Update training scripts

Author: constantinpape

scripts/README.md

@@ -13,6 +13,7 @@ conda install -c conda-forge mobie_utils
 ## Training
 
 Contains the scripts for training a U-Net that predicts foreground probabilties and normalized object distances.
+It also contains a documentation for how to run training on new annotated data.
 
 
 ## Prediction

scripts/data_transfer/README.md

@@ -33,3 +33,10 @@ Try to automate via https://github.com/jborean93/smbprotocol see `sync_smb.py` f
 
 For transfering back MoBIE results.
 ...
+
+# Data Transfer Huisken
+
+See "Transfer via smbclient" above:
+```
+smbclient \\\\wfs-biologie-spezial.top.gwdg.de\\UBM1-all\$\\ -U GWDG\\pape41
+```

scripts/training/check_training_data.py

@@ -1,42 +1,57 @@
+import argparse
 import os
-from glob import glob
 
 import imageio.v3 as imageio
 import napari
 import numpy as np
 
-root = "/home/pape/Work/data/moser/lightsheet"
+from train_distance_unet import get_image_and_label_paths
+from tqdm import tqdm
 
+# Root folder on my laptop.
+# This is just for convenience, so that I don't have to pass
+# the root argument during development.
+ROOT_CP = "/home/pape/Work/data/moser/lightsheet"
 
-def check_visually(check_downsampled=False):
-    if check_downsampled:
-        images = sorted(glob(os.path.join(root, "images_s2", "*.tif")))
-        masks = sorted(glob(os.path.join(root, "masks_s2", "*.tif")))
-    else:
-        images = sorted(glob(os.path.join(root, "images", "*.tif")))
-        masks = sorted(glob(os.path.join(root, "masks", "*.tif")))
-    assert len(images) == len(masks)
 
-    for im, mask in zip(images, masks):
-        print(im)
+def check_visually(images, labels):
+    for im, label in tqdm(zip(images, labels), total=len(images)):
 
         vol = imageio.imread(im)
-        seg = imageio.imread(mask).astype("uint32")
+        seg = imageio.imread(label).astype("uint32")
 
         v = napari.Viewer()
-        v.add_image(vol)
-        v.add_labels(seg)
+        v.add_image(vol, name="pv-channel")
+        v.add_labels(seg, name="annotations")
+        folder, name = os.path.split(im)
+        folder = os.path.basename(folder)
+        v.title = f"{folder}/{name}"
         napari.run()
 
 
-def check_labels():
-    masks = sorted(glob(os.path.join(root, "masks", "*.tif")))
-    for mask_path in masks:
-        labels = imageio.imread(mask_path)
+def check_labels(images, labels):
+    for label_path in labels:
+        labels = imageio.imread(label_path)
         n_labels = len(np.unique(labels))
-        print(mask_path, n_labels)
+        print(label_path, n_labels)
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--root", "-i", help="The root folder with the annotated training crops.",
+        default=ROOT_CP,
+    )
+    parser.add_argument("--check_labels", "-l", action="store_true")
+    args = parser.parse_args()
+    root = args.root
+
+    images, labels = get_image_and_label_paths(root)
+
+    check_visually(images, labels)
+    if args.check_labels:
+        check_labels(images, labels)
 
 
 if __name__ == "__main__":
-    check_visually(True)
-    # check_labels()
+    main()

scripts/training/train_distance_unet.py

@@ -1,54 +1,72 @@
+import argparse
 import os
+from datetime import datetime
 from glob import glob
 
 import torch_em
-
 from torch_em.model import UNet3d
 
-# DATA_ROOT = "/home/pape/Work/data/moser/lightsheet"
-DATA_ROOT = "/scratch-grete/usr/nimcpape/data/moser/lightsheet"
+ROOT_CLUSTER = "/scratch-grete/usr/nimcpape/data/moser/lightsheet"
+
+
+def get_image_and_label_paths(root):
+    exclude_names = ["annotations", "cp_masks"]
+    all_image_paths = sorted(glob(os.path.join(root, "**/**.tif"), recursive=True))
+    all_image_paths = [
+        path for path in all_image_paths if not any(exclude in path for exclude in exclude_names)
+    ]
+
+    image_paths, label_paths = [], []
+    label_extensions = ["_annotations.tif"]
+    for path in all_image_paths:
+        folder, fname = os.path.split(path)
+        fname = os.path.splitext(fname)[0]
+        label_path = None
+        for ext in label_extensions:
+            candidate_label_path = os.path.join(folder, f"{fname}{ext}")
+            if os.path.exists(candidate_label_path):
+                label_path = candidate_label_path
+                break
+
+        if label_path is None:
+            print("Did not find annotations for", path)
+            print("This image will not be used for training.")
+        else:
+            image_paths.append(path)
+            label_paths.append(label_path)
+
+    assert len(image_paths) == len(label_paths)
+    return image_paths, label_paths
 
 
-def get_paths(image_paths, label_paths, split, filter_empty):
+def select_paths(image_paths, label_paths, split, filter_empty):
     if filter_empty:
         image_paths = [imp for imp in image_paths if "empty" not in imp]
         label_paths = [imp for imp in label_paths if "empty" not in imp]
     assert len(image_paths) == len(label_paths)
 
     n_files = len(image_paths)
 
-    train_fraction = 0.8
-    val_fraction = 0.1
+    train_fraction = 0.85
 
     n_train = int(train_fraction * n_files)
-    n_val = int(val_fraction * n_files)
     if split == "train":
         image_paths = image_paths[:n_train]
         label_paths = label_paths[:n_train]
 
     elif split == "val":
-        image_paths = image_paths[n_train:(n_train + n_val)]
-        label_paths = label_paths[n_train:(n_train + n_val)]
+        image_paths = image_paths[n_train:]
+        label_paths = label_paths[n_train:]
 
     return image_paths, label_paths
 
 
-def get_loader(split, patch_shape, batch_size, filter_empty, train_on=["default"]):
-    image_paths, label_paths = [], []
-
-    if "default" in train_on:
-        all_image_paths = sorted(glob(os.path.join(DATA_ROOT, "images", "*.tif")))
-        all_label_paths = sorted(glob(os.path.join(DATA_ROOT, "masks", "*.tif")))
-        this_image_paths, this_label_paths = get_paths(all_image_paths, all_label_paths, split, filter_empty)
-        image_paths.extend(this_image_paths)
-        label_paths.extend(this_label_paths)
+def get_loader(root, split, patch_shape, batch_size, filter_empty):
+    image_paths, label_paths = get_image_and_label_paths(root)
+    this_image_paths, this_label_paths = select_paths(image_paths, label_paths, split, filter_empty)
 
-    if "downsampled" in train_on:
-        all_image_paths = sorted(glob(os.path.join(DATA_ROOT, "images_s2", "*.tif")))
-        all_label_paths = sorted(glob(os.path.join(DATA_ROOT, "masks_s2", "*.tif")))
-        this_image_paths, this_label_paths = get_paths(all_image_paths, all_label_paths, split, filter_empty)
-        image_paths.extend(this_image_paths)
-        label_paths.extend(this_label_paths)
+    assert len(this_image_paths) == len(this_label_paths)
+    assert len(this_image_paths) > 0
 
     label_transform = torch_em.transform.label.PerObjectDistanceTransform(
             distances=True, boundary_distances=True, foreground=True,
@@ -69,26 +87,40 @@ def get_loader(split, patch_shape, batch_size, filter_empty, train_on=["default"
     return loader
 
 
-def main(check_loaders=False):
-    # Parameters for training:
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--root", "-i", help="The root folder with the annotated training crops.",
+        default=ROOT_CLUSTER,
+    )
+    parser.add_argument(
+        "--check_loaders", "-l", action="store_true",
+        help="Visualize the data loader output instead of starting a training run."
+    )
+    parser.add_argument(
+        "--filter_empty", "-f", action="store_true",
+        help="Whether to exclude blocks with empty annotations from the training process."
+    )
+    parser.add_argument(
+        "--name", help="Optional name for the model to be trained. If not given the current date is used."
+    )
+    args = parser.parse_args()
+    root = args.root
+    check_loaders = args.check_loaders
+    filter_empty = args.filter_empty
+    run_name = datetime.now().strftime("%Y%m%d") if args.name is None else args.name
+
+    # Parameters for training on A100.
     n_iterations = 1e5
     batch_size = 8
-    filter_empty = False
-    train_on = ["downsampled"]
-    # train_on = ["downsampled", "default"]
-
-    patch_shape = (32, 128, 128) if "downsampled" in train_on else (64, 128, 128)
+    patch_shape = (64, 128, 128)
 
     # The U-Net.
     model = UNet3d(in_channels=1, out_channels=3, initial_features=32, final_activation="Sigmoid")
 
     # Create the training loader with train and val set.
-    train_loader = get_loader(
-        "train", patch_shape, batch_size, filter_empty=filter_empty, train_on=train_on
-    )
-    val_loader = get_loader(
-        "val", patch_shape, batch_size, filter_empty=filter_empty, train_on=train_on
-    )
+    train_loader = get_loader(root, "train", patch_shape, batch_size, filter_empty=filter_empty)
+    val_loader = get_loader(root, "val", patch_shape, batch_size, filter_empty=filter_empty)
 
     if check_loaders:
         from torch_em.util.debug import check_loader
@@ -99,12 +131,7 @@ def main(check_loaders=False):
     loss = torch_em.loss.distance_based.DiceBasedDistanceLoss(mask_distances_in_bg=True)
 
     # Create the trainer.
-    name = "cochlea_distance_unet"
-    if filter_empty:
-        name += "-filter-empty"
-    if train_on == ["downsampled"]:
-        name += "-train-downsampled"
-
+    name = f"cochlea_distance_unet_{run_name}"
     trainer = torch_em.default_segmentation_trainer(
         name=name,
         model=model,
@@ -123,4 +150,4 @@ def main(check_loaders=False):
 
 
 if __name__ == "__main__":
-    main(check_loaders=False)
+    main()