import torch
import numpy as np
from torch.nn.utils.rnn import pad_sequence
import gradio as gr
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
from sentence_transformers import SentenceTransformer

# Load the model and tokenizer
device = 'cuda' if torch.cuda.is_available() else 'cpu'
model_name = "google/flan-t5-large"
tokenizer = AutoTokenizer.from_pretrained(model_name)
embedding_model = SentenceTransformer('AnnaWegmann/Style-Embedding', device='cpu')
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
model.to(device)

def get_target_style_embeddings(target_texts_batch):
    all_target_texts = [target_text for target_texts in target_texts_batch for target_text in target_texts]
    embeddings = embedding_model.encode(all_target_texts, batch_size=len(all_target_texts), convert_to_tensor=True, show_progress_bar=False)
    lengths = [len(target_texts) for target_texts in target_texts_batch]
    split_embeddings = torch.split(embeddings, lengths)
    padded_embeddings = pad_sequence(split_embeddings, batch_first=True, padding_value=0.0)
    mask = (torch.arange(padded_embeddings.size(1))[None, :] < torch.tensor(lengths)[:, None]).to(torch.float32).unsqueeze(-1)
    mean_embeddings = torch.sum(padded_embeddings * mask, dim=1) / mask.sum(dim=1)
    return mean_embeddings.cpu().numpy()

def run_tinystyler_batch(source_texts, target_example_texts_batch, reranking, temperature, top_p):
    inputs = tokenizer(source_texts, return_tensors="pt")
    target_style_embeddings = get_target_style_embeddings(target_texts_batch)
    
    # Generate the output with specified temperature and top_p
    output = model.generate(
        inputs["input_ids"], 
        do_sample=True, 
        temperature=temperature, 
        top_p=top_p,
        max_length=1024
    )
    
    generated_texts = tokenizer.decode_batch(output, skip_special_tokens=True)
    return generated_texts
    
def run_tinystyler(source_text, target_texts, reranking, temperature, top_p):
    target_texts = [target_text.strip() for target_text in target_texts.split("\n")]
    return run_tinystyler_batch([source_text], [target_texts], reranking, temperature, top_p)[0]

# Preset examples with cached generations
preset_examples = {
    "Example 1": {
        "source_text": "Once upon a time in a small village",
        "target_texts": "In a land far away, there was a kingdom ruled by a wise king. Every day, the people of the kingdom would gather to listen to the king's stories, which were full of wisdom and kindness.",
        "reranking": 5,
        "temperature": 1.0,
        "top_p": 1.0,
        "output": "Once upon a time in a small village in a land far away, there was a kingdom ruled by a wise king. Every day, the people of the kingdom would gather to listen to the king's stories, which were full of wisdom and kindness."
    },
    "Example 2": {
        "source_text": "The quick brown fox",
        "target_texts": "A nimble, chocolate-colored fox swiftly darted through the emerald forest, weaving between trees with grace and agility.",
        "reranking": 5,
        "temperature": 0.9,
        "top_p": 0.9,
        "output": "The quick brown fox, a nimble, chocolate-colored fox, swiftly darted through the emerald forest, weaving between trees with grace and agility."
    }
}

# Define Gradio interface
with gr.Blocks(theme="ParityError/Interstellar@0.0.1") as demo:
    gr.Markdown("# TinyStyler Demo")
    gr.Markdown("Style transfer the source text into the target style, given some example texts of the target style. You can adjust re-ranking and top_p to your desire to control the quality of style transfer. A higher re-ranking value will generally result in better generations, at slower speed.")
    
    with gr.Row():
        example_dropdown = gr.Dropdown(label="Examples", choices=list(preset_examples.keys()))
    
    source_text = gr.Textbox(lines=3, placeholder="Enter the source text to transform into the target style...", label="Source Text")
    target_texts = gr.Textbox(lines=5, placeholder="Enter example texts of the target style (one per line)...", label="Example Texts of the Target Style")
    reranking = gr.Slider(1, 10, value=5, step=1, label="Re-ranking")
    temperature = gr.Slider(0.1, 2.0, value=1.0, step=0.1, label="Temperature")
    top_p = gr.Slider(0.0, 1.0, value=1.0, step=0.1, label="Top-P")
    
    output = gr.Textbox(lines=5, placeholder="Click 'Generate' to transform the source text into the target style.", label="Output", interactive=False)

    def set_example(example_name):
        example = preset_examples[example_name]
        return example["source_text"], example["target_example_texts"], example["reranking"], example["temperature"], example["top_p"], example["output"]

    example_dropdown.change(
        set_example,
        inputs=[example_dropdown],
        outputs=[source_text, target_example_texts, reranking, temperature, top_p, output]
    )
    
    btn = gr.Button("Generate")
    btn.click(run_tinystyler, [source_text, target_texts, reranking, temperature, top_p], output)

    # Initialize the fields with the first example
    example_dropdown.value, (source_text.value, target_texts.value, reranking.value, temperature.value, top_p.value, output.value) = list(preset_examples.keys())[0], set_example(list(preset_examples.keys())[0])

demo.launch()