Merge pull request #275 from VectorInstitute/dbe/updating_flamby_install

Updating the FLamby installation instructions to reflect some changes on their end.

research/flamby/README.md

@@ -1,20 +1,4 @@
-### Installing the Flamby dependencies from Fixed Requirements File
-
-__NOTE__: The standard workflow discussed by FLamby is in the next section, but is currently broken due to dependency changes.
-
-Create a python environment with your preferred env manager. We'll use conda below
-``` bash
-conda create -n flamby_fl4health python=3.10
-conda activate flamby_fl4health
-```
-Install the dependencies of both FLamby and FL4Health using the fixed requirements file at `research/flamby/flamby_requirements.txt`.
-```bash
-cd <fl4health_repository>
-pip install --src <path/to/clone/flamby/repository/to> -r research/flamby/flamby_requirements.txt
-```
-Note that this installation will clone the FLamby repo to the path provided to --src and then install FLamby as a package.
-
-### Installing the Flamby dependencies (Old Workflow)
+### Installing the Flamby dependencies
 
 __NOTE__: The workflow below is normally the smoothest way to construct the FLamby + FL4Health environment required to run the FLamby experiments. However, with a recent upgrade to MonAI, some of the functionality that FLamby depends on are broken. Until this is fixed, the workflow below will not work.
 
@@ -33,8 +17,56 @@ cd <fl4health_repository>
 pip install --upgrade pip poetry
 poetry install --with "dev, dev-local, test, codestyle"
 cd <FLamby_repository>
-pip install -e ".[all_extra]"
+pip install -e ".[cam16, heart, isic2019, ixi, lidc, tcga]"
+```
+__NOTE__: We avoid installing Fed-KITS2019, as it requires a fairly old version on nnUnet, which we no longer support in our library.
+
+In addition, you'll have to edit the code for `FedIXITiny` in `flamby/datasets/fed_ixi/datasets.py` replacing the following
+
+<table>
+<tr>
+<th>Old</th>
+<th>New</th>
+</tr>
+<tr>
+<td>
+
+``` python
+from monai.transforms import (
+    AddChannel,
+    ...
+```
+</td>
+<td>
+
+``` python
+from monai.transforms import (
+    EnsureChannelFirst,
+    ...
 ```
+</td>
+</tr>
+</tr>
+<tr>
+<td>
+
+```python
+default_transform = Compose(
+    [ToTensor(), AddChannel(), Resize(self.common_shape)]
+```
+</td>
+<td>
+
+```python
+default_transform = Compose(
+    [ToTensor(), EnsureChannelFirst(channel_dim="no_channel"), Resize(self.common_shape)]
+)
+```
+</td>
+</tr>
+</table>
+
+This is because AddChannel was removed in Version 1.3 of MonAI
 
 ### Downloading the Fed ISIC 2019 Dataset
 

research/flamby/single_node_trainer.py

@@ -95,7 +95,7 @@ def validate(self, val_metric_mngr: MetricManager) -> None:
             for input, target in self.val_loader:
                 input, target = input.to(self.device), target.to(self.device)
 
-                preds = self.model(input)
+                preds = {"predictions": self.model(input)}
                 batch_loss = self.criterion(preds["predictions"], target)
                 running_loss += batch_loss.item()
                 val_metric_mngr.update(preds, target)