Add preprocessing and postprocessing to forward method

tiktorch/server/grpc/training_servicer.py

@@ -78,7 +78,6 @@ def Export(self, request: training_pb2.ExportRequest, context):
         session.client.export(Path(request.filePath))
         return utils_pb2.Empty()
 
-
     def Predict(self, request: training_pb2.PredictRequest, context):
         session = self._getTrainerSession(context, request.sessionId.id)
         tensors = [torch.tensor(pb_tensor_to_numpy(pb_tensor)) for pb_tensor in request.tensors]

tiktorch/trainer.py

@@ -8,6 +8,7 @@
 
 import torch
 import yaml
+from pytorch3dunet.augment.transforms import Compose, Normalize, Standardize, ToTensor
 from pytorch3dunet.datasets.utils import get_train_loaders
 from pytorch3dunet.unet3d.losses import get_loss_criterion
 from pytorch3dunet.unet3d.metrics import get_evaluation_metric
@@ -97,6 +98,8 @@ def __init__(
         self,
         model,
         device,
+        in_channels,
+        out_channels,
         optimizer,
         lr_scheduler,
         loss_criterion,
@@ -139,6 +142,8 @@ def __init__(
             pre_trained=pre_trained,
             **kwargs,
         )
+        self._in_channels = in_channels
+        self._out_channels = out_channels
         self._device = device
         self.logs_callbacks: LogsCallbacks = BaseCallbacks()
         self.should_stop_callbacks: Callbacks = ShouldStopCallbacks()
@@ -170,16 +175,60 @@ def forward(self, input_tensors: List[torch.Tensor]):
         if self.is_2d_model() and z != 1:
             raise ValueError(f"2d model detected but z != 1 for tensor {input_tensor.shape}")
 
+        # todo: normalization need to be consistent with the training one (it should be retrieved by the config)
+        preprocessor = Compose([Standardize(), ToTensor(expand_dims=True)])
+        input_tensor = self._apply_transformation(compose=preprocessor, tensor=input_tensor)
+
+        def apply_final_activation(input_tensors) -> torch.Tensor:
+            if self.model.final_activation is not None:
+                return self.model.final_activation(input_tensors)
+            return input_tensors
+
         with torch.no_grad():
             if self.is_2d_model():
                 input_tensor = input_tensor.squeeze(dim=-3)  # b, c, [z], y, x
                 predictions = self.model(input_tensor.to(self._device))
+                predictions = apply_final_activation(predictions)
                 predictions = predictions.unsqueeze(dim=-3)  # for consistency
             else:
                 predictions = self.model(input_tensor.to(self._device))
+                predictions = apply_final_activation(predictions)
 
+        predictions = predictions.cpu()
+
+        # this needs to be exposed as well
+        # currently we scale the features from 0 - 1 (consistent scale for rendering across channels)
+        postprocessor = Compose([Normalize(norm01=True), ToTensor(expand_dims=True)])
+        predictions = self._apply_transformation(compose=postprocessor, tensor=predictions)
         return predictions
 
+    def _apply_transformation(self, compose: Compose, tensor: torch.Tensor) -> torch.Tensor:
+        """
+        To apply transformations pytorch 3d unet requires shape of DxHxW or CxDxHxW
+        """
+        b, c, z, y, x = tensor.shape
+        non_batch_tensors = []
+        for batch_idx in range(b):
+            # drop batch
+            non_batch_tensor = tensor[batch_idx, :]
+
+            # drop channel dim if single channel
+            dropped_channel = non_batch_tensor.squeeze(dim=-4) if self.is_input_single_channel() else non_batch_tensor
+
+            # adds channel back with the`expand_dims`
+            transformed_tensor = compose(dropped_channel.detach().cpu().numpy())
+
+            non_batch_tensors.append(transformed_tensor)
+
+        # add batch dim again
+        return torch.stack(non_batch_tensors, dim=0)
+
+    def is_input_single_channel(self) -> bool:
+        return self._in_channels == 1
+
+    def is_output_single_channel(self) -> bool:
+        return self._out_channels == 1
+
     @staticmethod
     def get_axes_from_tensor(tensor: torch.Tensor) -> Tuple[str, ...]:
         if tensor.ndim != 5:
@@ -238,6 +287,9 @@ def parse(self) -> Trainer:
 
         model = get_model(config["model"])
 
+        in_channels = config["model"]["in_channels"]
+        out_channels = config["model"]["out_channels"]
+
         if torch.cuda.device_count() > 1 and not config["device"] == "cpu":
             model = nn.DataParallel(model)
         if torch.cuda.is_available() and not config["device"] == "cpu":
@@ -266,6 +318,8 @@ def parse(self) -> Trainer:
 
         return Trainer(
             device=config["device"],
+            in_channels=in_channels,
+            out_channels=out_channels,
             model=model,
             optimizer=optimizer,
             lr_scheduler=lr_scheduler,