fiber inference

Author: Donglai Wei

CLAUDE.md

@@ -54,7 +54,7 @@ conda create -n pytc python=3.10 -y
 conda activate pytc
 
 # Install pre-built packages (avoids compilation)
-conda install -c conda-forge numpy h5py cython connected-components-3d -y
+conda install -c conda-forge numpy=1.23 h5py cython connected-components-3d mahotas -y
 
 # Install PyTorch (adjust for your CUDA version)
 pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121

QUICKSTART.md

@@ -47,6 +47,37 @@ pip install -e .
 
 ---
 
+### ❌ "AttributeError: module 'numpy' has no attribute 'float'"
+
+**Cause:** Mahotas version incompatibility with NumPy 2.0+. This occurs with older mahotas versions (< 1.4.18).
+
+**Solution 1 - Upgrade packages (recommended):**
+```bash
+pip install --upgrade numpy mahotas
+```
+
+**Solution 2 - Pin compatible versions:**
+```bash
+pip install numpy>=1.23.0 mahotas>=1.4.18
+```
+
+**Why?** Mahotas 1.4.18+ is compatible with NumPy 2.x. The deprecated `np.float` alias was removed in NumPy 2.0.
+
+---
+
+### ❌ "Matplotlib requires numpy>=1.23"
+
+**Cause:** Matplotlib requires NumPy 1.23 or higher for compatibility.
+
+**Solution:**
+```bash
+conda activate pytc
+conda install -c conda-forge matplotlib -y
+pip install -e . --no-build-isolation
+```
+
+---
+
 ### ❌ "Could not find a version that satisfies connected-components-3d"
 
 **Cause:** Python version incompatibility (cc3d requires Python 3.10).

TROUBLESHOOTING.md

@@ -518,7 +518,6 @@ class CheckpointConfig:
     save_top_k: int = 1
     save_last: bool = True
     save_every_n_epochs: int = 10
-    dirpath: str = "checkpoints/"
     checkpoint_filename: Optional[str] = None  # Auto-generated from monitor if None
     use_timestamp: bool = True  # Create timestamped subdirectories (YYYYMMDD_HHMMSS)
 
@@ -815,8 +814,14 @@ class AugmentationConfig:
 
 @dataclass
 class InferenceDataConfig:
-    """Inference data configuration."""
+    """Inference data configuration.
 
+    Output path is automatically computed from checkpoint directory:
+    - Checkpoint: outputs/experiment_name/YYYYMMDD_HHMMSS/checkpoints/last.ckpt
+    - Inference: outputs/experiment_name/YYYYMMDD_HHMMSS/inference/last.ckpt/{output_name}
+    """
+
+    test_path: str = ""  # Base path for test data (e.g., "/path/to/dataset/test/")
     test_image: Any = None  # str, List[str], or None - Can be single file or list of files
     test_label: Any = None  # str, List[str], or None - Can be single file or list of files
     test_mask: Any = None  # str, List[str], or None - Optional mask for inference
@@ -826,7 +831,9 @@ class InferenceDataConfig:
     test_transpose: List[int] = field(
         default_factory=list
     )  # Axis permutation for test data (e.g., [2,1,0] for xyz->zyx)
-    output_path: str = "results/"
+    output_name: str = (
+        "predictions.h5"  # Output filename (auto-pathed to inference/{checkpoint}/{output_name})
+    )
 
     # 2D data support
     do_2d: bool = False  # Enable 2D data processing for inference
@@ -838,8 +845,7 @@ class SlidingWindowConfig:
 
     window_size: Optional[List[int]] = None
     sw_batch_size: Optional[int] = None  # If None, will use system.inference.batch_size
-    overlap: Optional[float] = 0.5  # Overlap ratio (0-1), or None to use stride instead
-    stride: Optional[List[int]] = None  # Explicit stride (overrides overlap if set)
+    stride: Optional[List[int]] = None  # Explicit stride for controlling window movement
     blending: str = "gaussian"  # 'gaussian' or 'constant' - blending mode for overlapping patches
     sigma_scale: float = (
         0.125  # Gaussian sigma scale (only for blending='gaussian'); larger = smoother blending
@@ -860,7 +866,7 @@ class TestTimeAugmentationConfig:
         None  # Single activation for all channels: 'softmax', 'sigmoid', 'tanh', None (deprecated, use channel_activations)
     )
     channel_activations: Optional[List[Any]] = (
-        None  # Per-channel activations: [[0, 'sigmoid'], [1, 'sigmoid'], [2, 'tanh']]
+        None  # Per-channel activations: [[start_ch, end_ch, 'activation'], ...] e.g., [[0, 2, 'softmax'], [2, 3, 'sigmoid'], [3, 4, 'tanh']]
     )
     select_channel: Any = (
         None  # Channel selection: null (all), [1] (foreground), -1 (all) (applied even with null flip_axes)

connectomics/config/hydra_config.py

@@ -241,6 +241,12 @@ def resolve_data_paths(cfg: Config) -> Config:
     2. Expanding glob patterns to actual file lists
     3. Flattening nested lists from glob expansion
 
+    Supported paths:
+    - Training: cfg.data.train_path + cfg.data.train_image/train_label/train_mask
+    - Validation: cfg.data.val_path + cfg.data.val_image/val_label/val_mask
+    - Testing (legacy): cfg.data.test_path + cfg.data.test_image/test_label/test_mask
+    - Inference (primary): cfg.inference.data.test_path + cfg.inference.data.test_image/test_label/test_mask
+
     Args:
         cfg: Config object to resolve paths for
 
@@ -253,6 +259,12 @@ def resolve_data_paths(cfg: Config) -> Config:
         >>> resolve_data_paths(cfg)
         >>> print(cfg.data.train_image)
         ['/data/barcode/PT37/img1_raw.tif', '/data/barcode/PT37/img2_raw.tif', '/data/barcode/file.tif']
+
+        >>> cfg.inference.data.test_path = "/data/test/"
+        >>> cfg.inference.data.test_image = ["volume_*.tif"]
+        >>> resolve_data_paths(cfg)
+        >>> print(cfg.inference.data.test_image)
+        ['/data/test/volume_1.tif', '/data/test/volume_2.tif']
     """
     import os
     from glob import glob
@@ -304,18 +316,18 @@ def _combine_path(base_path: str, file_path: Optional[Union[str, List[str]]]) ->
         cfg.data.val_mask = _combine_path(cfg.data.val_path, cfg.data.val_mask)
         cfg.data.val_json = _combine_path(cfg.data.val_path, cfg.data.val_json)
 
-    # Resolve test paths
+    # Resolve test paths (legacy support for cfg.data.test_path)
     if cfg.data.test_path:
         cfg.data.test_image = _combine_path(cfg.data.test_path, cfg.data.test_image)
         cfg.data.test_label = _combine_path(cfg.data.test_path, cfg.data.test_label)
         cfg.data.test_mask = _combine_path(cfg.data.test_path, cfg.data.test_mask)
         cfg.data.test_json = _combine_path(cfg.data.test_path, cfg.data.test_json)
 
-    # Also resolve inference data paths
-    if cfg.data.test_path and cfg.inference.data:
-        cfg.inference.data.test_image = _combine_path(cfg.data.test_path, cfg.inference.data.test_image)
-        cfg.inference.data.test_label = _combine_path(cfg.data.test_path, cfg.inference.data.test_label)
-        cfg.inference.data.test_mask = _combine_path(cfg.data.test_path, cfg.inference.data.test_mask)
+    # Resolve inference data paths (primary location for test_path)
+    if hasattr(cfg.inference.data, 'test_path') and cfg.inference.data.test_path:
+        cfg.inference.data.test_image = _combine_path(cfg.inference.data.test_path, cfg.inference.data.test_image)
+        cfg.inference.data.test_label = _combine_path(cfg.inference.data.test_path, cfg.inference.data.test_label)
+        cfg.inference.data.test_mask = _combine_path(cfg.inference.data.test_path, cfg.inference.data.test_mask)
 
     return cfg
 

connectomics/config/hydra_utils.py

@@ -139,7 +139,8 @@ def watershed_split(
     mask = np.zeros(distance.shape, dtype=bool)
     mask[tuple(coords.T)] = True
     markers = cc3d.connected_components(mask)
-    split_objects = mahotas.cwatershed(-distance, markers, mask=cropped)
+    split_objects = mahotas.cwatershed(-distance, markers)
+    split_objects[~cropped] = 0  # Apply mask manually (mahotas 1.4.18 doesn't support mask parameter)
 
     seg_id = np.unique(split_objects)
     new_id = []