Minor updates to Zarr access patterns. (#50)

* Minor updates to Zarr access patterns.

* Zarr3 doesn't support structured dtype access - convert
  to numpy array first
* Zarr3 doesn't have create_dataset, replace with create_array.

* DEP: Add min version pin for zarr.

Author: rossbar

requirements.txt

@@ -5,7 +5,7 @@ scikit-image
 torch
 torchvision
 tqdm
-zarr
+zarr>2
 matplotlib
 seaborn
 networkx

torch_mesmer/eval.py

@@ -138,7 +138,9 @@ def main(device: str,
 
     X_test = z_test['X'][:]
     y_test = z_test['y'][:].astype(int)
-    mpps = z_test['meta']['pixel_size'][:]
+    # NOTE: Zarr doesn't support structured arrays - must be converted to numpy
+    # array explicitly before attempting to access fields
+    mpps = z_test['meta'][:]["pixel_size"]
 
     # Load model and application
     model = Mesmer(

torch_mesmer/preprocess.py

@@ -45,19 +45,17 @@ def convert_to_zarr(filename, out_dir=None):
     print(f"    Writing {split}.")
 
     # Store 'X' — chunked across C, H, W (one sample per chunk)
-    store.create_dataset(
+    store.create_array(
         "X",
         data=X,
         chunks=(1, C, H, W),  # chunk = one full image (all channels, full spatial dims)
-        dtype=X.dtype,
     )
 
     # Store 'y' — chunked across C, H, W (one sample per chunk)
-    store.create_dataset(
+    store.create_array(
         "y",
         data=y,
         chunks=(1, C, H, W),
-        dtype=y.dtype,
     )
 
     meta_dtype = np.dtype(
@@ -71,7 +69,7 @@ def convert_to_zarr(filename, out_dir=None):
     meta_ary["specimen"] = data["meta"][crop_val:, -1]
 
     # Store 'metadata' — no spatial chunking needed, one scalar per sample
-    store.create_dataset(
+    store.create_array(
         "meta",
         data=meta_ary,
     )