class DummyRegression(DummyRegressionPlainLightning, InnerEyeInference):
def __init__(self, in_features: int = 1, *args, **kwargs) -> None: # type: ignore
super().__init__(in_features=in_features, *args, **kwargs) # type: ignore
self.l_rate = 1e-1
self.dataset_split = ModelExecutionMode.TRAIN
activation = Identity()
layers = [
torch.nn.Linear(in_features=in_features, out_features=1, bias=True),
activation
]
self.model = torch.nn.Sequential(*layers) # type: ignore
def forward(self, x: Tensor) -> Tensor: # type: ignore
return self.model(x)
def training_step(self, batch, *args, **kwargs) -> torch.Tensor: # type: ignore
input, target = batch
prediction = self.forward(input)
loss = torch.nn.functional.mse_loss(prediction, target)
self.log("loss", loss, on_epoch=True, on_step=True)
return loss
def on_inference_start(self) -> None:
Path("on_inference_start.txt").touch()
self.inference_mse: Dict[ModelExecutionMode, float] = {}
def on_inference_epoch_start(self, dataset_split: ModelExecutionMode, is_ensemble_model: bool) -> None:
self.dataset_split = dataset_split
Path(f"on_inference_start_{self.dataset_split.value}.txt").touch()
self.mse = MeanSquaredError()
def inference_step(self, item: Tuple[Tensor, Tensor], batch_idx: int, model_output: torch.Tensor) -> None:
input, target = item
prediction = self.forward(input)
self.mse(prediction, target)
with Path(f"inference_step_{self.dataset_split.value}.txt").open(mode="a") as f:
f.write(f"{prediction.item()},{target.item()}\n")
def on_inference_epoch_end(self) -> None:
Path(f"on_inference_end_{self.dataset_split.value}.txt").touch()
self.inference_mse[self.dataset_split] = self.mse.compute().item()
self.mse.reset()
def on_inference_end(self) -> None:
Path("on_inference_end.txt").touch()
df = pd.DataFrame(columns=["Split", "MSE"],
data=[[split.value, mse] for split, mse in self.inference_mse.items()])
df.to_csv("metrics_per_split.csv", index=False)
class EncDecRegressionModel(_EncDecBaseModel):
"""Encoder decoder class for speech regression models.
Model class creates training, validation methods for setting up data
performing model forward pass.
"""
@classmethod
def list_available_models(cls) -> List[PretrainedModelInfo]:
"""
This method returns a list of pre-trained model which can be instantiated directly from NVIDIA's NGC cloud.
Returns:
List of available pre-trained models.
"""
result = []
return result
def __init__(self, cfg: DictConfig, trainer: Trainer = None):
if not cfg.get('is_regression_task', False):
raise ValueError(f"EndDecRegressionModel requires the flag is_regression_task to be set as true")
super().__init__(cfg=cfg, trainer=trainer)
def _setup_preprocessor(self):
return EncDecRegressionModel.from_config_dict(self._cfg.preprocessor)
def _setup_encoder(self):
return EncDecRegressionModel.from_config_dict(self._cfg.encoder)
def _setup_decoder(self):
return EncDecRegressionModel.from_config_dict(self._cfg.decoder)
def _setup_loss(self):
return MSELoss()
def _setup_metrics(self):
self._mse = MeanSquaredError()
self._mae = MeanAbsoluteError()
@property
def output_types(self) -> Optional[Dict[str, NeuralType]]:
return {"preds": NeuralType(tuple('B'), RegressionValuesType())}
@typecheck()
def forward(self, input_signal, input_signal_length):
logits = super().forward(input_signal=input_signal, input_signal_length=input_signal_length)
return logits.view(-1)
# PTL-specific methods
def training_step(self, batch, batch_idx):
audio_signal, audio_signal_len, targets, targets_len = batch
logits = self.forward(input_signal=audio_signal, input_signal_length=audio_signal_len)
loss = self.loss(preds=logits, labels=targets)
train_mse = self._mse(preds=logits, target=targets)
train_mae = self._mae(preds=logits, target=targets)
tensorboard_logs = {
'train_loss': loss,
'train_mse': train_mse,
'train_mae': train_mae,
'learning_rate': self._optimizer.param_groups[0]['lr'],
}
self.log_dict(tensorboard_logs)
return {'loss': loss}
def validation_step(self, batch, batch_idx, dataloader_idx: int = 0):
audio_signal, audio_signal_len, targets, targets_len = batch
logits = self.forward(input_signal=audio_signal, input_signal_length=audio_signal_len)
loss_value = self.loss(preds=logits, labels=targets)
val_mse = self._mse(preds=logits, target=targets)
val_mae = self._mae(preds=logits, target=targets)
return {'val_loss': loss_value, 'val_mse': val_mse, 'val_mae': val_mae}
def test_step(self, batch, batch_idx, dataloader_idx: int = 0):
logs = self.validation_step(batch, batch_idx, dataloader_idx)
return {'test_loss': logs['val_loss'], 'test_mse': logs['test_mse'], 'test_mae': logs['val_mae']}
def multi_validation_epoch_end(self, outputs, dataloader_idx: int = 0):
val_loss_mean = torch.stack([x['val_loss'] for x in outputs]).mean()
val_mse = self._mse.compute()
self._mse.reset()
val_mae = self._mae.compute()
self._mae.reset()
tensorboard_logs = {'val_loss': val_loss_mean, 'val_mse': val_mse, 'val_mae': val_mae}
return {'val_loss': val_loss_mean, 'val_mse': val_mse, 'val_mae': val_mae, 'log': tensorboard_logs}
def multi_test_epoch_end(self, outputs, dataloader_idx: int = 0):
test_loss_mean = torch.stack([x['test_loss'] for x in outputs]).mean()
test_mse = self._mse.compute()
self._mse.reset()
test_mae = self._mae.compute()
self._mae.reset()
tensorboard_logs = {'test_loss': test_loss_mean, 'test_mse': test_mse, 'test_mae': test_mae}
return {'test_loss': test_loss_mean, 'test_mse': test_mse, 'test_mae': test_mae, 'log': tensorboard_logs}
@torch.no_grad()
def transcribe(self, paths2audio_files: List[str], batch_size: int = 4) -> List[float]:
"""
Generate class labels for provided audio files. Use this method for debugging and prototyping.
Args:
paths2audio_files: (a list) of paths to audio files. \
Recommended length per file is approximately 1 second.
batch_size: (int) batch size to use during inference. \
Bigger will result in better throughput performance but would use more memory.
Returns:
A list of predictions in the same order as paths2audio_files
"""
predictions = super().transcribe(paths2audio_files, batch_size, logprobs=True)
return [float(pred) for pred in predictions]
def _update_decoder_config(self, labels, cfg):
OmegaConf.set_struct(cfg, False)
if 'params' in cfg:
cfg.params.num_classes = 1
else:
cfg.num_classes = 1
OmegaConf.set_struct(cfg, True)