Models refactoring

.github/workflows/test.yml

@@ -115,7 +115,6 @@ jobs:
         env:
           CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}
 
-
   deploy:
       # this will run when you have tagged a commit, starting with "v*"
       # and requires that you have put your twine API key in your 

.gitignore

@@ -153,4 +153,6 @@ dmypy.json
 .idea/
 
 .DS_Store
+data/
+BentoML/
 

docs/source/dcp_server_installation.rst

@@ -83,17 +83,18 @@ The models are currently integrated into DCP:
 
 - **Instance** Segmentation: 
 
-  - ``CustomCellposeModel``: Inherits from cellpose.models.CellposeModel, see `here <https://cellpose.readthedocs.io/en/latest/api.html#cellposemodel>`__ for more information.
+  - ``CustomCellpose``: Inherits from cellpose.models.CellposeModel, see `here <https://cellpose.readthedocs.io/en/latest/api.html#cellposemodel>`__ for more information.
 - **Semantic** Segmentation: 
 
   - ``UNet``: A vanilla U-Net model, trained on the full images
-- **Panoptic** Segmentation: 
+- **Multi Class Instance** Segmentation: 
 
-  - ``CellposePatchCNN``: Includes a segmentor for instance segmentation, sequentially followed by a classifier for semantic segmentation. The segmentor can only be ``CustomCellposeModel`` model, while the classifier can be one of:
+  - ``Inst2MultiSeg``: Includes a segmentor for instance segmentation, sequentially followed by a classifier for semantic segmentation. The segmentor can only be ``CustomCellposeModel`` model, while the classifier can be one of:
 
-    - ``CellClassifierFCNN`` or "FCNN" (in config): A CNN model for obtaining class labels, trained on images patches of individual objects, extarcted using the instance mask from the previous step
-    - ``CellClassifierShallowModel`` or "RandomForest" (in config): A Random Forest model for obtaining class labels, trained on shape and intensity features of the objects, extracted using the instance mask from the previous step.
-  - UNet: If the post-processing argument is set, then the instance mask is deduced from the labels mask. Will not be able to handle touching objects 
+    - ``PatchClassifier`` or "FCNN" (in config): A CNN model for obtaining class labels, trained on images patches of individual objects, extarcted using the instance mask from the previous step
+    - ``FeatureClassifier`` or "RandomForest" (in config): A Random Forest model for obtaining class labels, trained on shape and intensity features of the objects, extracted using the instance mask from the previous step.
+  - ``MultiCellpose``: Includes **n** CustomCellpose models, where n equals the number of classes, stacked such that each model predicts only the object corresponding to each class.
+  - ``UNet``: If the post-processing argument is set, then the instance mask is deduced from the labels mask. Will not be able to handle touching objects 
 
 
 Running with Docker 

docs/source/index.rst

@@ -31,7 +31,7 @@ DCP handles all kinds of **segmentation tasks**! Try it out if you need to do:
 
 - **Instance** segmentation
 - **Semantic** segmentation
-- **Panoptic** segmentation
+- **Multi-class instance** segmentation
 
 Toy data
 --------

src/client/dcp_client/gui/napari_window.py

@@ -44,7 +44,7 @@ def __init__(self, app: Application):
         layout.addWidget(main_window, 0, 0, 1, 4)
 
         # select the first seg as the currently selected layer if there are any segs
-        if len(self.seg_files):
+        if len(self.seg_files) and len(self.viewer.layers[get_path_stem(self.seg_files[0])].data.shape) > 2:
             self.cur_selected_seg = self.viewer.layers.selection.active.name
             self.layer = self.viewer.layers[self.cur_selected_seg]
             self.viewer.layers.selection.events.changed.connect(self.on_seg_channel_changed)
@@ -68,7 +68,7 @@ def __init__(self, app: Application):
 
             self.qctrl = self.viewer.window.qt_viewer.controls.widgets[self.layer]
 
-            if self.layer.data.shape[0] >= 2:
+            if len(self.layer.data.shape) > 2:
                 # User hint
                 message_label = QLabel('Choose an active mask')
                 message_label.setAlignment(Qt.AlignRight)

src/client/dcp_client/main.py

@@ -17,7 +17,7 @@
 def main():
     settings.init()
     dir_name = path.dirname(path.abspath(sys.argv[0]))
-    server_config = read_config('server', config_path = path.join(dir_name, 'config.cfg'))
+    server_config = read_config('server', config_path = path.join(dir_name, 'config.yaml'))
 
     image_storage = FilesystemImageStorage()
     ml_model = BentomlModel()

src/client/dcp_client/utils/utils.py

@@ -6,7 +6,7 @@
 from skimage.draw import polygon_perimeter
 
 from pathlib import Path, PurePath
-import json
+import yaml
 
 from dcp_client.utils import settings
 
@@ -36,18 +36,18 @@
         else:
             return super().icon(type)
 
-def read_config(name, config_path = 'config.cfg') -> dict:   
-    """ Reads the configuration file
+def read_config(name, config_path = 'config.yaml') -> dict:   
+    """Reads the configuration file
 
     :param name: name of the section you want to read (e.g. 'setup','train')
     :type name: string
-    :param config_path: path to the configuration file, defaults to 'config.cfg'
+    :param config_path: path to the configuration file, defaults to 'config.yaml'
     :type config_path: str, optional
     :return: dictionary from the config section given by name
     :rtype: dict
     """     
     with open(config_path) as config_file:
-        config_dict = json.load(config_file)
+        config_dict =  yaml.safe_load(config_file) # json.load(config_file) for .cfg file
         # Check if config file has main mandatory keys
         assert all([i in config_dict.keys() for i in ['server']])
         return config_dict[name]
@@ -140,16 +140,18 @@
             # get a binary mask only of object
             single_obj_mask = np.zeros_like(instance_mask)
             single_obj_mask[instance_mask==instance_id] = 1
-            # compute contours for mask
-            contours = find_contours(single_obj_mask, contours_level)
-            # sometimes little dots appeas as additional contours so remove these
-            if len(contours)>1: 
-                contour_sizes = [contour.shape[0] for contour in contours]
-                contour = contours[contour_sizes.index(max(contour_sizes))].astype(int)
-            else: contour = contours[0]
-            # and draw onto contours mask
-            rr, cc = polygon_perimeter(contour[:, 0], contour[:, 1], contour_mask.shape)
-            contour_mask[rr, cc] = instance_id
+            try:
+                # compute contours for mask
+                contours = find_contours(single_obj_mask, contours_level)
+                # sometimes little dots appeas as additional contours so remove these
+                if len(contours)>1: 
+                    contour_sizes = [contour.shape[0] for contour in contours]
+                    contour = contours[contour_sizes.index(max(contour_sizes))].astype(int)
+                else: contour = contours[0]
+                # and draw onto contours mask
+                rr, cc = polygon_perimeter(contour[:, 0], contour[:, 1], contour_mask.shape)
+                contour_mask[rr, cc] = instance_id
+            except: print("Could not create contour for instance id", instance_id)
         return contour_mask
 
     @staticmethod

src/server/dcp_server/config.cfg → src/server/dcp_server/config.yaml

@@ -1,34 +1,40 @@
 {
     "setup": {
         "segmentation": "GeneralSegmentation",
-        "model_to_use": "CellposePatchCNN", 
+        "model_to_use": "Inst2MultiSeg", 
         "accepted_types": [".jpg", ".jpeg", ".png", ".tiff", ".tif"],
         "seg_name_string": "_seg"
     },
 
     "service": {
         "runner_name": "bento_runner",
-        "bento_model_path": "cppcnn",
+        "bento_model_path": "cells",
         "service_name": "data-centric-platform",
         "port": 7010
     },
 
     "model": { 
+        "segmentor_name": "Cellpose",
         "segmentor": {
             "model_type": "cyto"
 
         },
+        "classifier_name": "PatchClassifier",
         "classifier":{
-            "model_class": "FCNN",
             "in_channels": 1,
             "num_classes": 2,
-            "black_bg": "False",
-            "include_mask": "False"
+            "features":[64,128,256,512],
+            "black_bg": False,
+            "include_mask": True
         }
     },
 
     "data": {
-        "data_root": "data"
+        "data_root": "data",
+        "patch_size": 64,
+        "noise_intensity": 5,
+        "gray": True,
+        "rescale": True
     },
 
     "train":{
@@ -38,12 +44,7 @@
             "min_train_masks": 1
         },
         "classifier":{
-            "train_data":{
-                "patch_size": 64,
-                "noise_intensity": 5,
-                "num_classes": 2
-            },
-            "n_epochs": 10,
+            "n_epochs": 100,
             "lr": 0.001,
             "batch_size": 1,
             "optimizer": "Adam"
@@ -58,10 +59,6 @@
             "batch_size": 1
         },
         "classifier": {
-            "data":{
-                "patch_size": 64,
-                "noise_intensity": 5
-            }
         },
         "mask_channel_axis": 0
     }

src/server/dcp_server/config_instance.cfg → src/server/dcp_server/config_instance.yaml

@@ -1,14 +1,14 @@
 {
     "setup": {
         "segmentation": "GeneralSegmentation",
-        "model_to_use": "CustomCellposeModel", 
+        "model_to_use": "CustomCellpose", 
         "accepted_types": [".jpg", ".jpeg", ".png", ".tiff", ".tif"],
         "seg_name_string": "_seg"
     },
 
     "service": {
-        "runner_name": "cellpose_runner",
-        "bento_model_path": "cells",
+        "runner_name": "bento_runner",
+        "bento_model_path": "cells",   
         "service_name": "data-centric-platform",
         "port": 7010
     },
@@ -20,12 +20,14 @@
     },
 
     "data": {
-        "data_root": "data"
+        "data_root": "data",
+        "gray": True,
+        "rescale": True
     },
 
     "train":{
         "segmentor":{
-            "n_epochs": 10,
+            "n_epochs": 5,
             "channels": [0,0],
             "min_train_masks": 1
         }

src/server/dcp_server/config_semantic.cfg → src/server/dcp_server/config_semantic.yaml

@@ -15,15 +15,16 @@
 
     "model": { 
         "classifier":{
-            "model_class": "UNet",
             "in_channels": 1,
-            "num_classes": 3,
+            "num_classes": 2,
             "features":[64,128,256,512]
         }
     },
 
     "data": {
-        "data_root": "data"
+        "data_root": "data",
+        "gray": True,
+        "rescale": True
     },
 
     "train":{
@@ -37,7 +38,9 @@
 
     "eval":{
         "classifier": {
+
         },
-        "mask_channel_axis": null
+        "compute_instance": True,
+        "mask_channel_axis": 0
     }
 }

src/server/dcp_server/main.py

@@ -1,7 +1,7 @@
 import subprocess
 from os import path
 import sys
-from utils import read_config
+from dcp_server.utils.helpers import read_config
 
 def main(): 
     """
@@ -17,7 +17,7 @@ def main():
 
     local_path = path.join(__file__, '..')
     dir_name = path.dirname(path.abspath(sys.argv[0]))
-    service_config = read_config('service', config_path = path.join(dir_name, 'config.cfg'))
+    service_config = read_config('service', config_path = path.join(dir_name, 'config.yaml'))
     port = str(service_config['port'])
 
     subprocess.run([