Remove deterministic argument from som function in pyFlowSOM (#893)

* Remove deterministic argument from SOM

* Add explicit seed setting for Pixie SOM training

* Update pyFlowSOM to 0.1.14

* Remove requirements.txt as well

* Update pyFlowSOM to 0.1.14

* Abstract seed to PixieSOMCluster and ensure it actually gets passed to pyFlowSOM.som

* Add explicit seed params to pixel and cell clustering notebooks

pyproject.toml

@@ -28,7 +28,7 @@ dependencies = [
     "palettable>=3.3.0,<4",
     "pandas>=1.3,<2",
     "pillow>=9,<10",
-    "pyFlowSOM==0.1.13",
+    "pyFlowSOM==0.1.14",
     "requests>=2.20,<3",
     "scikit-image>=0.19,<0.20",
     "scikit-learn>=1.1,<2",

src/ark/phenotyping/cell_cluster_utils.py

@@ -417,7 +417,7 @@ def train_cell_som(fovs, channels, base_dir, pixel_data_dir, cell_table_path,
                    pc_chan_avg_name='pc_chan_avg.csv',
                    som_weights_name='cell_som_weights.feather',
                    weighted_cell_channel_name='weighted_cell_channel.feather',
-                   xdim=10, ydim=10, lr_start=0.05, lr_end=0.01, num_passes=1):
+                   xdim=10, ydim=10, lr_start=0.05, lr_end=0.01, num_passes=1, seed=42):
     """Run the SOM training on the number of pixel/meta clusters in each cell of each fov
 
     Saves the SOM weights to `base_dir/som_weights_name`. Computes and saves weighted
@@ -461,6 +461,8 @@ def train_cell_som(fovs, channels, base_dir, pixel_data_dir, cell_table_path,
             The end learning rate for the SOM, decays from `lr_start`
         num_passes (int):
             The number of training passes to make through the dataset
+        seed (int):
+            The random seed to use for training the SOM
 
     Returns:
         cluster_helpers.CellSOMCluster:
@@ -507,7 +509,8 @@ def train_cell_som(fovs, channels, base_dir, pixel_data_dir, cell_table_path,
     # define the cell SOM cluster object
     cell_pysom = cluster_helpers.CellSOMCluster(
         cluster_counts_size_norm_path, som_weights_path, fovs, cluster_count_cols,
-        num_passes=num_passes, xdim=xdim, ydim=ydim, lr_start=lr_start, lr_end=lr_end
+        num_passes=num_passes, xdim=xdim, ydim=ydim, lr_start=lr_start, lr_end=lr_end,
+        seed=seed
     )
 
     # train the SOM weights

src/ark/phenotyping/cluster_helpers.py

@@ -19,7 +19,8 @@
 class PixieSOMCluster(ABC):
     @abstractmethod
     def __init__(self, weights_path: pathlib.Path, columns: List[str], num_passes: int = 1,
-                 xdim: int = 10, ydim: int = 10, lr_start: float = 0.05, lr_end: float = 0.01):
+                 xdim: int = 10, ydim: int = 10, lr_start: float = 0.05, lr_end: float = 0.01,
+                 seed=42):
         """Generic implementation of a pyFlowSOM runner
 
         Args:
@@ -37,6 +38,8 @@ def __init__(self, weights_path: pathlib.Path, columns: List[str], num_passes: i
                 The initial learning rate.
             lr_end (float):
                 The learning rate to decay to
+            seed (int):
+                The random seed to use for training.
         """
         self.weights_path = weights_path
         self.weights = None if not os.path.exists(weights_path) else feather.read_dataframe(
@@ -48,6 +51,7 @@ def __init__(self, weights_path: pathlib.Path, columns: List[str], num_passes: i
         self.ydim = ydim
         self.lr_start = lr_start
         self.lr_end = lr_end
+        self.seed = seed
 
     @abstractmethod
     def normalize_data(self) -> pd.DataFrame:
@@ -65,10 +69,10 @@ def train_som(self, data: pd.DataFrame):
             data (pandas.DataFrame):
                 The input data to train the SOM on.
         """
-        # make sure to run a deterministic SOM for reproducibility purposes
+
         som_weights = som(
             data=data.values, xdim=self.xdim, ydim=self.ydim, rlen=self.num_passes,
-            alpha_range=(self.lr_start, self.lr_end), deterministic=True
+            alpha_range=(self.lr_start, self.lr_end), seed=self.seed
         )
 
         # ensure dimensions of weights are flattened
@@ -118,7 +122,7 @@ class PixelSOMCluster(PixieSOMCluster):
     def __init__(self, pixel_subset_folder: pathlib.Path, norm_vals_path: pathlib.Path,
                  weights_path: pathlib.Path, fovs: List[str], columns: List[str],
                  num_passes: int = 1, xdim: int = 10, ydim: int = 10,
-                 lr_start: float = 0.05, lr_end: float = 0.01):
+                 lr_start: float = 0.05, lr_end: float = 0.01, seed=42):
         """Creates a pixel SOM cluster object derived from the abstract PixieSOMCluster
 
         Args:
@@ -141,10 +145,12 @@ def __init__(self, pixel_subset_folder: pathlib.Path, norm_vals_path: pathlib.Pa
             lr_start (float):
                 The initial learning rate.
             lr_end (float):
-                The learning rate to decay to
+                The learning rate to decay to.
+            seed (int):
+                The random seed to use.
         """
         super().__init__(
-            weights_path, columns, num_passes, xdim, ydim, lr_start, lr_end
+            weights_path, columns, num_passes, xdim, ydim, lr_start, lr_end, seed
         )
 
         # path validation
@@ -232,7 +238,8 @@ def assign_som_clusters(self, external_data: pd.DataFrame) -> pd.DataFrame:
 class CellSOMCluster(PixieSOMCluster):
     def __init__(self, cell_data_path: pathlib.Path, weights_path: pathlib.Path,
                  fovs: List[str], columns: List[str], num_passes: int = 1,
-                 xdim: int = 10, ydim: int = 10, lr_start: float = 0.05, lr_end: float = 0.01):
+                 xdim: int = 10, ydim: int = 10, lr_start: float = 0.05, lr_end: float = 0.01,
+                 seed=42):
         """Creates a cell SOM cluster object derived from the abstract PixieSOMCluster
 
         Args:
@@ -253,10 +260,12 @@ def __init__(self, cell_data_path: pathlib.Path, weights_path: pathlib.Path,
             lr_start (float):
                 The initial learning rate.
             lr_end (float):
-                The learning rate to decay to
+                The learning rate to decay to.
+            seed (int):
+                The random seed to use.
         """
         super().__init__(
-            weights_path, columns, num_passes, xdim, ydim, lr_start, lr_end
+            weights_path, columns, num_passes, xdim, ydim, lr_start, lr_end, seed
         )
 
         # path validation

src/ark/phenotyping/pixel_cluster_utils.py

@@ -874,7 +874,7 @@ def train_pixel_som(fovs, channels, base_dir,
                     subset_dir='pixel_mat_subsetted',
                     norm_vals_name='post_rowsum_chan_norm.feather',
                     som_weights_name='pixel_som_weights.feather', xdim=10, ydim=10,
-                    lr_start=0.05, lr_end=0.01, num_passes=1):
+                    lr_start=0.05, lr_end=0.01, num_passes=1, seed=42):
     """Run the SOM training on the subsetted pixel data.
 
     Saves SOM weights to `base_dir/som_weights_name`.
@@ -902,6 +902,8 @@ def train_pixel_som(fovs, channels, base_dir,
             The end learning rate for the SOM, decays from `lr_start`
         num_passes (int):
             The number of training passes to make through the dataset
+        seed (int):
+            The random seed to use for training the SOM
 
     Returns:
         cluster_helpers.PixelSOMCluster:
@@ -930,7 +932,8 @@ def train_pixel_som(fovs, channels, base_dir,
     # define the pixel SOM cluster object
     pixel_pysom = cluster_helpers.PixelSOMCluster(
         subsetted_path, norm_vals_path, som_weights_path, fovs, channels,
-        num_passes=num_passes, xdim=xdim, ydim=ydim, lr_start=lr_start, lr_end=lr_end
+        num_passes=num_passes, xdim=xdim, ydim=ydim, lr_start=lr_start, lr_end=lr_end,
+        seed=seed
     )
 
     # train the SOM weights

templates/2_Pixie_Cluster_Pixels.ipynb

@@ -451,7 +451,8 @@
     "    subset_dir=pixel_subset_dir,\n",
     "    norm_vals_name=norm_vals_name,\n",
     "    som_weights_name=pixel_som_weights_name,\n",
-    "    num_passes=1\n",
+    "    num_passes=1,\n",
+    "    seed=42\n",
     ")"
    ]
   },
@@ -878,7 +879,7 @@
  ],
  "metadata": {
   "kernelspec": {
-   "display_name": "ark38",
+   "display_name": "Python 3 (ipykernel)",
    "language": "python",
    "name": "python3"
   },
@@ -892,7 +893,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.8.16"
+   "version": "3.8.15"
   },
   "toc-autonumbering": false,
   "toc-showcode": true,

templates/3_Pixie_Cluster_Cells.ipynb

@@ -333,7 +333,8 @@
     "    pc_chan_avg_name=pc_chan_avg_name,\n",
     "    som_weights_name=cell_som_weights_name,\n",
     "    weighted_cell_channel_name=weighted_cell_channel_name,\n",
-    "    num_passes=1\n",
+    "    num_passes=1,\n",
+    "    seed=42\n",
     ")"
    ]
   },
@@ -749,7 +750,7 @@
  ],
  "metadata": {
   "kernelspec": {
-   "display_name": "ark38",
+   "display_name": "Python 3 (ipykernel)",
    "language": "python",
    "name": "python3"
   },
@@ -763,7 +764,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.8.16"
+   "version": "3.8.15"
   },
   "vscode": {
    "interpreter": {