Список проголосовавших за комментарий "<p>Михаил, дак мне и в предыдущем посте кидали такие задачи — мол вот тут ошибка, реши. </p><p>Вот как Grok предлагает исправить ваш код. </p><p>import asyncio<br/> import collections<br/> import itertools<br/> import logging<br/> import time<br/> from typing import Literal, cast</p><p>import torch<br/> import tqdm<br/> from pydantic import BaseModel<br/> from torch import optim</p><p>from poptimizer import consts, errors<br/> from poptimizer.domain.dl import data_loaders, datasets, ledoit_wolf, risk<br/> from poptimizer.domain.dl.wave_net import backbone, wave_net<br/> from poptimizer.domain.evolve import evolve<br/> from poptimizer.use_cases import handler<br/> from poptimizer.use_cases.dl import builder</p><p><br/> class Optimizer(BaseModel):<br/> lr: float<br/> beta1: float<br/> beta2: float<br/> eps: float<br/> weight_decay: float<br/> momentum_decay: float<br/> decoupled_weight_decay: bool</p><p><br/> class Scheduler(BaseModel):<br/> max_lr: float<br/> epochs: float<br/> pct_start: float<br/> anneal_strategy: Literal[«linear», «cos»]<br/> cycle_momentum: bool<br/> base_momentum: float<br/> max_momentum: float<br/> div_factor: float<br/> final_div_factor: float<br/> three_phase: bool</p><p><br/> class Cfg(BaseModel):<br/> batch: builder.Batch<br/> net: backbone.Cfg<br/> optimizer: Optimizer<br/> scheduler: Scheduler<br/> risk: risk.Cfg</p><p><br/> class RunningMean:<br/> def __init__(self, window_size: int) -> None:<br/> self._sum: float = 0<br/> self._que: collections.deque[float] = collections.deque([0], maxlen=window_size)</p><p>def append(self, num: float) -> None:<br/> self._sum += num — self._que[0]<br/> self._que.append(num)</p><p>def running_avg(self) -> float:<br/> return self._sum / len(self._que)</p><p><br/> def _get_device() -> Literal[«cpu», «cuda», «mps»]:<br/> if torch.cuda.is_available():<br/> return «cuda»<br/> if torch.backends.mps.is_available():<br/> return «mps»<br/> return «cpu»</p><p><br/> class Trainer:<br/> def __init__(self, builder: builder.Builder) -> None:<br/> self._lgr = logging.getLogger()<br/> self._builder = builder<br/> self._device = _get_device()<br/> self._stopping = False</p><p>async def update_model_metrics(<br/> self,<br/> ctx: handler.Ctx,<br/> model: evolve.Model,<br/> test_days: int,<br/> ) -> None:<br/> start = time.monotonic()</p><p>cfg = Cfg.model_validate(model.phenotype)<br/> days = datasets.Days(<br/> history=cfg.batch.history_days,<br/> forecast=model.forecast_days,<br/> test=test_days,<br/> )<br/> data, emb_size, emb_seq_size = await self._builder.build(<br/> ctx,<br/> model.day,<br/> model.tickers,<br/> days,<br/> cfg.batch,<br/> )</p><p>try:<br/> await asyncio.to_thread(<br/> self._run,<br/> model,<br/> data,<br/> emb_size,<br/> emb_seq_size,<br/> cfg,<br/> model.forecast_days,<br/> )<br/> except asyncio.CancelledError:<br/> self._stopping = True<br/> if self._device == «cuda»:<br/> torch.cuda.empty_cache()<br/> raise<br/> finally:<br/> # Освобождаем данные<br/> del data<br/> if self._device == «cuda»:<br/> torch.cuda.empty_cache()</p><p>model.risk_tolerance = cfg.risk.risk_tolerance<br/> model.duration = time.monotonic() — start</p><p>def _run(<br/> self,<br/> model: evolve.Model,<br/> data: list[datasets.TickerData],<br/> emb_size: list[int],<br/> emb_seq_size: list[int],<br/> cfg: Cfg,<br/> forecast_days: int,<br/> ) -> None:<br/> net = self._prepare_net(cfg, emb_size, emb_seq_size)<br/> try:<br/> self._train(net, cfg.optimizer, cfg.scheduler, data, cfg.batch.size)<br/> model.alfa, model.llh = self._test(net, cfg, forecast_days, data)<br/> model.mean, model.cov = self._forecast(net, forecast_days, data)<br/> finally:<br/> del net<br/> if self._device == «cuda»:<br/> torch.cuda.empty_cache()</p><p>def _train(<br/> self,<br/> net: wave_net.Net,<br/> optimizer: Optimizer,<br/> scheduler: Scheduler,<br/> data: list[datasets.TickerData],<br/> batch_size: int,<br/> ) -> None:<br/> train_dl = data_loaders.train(data, batch_size)<br/> opt = optim.NAdam(<br/> net.parameters(),<br/> lr=optimizer.lr,<br/> betas=(optimizer.beta1, optimizer.beta2),<br/> eps=optimizer.eps,<br/> weight_decay=optimizer.weight_decay,<br/> momentum_decay=optimizer.momentum_decay,<br/> decoupled_weight_decay=optimizer.decoupled_weight_decay,<br/> )</p><p>steps_per_epoch = len(train_dl)<br/> total_steps = int(steps_per_epoch * scheduler.epochs)</p><p>sch = optim.lr_scheduler.OneCycleLR(<br/> opt,<br/> max_lr=scheduler.max_lr,<br/> total_steps=total_steps,<br/> pct_start=scheduler.pct_start,<br/> anneal_strategy=scheduler.anneal_strategy,<br/> cycle_momentum=scheduler.cycle_momentum,<br/> base_momentum=scheduler.base_momentum,<br/> max_momentum=scheduler.max_momentum,<br/> div_factor=scheduler.div_factor,<br/> final_div_factor=scheduler.final_div_factor,<br/> three_phase=scheduler.three_phase,<br/> )</p><p>self._log_net_stats(net, scheduler.epochs, len(train_dl.dataset))</p><p>avg_llh = RunningMean(steps_per_epoch)<br/> net.train()</p><p>try:<br/> with tqdm.tqdm(total=total_steps, desc=«Train») as progress_bar:<br/> for epoch in range(int(scheduler.epochs)):<br/> for batch in train_dl:<br/> if self._stopping:<br/> return</p><p>opt.zero_grad()</p><p>loss = -net.llh(<br/> batch.num_feat.to(self._device),<br/> batch.emb_feat.to(self._device),<br/> batch.emb_seq_feat.to(self._device),<br/> batch.labels.to(self._device),<br/> )<br/> loss.backward()<br/> opt.step()<br/> sch.step()</p><p>avg_llh.append(-loss.item())<br/> progress_bar.set_postfix_str(f"{avg_llh.running_avg():.5f}")<br/> progress_bar.update(1)</p><p>loss = None # Освобождаем тензор<br/> if self._device == «cuda»:<br/> torch.cuda.empty_cache()<br/> # self._lgr.debug(<br/> # «Memory allocated: %.2f MB»,<br/> # torch.cuda.memory_allocated() / 1024**2,<br/> # )<br/> finally:<br/> del train_dl<br/> del opt<br/> del sch<br/> del avg_llh<br/> if self._device == «cuda»:<br/> torch.cuda.empty_cache()</p><p>def _test(<br/> self,<br/> net: wave_net.Net,<br/> cfg: Cfg,<br/> forecast_days: int,<br/> data: list[datasets.TickerData],<br/> ) -> tuple[list[float], list[float]]:<br/> with torch.inference_mode():<br/> net.eval()<br/> test_dl = data_loaders.test(data)<br/> alfa_sum = 0.0<br/> llh_sum = 0.0<br/> count = 0</p><p>try:<br/> for batch in test_dl:<br/> if self._stopping:<br/> break</p><p>loss, mean, std = net.loss_and_forecast_mean_and_std(<br/> batch.num_feat.to(self._device),<br/> batch.emb_feat.to(self._device),<br/> batch.emb_seq_feat.to(self._device),<br/> batch.labels.to(self._device),<br/> )<br/> rez = risk.optimize(<br/> mean,<br/> std,<br/> batch.labels.numpy() — 1,<br/> batch.returns.numpy(),<br/> cfg.risk,<br/> forecast_days,<br/> )</p><p>self._lgr.info(«Return: %.4f, LLH: %.4f», rez.ret — rez.avr, loss)</p><p>alfa_sum += rez.ret — rez.avr<br/> llh_sum += loss<br/> count += 1</p><p>if self._device == «cuda»:<br/> torch.cuda.empty_cache()<br/> finally:<br/> del test_dl<br/> if self._device == «cuda»:<br/> torch.cuda.empty_cache()</p><p>if count == 0:<br/> raise errors.UseCasesError(«No test batches processed»)</p><p>return [alfa_sum / count], [llh_sum / count]</p><p>def _forecast(<br/> self,<br/> net: wave_net.Net,<br/> forecast_days: int,<br/> data: list[datasets.TickerData],<br/> ) -> tuple[list[list[float]], list[list[float]]]:<br/> with torch.inference_mode():<br/> net.eval()<br/> forecast_dl = data_loaders.forecast(data)<br/> try:<br/> if len(forecast_dl) != 1:<br/> raise errors.UseCasesError(«Invalid forecast dataloader»)</p><p>batch = next(iter(forecast_dl))<br/> mean, std = net.forecast_mean_and_std(<br/> batch.num_feat.to(self._device),<br/> batch.emb_feat.to(self._device),<br/> batch.emb_seq_feat.to(self._device),<br/> )</p><p>year_multiplier = consts.YEAR_IN_TRADING_DAYS / forecast_days<br/> mean *= year_multiplier<br/> std *= year_multiplier**0.5</p><p>total_ret = batch.returns.numpy()<br/> cov = std.T * ledoit_wolf.ledoit_wolf_cor(total_ret)[0] * std</p><p>return cast(«list[list[float]]», mean.tolist()), cov.tolist()<br/> finally:<br/> del forecast_dl<br/> del batch<br/> if self._device == «cuda»:<br/> torch.cuda.empty_cache()</p><p>def _log_net_stats(self, net: wave_net.Net, epochs: float, steps_per_epoch: int) -> None:<br/> self._lgr.info(«Epochs: %.2f, Train size: %s», epochs, steps_per_epoch)<br/> modules = sum(1 for _ in net.modules())<br/> model_params = sum(tensor.numel() for tensor in net.parameters())<br/> self._lgr.info(«Layers: %d, Parameters: %d», modules, model_params)</p><p>def _prepare_net(self, cfg: Cfg, emb_size: list[int], emb_seq_size: list[int]) -> wave_net.Net:<br/> net = wave_net.Net(<br/> cfg=cfg.net,<br/> history_days=cfg.batch.history_days,<br/> num_feat_count=cfg.batch.num_feat_count,<br/> emb_size=emb_size,<br/> emb_seq_size=emb_seq_size,<br/> ).to(self._device)<br/> return net</p><p> </p><h3>Если проблемы сохраняются</h3><ul><li>Поделитесь логами профилирования памяти (tracemalloc или torch.cuda.memory_allocated).</li><li>Укажите, возникает ли утечка при конкретном методе (_train, _test, _forecast).</li><li>Предоставьте реализацию data_loaders или wave_net.Net, если они могут быть источником проблемы.</li></ul><p>Код выше должен значительно снизить риск утечек памяти.</p>"