app.py

import torch
import torch.nn as nn
from nltk.tokenize import word_tokenize
import nltk
from flask import Flask, request, jsonify

nltk.download('punkt')

# -------------------------
# MODEL ARKİTEKTÜRÜ
# -------------------------

class TransformerModel(nn.Module):
    def __init__(self, vocab_size, n_embd=512, n_head=16, n_layer=10, block_size=256):
        super().__init__()
        self.block_size = block_size

        self.token_embedding_table = nn.Embedding(vocab_size, n_embd)
        self.position_embedding_table = nn.Embedding(block_size, n_embd)

        self.layers = nn.ModuleList([
            nn.TransformerEncoderLayer(
                d_model=n_embd, 
                nhead=n_head, 
                dim_feedforward=n_embd * 4,
                dropout=0.1,
                activation="gelu",
                batch_first=True
            ) for _ in range(n_layer)
        ])

        self.ln_f = nn.LayerNorm(n_embd)
        self.head = nn.Linear(n_embd, vocab_size)

    def forward(self, idx):
        B, T = idx.shape

        tok_emb = self.token_embedding_table(idx)
        pos_emb = self.position_embedding_table(torch.arange(T, device=idx.device))
        x = tok_emb + pos_emb

        for layer in self.layers:
            x = layer(x)

        x = self.ln_f(x)
        logits = self.head(x)

        return logits

# -------------------------
# TOKENIZER
# -------------------------

class WordTokenizer:
    def __init__(self, vocab_path):
        import json
        with open(vocab_path, "r", encoding="utf-8") as f:
            obj = json.load(f)
        self.word_to_id = obj["word_to_id"]
        self.id_to_word = {int(v): k for k, v in self.word_to_id.items()}

    def encode(self, text):
        words = word_tokenize(text.lower())
        return [self.word_to_id.get(w, self.word_to_id["<unk>"]) for w in words]

    def decode(self, ids):
        return " ".join([self.id_to_word.get(i, "<unk>") for i in ids])

# -------------------------
# YÜKLEME
# -------------------------

tokenizer = WordTokenizer("vocab.json")
VOCAB_SIZE = len(tokenizer.word_to_id)

model = TransformerModel(VOCAB_SIZE)
model.load_state_dict(torch.load("pytorch_model.bin", map_location="cpu"))
model.eval()

app = Flask(__name__)

# -------------------------
# GENERATE FONKSİYONU
# -------------------------

def generate(text, max_new_tokens=50):
    ids = tokenizer.encode(text)
    x = torch.tensor([ids], dtype=torch.long)

    for _ in range(max_new_tokens):
        logits = model(x)
        last = logits[0, -1]
        probs = torch.softmax(last, dim=0)
        next_id = torch.multinomial(probs, num_samples=1)
        x = torch.cat([x, next_id.unsqueeze(0)], dim=1)

    out = x[0].tolist()
    return tokenizer.decode(out)

# -------------------------
# API
# -------------------------

@app.route("/chat", methods=["POST"])
def chat_api():
    data = request.get_json()
    user_text = data["text"]
    response = generate(user_text, max_new_tokens=40)
    return jsonify({"response": response})

if __name__ == "__main__":
    app.run()