input_en}\n", None), (answer_without_prompt_en, "")] |
|
|
|
|
|
with gr.Blocks(css="#answer_id span {white-space: pre-line} #answer_en span {white-space: pre-line}") as demo: |
|
with gr.Row(): |
|
gr.Markdown("""## IndoChat |
|
|
|
A Prove of Concept of a multilingual Chatbot (in this case a bilingual, English and Indonesian), fine-tuned with |
|
multilingual instructions dataset. The base model is a GPT2-Medium (340M params) which was pretrained with 75GB |
|
of Indonesian and English dataset, where English part is only less than 1% of the whole dataset. |
|
""") |
|
with gr.Row(): |
|
with gr.Column(): |
|
user_input = gr.inputs.Textbox(placeholder="", |
|
label="Ask me something in Indonesian or English", |
|
default="Bagaimana cara mendidik anak supaya tidak berbohong?") |
|
decoding_methods = gr.inputs.Dropdown(["Beam Search", "Sampling", "Contrastive Search"], |
|
default="Sampling", label="Decoding Method") |
|
num_beams = gr.inputs.Slider(label="Number of beams for beam search", |
|
default=1, minimum=1, maximum=10, step=1) |
|
top_k = gr.inputs.Slider(label="Top K", |
|
default=30, maximum=50, minimum=1, step=1) |
|
top_p = gr.inputs.Slider(label="Top P", default=0.9, step=0.05, minimum=0.1, maximum=1.0) |
|
temperature = gr.inputs.Slider(label="Temperature", default=0.5, step=0.05, minimum=0.1, maximum=1.0) |
|
repetition_penalty = gr.inputs.Slider(label="Repetition Penalty", default=1.1, step=0.05, minimum=1.0, maximum=2.0) |
|
penalty_alpha = gr.inputs.Slider(label="The penalty alpha for contrastive search", |
|
default=0.5, step=0.05, minimum=0.05, maximum=1.0) |
|
with gr.Row(): |
|
button_generate_story = gr.Button("Submit") |
|
with gr.Column(): |
|
|
|
generated_answer = gr.HighlightedText( |
|
elem_id="answer_id", |
|
label="Generated Text", |
|
combine_adjacent=True, |
|
css="#htext span {white-space: pre-line}", |
|
).style(color_map={"": "blue", "-": "green"}) |
|
generated_answer_en = gr.HighlightedText( |
|
elem_id="answer_en", |
|
label="Translation", |
|
combine_adjacent=True, |
|
).style(color_map={"": "blue", "-": "green"}) |
|
with gr.Row(): |
|
gr.Markdown("![visitor badge](https://visitor-badge.glitch.me/badge?page_id=cahya_indochat)") |
|
|
|
button_generate_story.click(get_answer, |
|
inputs=[user_input, decoding_methods, num_beams, top_k, top_p, temperature, |
|
repetition_penalty, penalty_alpha], |
|
outputs=[generated_answer, generated_answer_en]) |
|
|
|
demo.launch(enable_queue=False) |