diff options
Diffstat (limited to 'text_recognizer')
| -rw-r--r-- | text_recognizer/callbacks/wandb_callbacks.py | 9 | ||||
| -rw-r--r-- | text_recognizer/models/base.py | 83 | ||||
| -rw-r--r-- | text_recognizer/models/transformer.py | 26 | ||||
| -rw-r--r-- | text_recognizer/networks/encoders/efficientnet/efficientnet.py | 10 |
4 files changed, 61 insertions, 67 deletions
diff --git a/text_recognizer/callbacks/wandb_callbacks.py b/text_recognizer/callbacks/wandb_callbacks.py index 3936aaf..900c3b1 100644 --- a/text_recognizer/callbacks/wandb_callbacks.py +++ b/text_recognizer/callbacks/wandb_callbacks.py @@ -29,6 +29,9 @@ class WatchModel(Callback): log: str = attr.ib(default="gradients") log_freq: int = attr.ib(default=100) + def __attrs_pre_init__(self): + super().__init__() + def on_train_start(self, trainer: Trainer, pl_module: LightningModule) -> None: """Watches model weights with wandb.""" logger = get_wandb_logger(trainer) @@ -41,6 +44,9 @@ class UploadCodeAsArtifact(Callback): project_dir: Path = attr.ib(converter=Path) + def __attrs_pre_init__(self): + super().__init__() + def on_train_start(self, trainer: Trainer, pl_module: LightningModule) -> None: """Uploads project code as an artifact.""" logger = get_wandb_logger(trainer) @@ -59,6 +65,9 @@ class UploadCheckpointAsArtifact(Callback): ckpt_dir: Path = attr.ib(converter=Path) upload_best_only: bool = attr.ib() + def __attrs_pre_init__(self): + super().__init__() + def on_train_end(self, trainer: Trainer, pl_module: LightningModule) -> None: """Uploads model checkpoint to W&B.""" logger = get_wandb_logger(trainer) diff --git a/text_recognizer/models/base.py b/text_recognizer/models/base.py index 88ffde6..4e803eb 100644 --- a/text_recognizer/models/base.py +++ b/text_recognizer/models/base.py @@ -1,8 +1,10 @@ """Base PyTorch Lightning model.""" from typing import Any, Dict, List, Union, Tuple, Type -import madgrad -from omegaconf import DictConfig, OmegaConf +import attr +import hydra +import loguru.logger as log +from omegaconf import DictConfig import pytorch_lightning as pl import torch from torch import nn @@ -10,23 +12,29 @@ from torch import Tensor import torchmetrics +@attr.s class LitBaseModel(pl.LightningModule): """Abstract PyTorch Lightning class.""" - def __init__( - self, - network: Type[nn.Module], - optimizer: Union[DictConfig, Dict], - lr_scheduler: Union[DictConfig, Dict], - criterion: Union[DictConfig, Dict], - monitor: str = "val_loss", - ) -> None: + network: Type[nn.Module] = attr.ib() + criterion_config: DictConfig = attr.ib(converter=DictConfig) + optimizer_config: DictConfig = attr.ib(converter=DictConfig) + lr_scheduler_config: DictConfig = attr.ib(converter=DictConfig) + + interval: str = attr.ib() + monitor: str = attr.ib(default="val/loss") + + loss_fn = attr.ib(init=False) + + train_acc = attr.ib(init=False) + val_acc = attr.ib(init=False) + test_acc = attr.ib(init=False) + + def __attrs_pre_init__(self): super().__init__() - self.monitor = monitor - self.network = network - self._optimizer = OmegaConf.create(optimizer) - self._lr_scheduler = OmegaConf.create(lr_scheduler) - self.loss_fn = self.configure_criterion(criterion) + + def __attrs_post_init__(self): + self.loss_fn = self.configure_criterion() # Accuracy metric self.train_acc = torchmetrics.Accuracy() @@ -34,11 +42,10 @@ class LitBaseModel(pl.LightningModule): self.test_acc = torchmetrics.Accuracy() @staticmethod - def configure_criterion(criterion: Union[DictConfig, Dict]) -> Type[nn.Module]: + def configure_criterion(self) -> Type[nn.Module]: """Returns a loss functions.""" - criterion = OmegaConf.create(criterion) - args = {} or criterion.args - return getattr(nn, criterion.type)(**args) + log.info(f"Instantiating criterion <{self.criterion_config._target_}>") + return hydra.utils.instantiate(self.criterion_config) def optimizer_zero_grad( self, @@ -51,27 +58,23 @@ class LitBaseModel(pl.LightningModule): def _configure_optimizer(self) -> Type[torch.optim.Optimizer]: """Configures the optimizer.""" - args = {} or self._optimizer.args - if self._optimizer.type == "MADGRAD": - optimizer_class = madgrad.MADGRAD - else: - optimizer_class = getattr(torch.optim, self._optimizer.type) - return optimizer_class(params=self.parameters(), **args) + log.info(f"Instantiating optimizer <{self.optimizer_config._target_}>") + return hydra.utils.instantiate(self.optimizer_config, params=self.parameters()) def _configure_lr_scheduler( self, optimizer: Type[torch.optim.Optimizer] ) -> Dict[str, Any]: """Configures the lr scheduler.""" - scheduler = {"monitor": self.monitor} - args = {} or self._lr_scheduler.args - - if "interval" in args: - scheduler["interval"] = args.pop("interval") - - scheduler["scheduler"] = getattr( - torch.optim.lr_scheduler, self._lr_scheduler.type - )(optimizer, **args) - + log.info( + f"Instantiating learning rate scheduler <{self.lr_scheduler_config._target_}>" + ) + scheduler = { + "monitor": self.monitor, + "interval": self.interval, + "scheduler": hydra.utils.instantiate( + self.lr_scheduler_config, optimizer=optimizer + ), + } return scheduler def configure_optimizers(self) -> Tuple[List[type], List[Dict[str, Any]]]: @@ -90,9 +93,9 @@ class LitBaseModel(pl.LightningModule): data, targets = batch logits = self(data) loss = self.loss_fn(logits, targets) - self.log("train_loss", loss) + self.log("train/loss", loss) self.train_acc(logits, targets) - self.log("train_acc", self.train_acc, on_step=False, on_epoch=True) + self.log("train/acc", self.train_acc, on_step=False, on_epoch=True) return loss def validation_step(self, batch: Tuple[Tensor, Tensor], batch_idx: int) -> None: @@ -100,13 +103,13 @@ class LitBaseModel(pl.LightningModule): data, targets = batch logits = self(data) loss = self.loss_fn(logits, targets) - self.log("val_loss", loss, prog_bar=True) + self.log("val/loss", loss, prog_bar=True) self.val_acc(logits, targets) - self.log("val_acc", self.val_acc, on_step=False, on_epoch=True, prog_bar=True) + self.log("val/acc", self.val_acc, on_step=False, on_epoch=True, prog_bar=True) def test_step(self, batch: Tuple[Tensor, Tensor], batch_idx: int) -> None: """Test step.""" data, targets = batch logits = self(data) self.test_acc(logits, targets) - self.log("test_acc", self.test_acc, on_step=False, on_epoch=True) + self.log("test/acc", self.test_acc, on_step=False, on_epoch=True) diff --git a/text_recognizer/models/transformer.py b/text_recognizer/models/transformer.py index bc7e313..6be0ac5 100644 --- a/text_recognizer/models/transformer.py +++ b/text_recognizer/models/transformer.py @@ -2,9 +2,7 @@ from typing import Dict, List, Optional, Union, Tuple, Type from omegaconf import DictConfig -from torch import nn -from torch import Tensor -import wandb +from torch import nn, Tensor from text_recognizer.data.emnist import emnist_mapping from text_recognizer.models.metrics import CharacterErrorRate @@ -44,24 +42,12 @@ class LitTransformerModel(LitBaseModel): # TODO: add case for sentence pieces return mapping, ignore_tokens - def _log_prediction(self, data: Tensor, pred: Tensor) -> None: - """Logs prediction on image with wandb.""" - pred_str = "".join( - self.mapping[i] for i in pred[0].tolist() if i != 3 - ) # pad index is 3 - try: - self.logger.experiment.log( - {"val_pred_examples": [wandb.Image(data[0], caption=pred_str)]} - ) - except AttributeError: - pass - def training_step(self, batch: Tuple[Tensor, Tensor], batch_idx: int) -> Tensor: """Training step.""" data, targets = batch logits = self.network(data, targets[:, :-1]) loss = self.loss_fn(logits, targets[:, 1:]) - self.log("train_loss", loss) + self.log("train/loss", loss) return loss def validation_step(self, batch: Tuple[Tensor, Tensor], batch_idx: int) -> None: @@ -70,17 +56,15 @@ class LitTransformerModel(LitBaseModel): logits = self.network(data, targets[:-1]) loss = self.loss_fn(logits, targets[1:]) - self.log("val_loss", loss, prog_bar=True) + self.log("val/loss", loss, prog_bar=True) pred = self.network.predict(data) - self._log_prediction(data, pred) self.val_cer(pred, targets) - self.log("val_cer", self.val_cer, on_step=False, on_epoch=True, prog_bar=True) + self.log("val/cer", self.val_cer, on_step=False, on_epoch=True, prog_bar=True) def test_step(self, batch: Tuple[Tensor, Tensor], batch_idx: int) -> None: """Test step.""" data, targets = batch pred = self.network.predict(data) - self._log_prediction(data, pred) self.test_cer(pred, targets) - self.log("test_cer", self.test_cer, on_step=False, on_epoch=True, prog_bar=True) + self.log("test/cer", self.test_cer, on_step=False, on_epoch=True, prog_bar=True) diff --git a/text_recognizer/networks/encoders/efficientnet/efficientnet.py b/text_recognizer/networks/encoders/efficientnet/efficientnet.py index a59abf8..fb4f002 100644 --- a/text_recognizer/networks/encoders/efficientnet/efficientnet.py +++ b/text_recognizer/networks/encoders/efficientnet/efficientnet.py @@ -27,7 +27,7 @@ class EfficientNet(nn.Module): def __init__( self, arch: str, - out_channels: int = 256, + out_channels: int = 1280, stochastic_dropout_rate: float = 0.2, bn_momentum: float = 0.99, bn_eps: float = 1.0e-3, @@ -37,7 +37,7 @@ class EfficientNet(nn.Module): self.arch = self.archs[arch] self.out_channels = out_channels self.stochastic_dropout_rate = stochastic_dropout_rate - self.bn_momentum = 1 - bn_momentum + self.bn_momentum = bn_momentum self.bn_eps = bn_eps self._conv_stem: nn.Sequential = None self._blocks: nn.Sequential = None @@ -70,9 +70,7 @@ class EfficientNet(nn.Module): for _ in range(args.num_repeats): self._blocks.append( MBConvBlock( - **args, - bn_momentum=self.bn_momentum, - bn_eps=self.bn_eps, + **args, bn_momentum=self.bn_momentum, bn_eps=self.bn_eps, ) ) args.in_channels = args.out_channels @@ -94,7 +92,7 @@ class EfficientNet(nn.Module): if self.stochastic_dropout_rate: stochastic_dropout_rate *= i / len(self._blocks) x = block(x, stochastic_dropout_rate=stochastic_dropout_rate) - self._conv_head(x) + x = self._conv_head(x) return x def forward(self, x: Tensor) -> Tensor: |