Merge branch 'intensity-masking' of https://github.com/computational-cell-analytics/lightsheet-moser into intensity-masking

Author: constantinpape

reproducibility/label_components/IHC_v4c_fig2.json

@@ -1,6 +1,8 @@
 import napari
 import zarr
 
+from skimage.feature import peak_local_max
+
 
 resolution = [3.0, 1.887779, 1.887779]
 positions = [
@@ -29,7 +31,7 @@ def _load_prediction(bb):
 
 
 def _load_prediction_debug():
-    path = "./debug-pred/pred-v5.h5"
+    path = "./debug-pred/pred-v7.h5"
     with zarr.open(path, "r") as f:
         pred = f["pred"][:]
     return pred
@@ -45,10 +47,14 @@ def check_detection(position, halo=[32, 384, 384]):
     # pred = _load_prediction(bb)
     pred = _load_prediction_debug()
 
+    print("Runnign detection")
+    det_new = peak_local_max(pred, min_distance=4, threshold_abs=0.5)
+
     v = napari.Viewer()
     v.add_image(pv)
     v.add_image(pred)
     v.add_labels(detections_mobie)
+    v.add_points(det_new)
     napari.run()
 
 

scripts/la-vision/check_detections.py

@@ -30,21 +30,23 @@ def run_prediction(position, halo=[32, 384, 384]):
     print(mean, std)
     preproc = partial(standardize, mean=mean, std=std)
 
-    block_shape = (24, 256, 256)
-    halo = (8, 64, 64)
+    block_shape = (12, 128, 128)
+    halo = (10, 64, 64)
 
-    model_path = "/mnt/vast-nhr/projects/nim00007/data/moser/cochlea-lightsheet/trained_models/SGN/sgn-detection-v1.pt"
+    # model_path = "/mnt/vast-nhr/projects/nim00007/data/moser/cochlea-lightsheet/trained_models/SGN/sgn-detection-v1.pt"
+    model_path = "/mnt/vast-nhr/home/pape41/u12086/Work/my_projects/flamingo-tools/scripts/la-vision/checkpoints/sgn-detection-v3.pt"
     model = torch.load(model_path, weights_only=False)
 
-    def postproc(x):
-        x = np.clip(x, 0, 1)
-        max_ = np.percentile(x, 99)
-        x = x / max_
-        return x
+    # def postproc(x):
+    #     x = np.clip(x, 0, 1)
+    #     max_ = np.percentile(x, 99)
+    #     x = x / max_
+    #     return x
+    postproc = None
 
     pred = predict_with_halo(pv, model, [0], block_shape, halo, preprocess=preproc, postprocess=postproc).squeeze()
 
-    pred_name = "pred-v5"
+    pred_name = "pred-v7"
     out_folder = "./debug-pred"
     os.makedirs(out_folder, exist_ok=True)
 

scripts/la-vision/debug_prediction.py

@@ -9,17 +9,18 @@
 sys.path.append("../synapse_marker_detection")
 
 
-def export_model(input_, output):
-    model = load_model(input_, device="cpu")
+def export_model(input_, output, latest):
+    model = load_model(input_, device="cpu", name="latest" if latest else "best")
     torch.save(model, output)
 
 
 def main():
     parser = argparse.ArgumentParser()
     parser.add_argument("-i", "--input", required=True)
     parser.add_argument("-o", "--output", required=True)
+    parser.add_argument("--latest", action="store_true")
     args = parser.parse_args()
-    export_model(args.input, args.output)
+    export_model(args.input, args.output, latest=args.latest)
 
 
 if __name__ == "__main__":

scripts/la-vision/export_model.py

@@ -18,7 +18,7 @@
 TRAIN = os.path.join(ROOT, "images")
 TRAIN_EMPTY = os.path.join(ROOT, "empty_images")
 
-LABEL = os.path.join(ROOT, "centroids")
+LABEL = os.path.join(ROOT, "centroids_v2")
 LABEL_EMPTY = os.path.join(ROOT, "empty_centroids")
 
 
@@ -46,13 +46,13 @@ def _get_paths(split, train_folder, label_folder, n=None):
 
 def get_paths(split):
     image_paths, label_paths = _get_paths(split, TRAIN, LABEL)
-    empty_image_paths, empty_label_paths = _get_paths(split, TRAIN_EMPTY, LABEL_EMPTY, n=4)
+    empty_image_paths, empty_label_paths = _get_paths(split, TRAIN_EMPTY, LABEL_EMPTY, n=6)
     return image_paths + empty_image_paths, label_paths + empty_label_paths
 
 
 def train():
 
-    model_name = "sgn-low-res-detection-v1"
+    model_name = "sgn-low-res-detection-v6"
 
     train_paths, train_label_paths = get_paths("train")
     val_paths, val_label_paths = get_paths("val")
@@ -63,8 +63,8 @@ def train():
     print(len(train_paths), "tomograms for training")
     print(len(val_paths), "tomograms for validation")
 
-    patch_shape = [48, 256, 256]
-    batch_size = 8
+    patch_shape = [32, 144, 144]
+    batch_size = 16
     check = False
 
     checkpoint_path = f"./checkpoints/{model_name}"
@@ -78,6 +78,10 @@ def train():
             f, indent=2, sort_keys=True
         )
 
+    # For marmoset model
+    sigma = (0.6, 3, 3)
+    # For mouse model
+    # sigma = (1, 4, 4)
     supervised_training(
         name=model_name,
         train_paths=train_paths,
@@ -92,7 +96,7 @@ def train():
         out_channels=1,
         augmentations=None,
         eps=1e-5,
-        sigma=4,
+        sigma=sigma,
         lower_bound=None,
         upper_bound=None,
         test_paths=test_paths,

scripts/la-vision/train_sgn_detection.py

@@ -206,12 +206,14 @@ def _get_sample(self, index):
                     raise RuntimeError(f"Could not sample a valid batch in {self.max_sampling_attempts} attempts")
 
         # For synapse detection.
-        label = process_labels(coords, shape, self.sigma, self.eps, bb=bb)
+        # label = process_labels(coords, shape, self.sigma, self.eps, bb=bb)
 
         # For SGN detection with data specfic hacks
-        # label = process_labels_hacky(coords, shape, self.sigma, self.eps, bb=bb)
-        # gap = 8
-        # raw_patch, label = raw_patch[gap:-gap], label[gap:-gap]
+        label = process_labels_hacky(coords, shape, self.sigma, self.eps, bb=bb)
+        # Having this halo actually makes sense in general!
+        gap = 8
+        gap_bb = np.s_[gap:-gap, gap:-gap, gap:-gap]
+        raw_patch, label = raw_patch[gap_bb], label[gap_bb]
 
         have_label_channels = label.ndim == 4
         if have_label_channels: