app.py

"""
AutoTrain Mini — minimal FastAPI service for HuggingFace Spaces (CPU basic).

Endpoints:
  POST /train   { rows:[{text,label}], epochs, base_model }
  GET  /status
  POST /infer   { text }

CORS is wide open so the Lovable web app can call directly.
Training runs in a background thread (one job at a time).
"""
import os
import json
import threading
import traceback
from typing import List, Optional

from fastapi import FastAPI, HTTPException
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel

import torch
from torch.utils.data import Dataset
from transformers import (
    AutoTokenizer,
    AutoModelForSequenceClassification,
    Trainer,
    TrainingArguments,
    TrainerCallback,
)

MODEL_DIR = "/tmp/model"
os.makedirs(MODEL_DIR, exist_ok=True)

app = FastAPI()
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_methods=["*"],
    allow_headers=["*"],
)

STATE = {
    "state": "idle",       # idle | running | done | error
    "progress": 0,
    "epoch": 0,
    "loss": None,
    "accuracy": None,
    "message": "",
    "logs": [],
    "labels": [],
    "model_id": None,
}
LOCK = threading.Lock()
_pipeline = None  # cached inference pipeline


def log(msg: str):
    print(msg, flush=True)
    STATE["logs"].append(msg)
    if len(STATE["logs"]) > 200:
        STATE["logs"] = STATE["logs"][-200:]


class Row(BaseModel):
    text: str
    label: str


class TrainReq(BaseModel):
    rows: List[Row]
    epochs: int = 3
    base_model: str = "distilbert-base-uncased"


class InferReq(BaseModel):
    text: str


class TextDataset(Dataset):
    def __init__(self, encodings, labels):
        self.encodings = encodings
        self.labels = labels

    def __len__(self):
        return len(self.labels)

    def __getitem__(self, i):
        item = {k: torch.tensor(v[i]) for k, v in self.encodings.items()}
        item["labels"] = torch.tensor(self.labels[i])
        return item


class ProgressCB(TrainerCallback):
    def on_log(self, args, state, control, logs=None, **kw):
        if not logs:
            return
        if "loss" in logs:
            STATE["loss"] = float(logs["loss"])
        if "epoch" in logs:
            STATE["epoch"] = float(logs["epoch"])
        if state.max_steps:
            STATE["progress"] = int(100 * state.global_step / state.max_steps)
        log(f"step={state.global_step} {logs}")


def do_train(req: TrainReq):
    global _pipeline
    try:
        STATE.update(state="running", progress=0, message="Preparing data", logs=[], loss=None, accuracy=None)
        log(f"Rows: {len(req.rows)}  base_model={req.base_model}  epochs={req.epochs}")

        texts = [r.text for r in req.rows]
        raw_labels = [r.label for r in req.rows]
        label_list = sorted(set(raw_labels))
        label2id = {l: i for i, l in enumerate(label_list)}
        id2label = {i: l for l, i in label2id.items()}
        STATE["labels"] = label_list
        log(f"Labels ({len(label_list)}): {label_list}")

        y = [label2id[l] for l in raw_labels]

        STATE["message"] = "Loading tokenizer/model"
        tok = AutoTokenizer.from_pretrained(req.base_model)
        model = AutoModelForSequenceClassification.from_pretrained(
            req.base_model,
            num_labels=len(label_list),
            id2label=id2label,
            label2id=label2id,
        )

        STATE["message"] = "Tokenizing"
        enc = tok(texts, truncation=True, padding=True, max_length=128)
        ds = TextDataset(enc, y)

        args = TrainingArguments(
            output_dir="/tmp/out",
            num_train_epochs=req.epochs,
            per_device_train_batch_size=8,
            learning_rate=5e-5,
            logging_steps=5,
            save_strategy="no",
            report_to=[],
            disable_tqdm=True,
        )
        trainer = Trainer(model=model, args=args, train_dataset=ds, callbacks=[ProgressCB()])

        STATE["message"] = "Training"
        trainer.train()

        STATE["message"] = "Evaluating (train acc)"
        preds = trainer.predict(ds)
        pred_ids = preds.predictions.argmax(-1)
        acc = float((pred_ids == preds.label_ids).mean())
        STATE["accuracy"] = acc
        log(f"Train accuracy: {acc:.4f}")

        STATE["message"] = "Saving"
        model.save_pretrained(MODEL_DIR)
        tok.save_pretrained(MODEL_DIR)
        with open(os.path.join(MODEL_DIR, "labels.json"), "w") as f:
            json.dump(label_list, f)

        _pipeline = None  # force reload
        STATE.update(state="done", progress=100, message="Done", model_id="local")
        log("Training complete.")
    except Exception as e:
        STATE.update(state="error", message=str(e))
        log("ERROR: " + traceback.format_exc())


@app.get("/")
def root():
    return {"ok": True, "service": "autotrain-mini", "state": STATE["state"]}


@app.get("/status")
def status():
    return STATE


@app.post("/train")
def train(req: TrainReq):
    with LOCK:
        if STATE["state"] == "running":
            raise HTTPException(409, "A training job is already running")
        t = threading.Thread(target=do_train, args=(req,), daemon=True)
        t.start()
    return {"ok": True, "message": "started"}


def get_pipeline():
    global _pipeline
    if _pipeline is not None:
        return _pipeline
    if not os.path.exists(os.path.join(MODEL_DIR, "config.json")):
        return None
    from transformers import pipeline
    _pipeline = pipeline("text-classification", model=MODEL_DIR, tokenizer=MODEL_DIR, top_k=None)
    return _pipeline


@app.post("/infer")
def infer(req: InferReq):
    pipe = get_pipeline()
    if pipe is None:
        raise HTTPException(400, "No trained model yet. Call /train first.")
    out = pipe(req.text)
    return {"input": req.text, "predictions": out}