main.py

import torch
import torch.nn as nn
import re
from fastapi import FastAPI
from pydantic import BaseModel
from fastapi.middleware.cors import CORSMiddleware

# =====================================================
# 1. FastAPI App
# =====================================================
app = FastAPI(
    title="Khmer Spell Correction API",
    version="1.0"
)

# Allow CORS for testing
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_methods=["*"],
    allow_headers=["*"],
)

# =====================================================
# 2. Utils
# =====================================================
def preprocess_khmer_text(text: str) -> str:
    """Clean and normalize Khmer text."""
    text = re.sub(r'\s+', ' ', text)
    text = re.sub(r'[^\u1780-\u17FF\u200B\u0020-\u007E]', '', text)
    return text.strip()

# =====================================================
# 3. Model Definition
# =====================================================
class KhmerSpellLSTM(nn.Module):
    def __init__(self, vocab_size, embedding_dim, hidden_dim, num_layers=2, dropout=0.3):
        super().__init__()
        self.embedding = nn.Embedding(vocab_size, embedding_dim, padding_idx=0)
        self.lstm = nn.LSTM(
            embedding_dim,
            hidden_dim,
            num_layers=num_layers,
            batch_first=True,
            dropout=dropout if num_layers > 1 else 0,
            bidirectional=True
        )
        # Match checkpoint fc
        self.fc = nn.Sequential(
            nn.Linear(hidden_dim * 2, hidden_dim),
            nn.ReLU(),
            nn.Dropout(dropout),
            nn.Linear(hidden_dim, vocab_size)
        )

    def forward(self, x):
        emb = self.embedding(x)
        out, _ = self.lstm(emb)
        return self.fc(out)

# =====================================================
# 4. Load Model ONCE
# =====================================================
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

checkpoint = torch.load("result/khmer_spell_lstm.pth", map_location=device)

char_to_idx = checkpoint["char_to_idx"]
vocab = checkpoint.get("vocab", char_to_idx.keys())
max_length = checkpoint["max_length"]
idx_to_char = {i: c for c, i in char_to_idx.items()}

model = KhmerSpellLSTM(
    vocab_size=len(vocab),
    embedding_dim=128,
    hidden_dim=256
).to(device)

# Load weights
model.load_state_dict(checkpoint["model_state_dict"])
model.eval()
print("✅ Khmer Spell LSTM loaded successfully")

# =====================================================
# 5. Inference Function
# =====================================================
def predict(text: str) -> str:
    text = preprocess_khmer_text(text)
    input_len = len(text)

    seq = [char_to_idx.get(c, char_to_idx["<UNK>"]) for c in text]
    seq += [char_to_idx["<PAD>"]] * (max_length - len(seq))
    seq = torch.tensor(seq[:max_length]).unsqueeze(0).to(device)

    with torch.no_grad():
        out = model(seq)
        pred = torch.argmax(out, dim=-1)[0]

    # Keep the prediction same length as input
    pred = pred[:input_len]

    return "".join(idx_to_char[i.item()] for i in pred)

# =====================================================
# 6. API Schema
# =====================================================
class TextInput(BaseModel):
    text: str

# =====================================================
# 7. Routes
# =====================================================
@app.get("/")
def health_check():
    return {"status": "Khmer Spell API running"}

@app.post("/predict")
def spell_correct(data: TextInput):
    corrected_text = predict(data.text)
    return {
        "input": data.text,
        "output": corrected_text
    }