TheAiCollectiveART's picture
Initial release of Language U Microscopy submission framework
2fe488a verified
Raw
History Blame Contribute Delete
13.5 kB
"""I/O utilities for tracking challenge datasets."""
from dataclasses import dataclass, field
from pathlib import Path
import dask.array as da
import numpy as np
import polars as pl
import torch
import tracksdata as td
import zarr
DEFAULT_SCALE: tuple[float, float, float] = (1.625, 0.40625, 0.40625)
@dataclass
class Dataset:
path: Path
image: da.Array | np.ndarray | torch.Tensor | None
tracks: td.graph.IndexedRXGraph | None
scale: tuple[float, float, float]
original_scale: tuple[float, float, float] | None = None
image_shape: tuple[int, ...] | None = None # (T, Z, Y, X), always available
zarr_path: Path | None = None
quantiles: dict[str, float] = field(default_factory=dict)
def napari_tracks(self) -> tuple[np.ndarray, dict[int, int]]:
return td.functional.to_napari_format(
self.tracks,
self.image.shape,
solution_key=None,
output_tracklet_id_key="track_id",
mask_key=None,
)
def open_dataset(ds_path: Path | str,
target_scale: tuple[float, float, float] | None = None,
normalize: bool = True,
gamma: float = 1.0,
device: str = "cuda",
require_tracks: bool = False,
load_image: bool = True,
downsample: tuple[int, ...] | None = None) -> Dataset:
"""Open a dataset from a zarr file and optionally a geff tracks file.
Parameters
----------
ds_path : Path or str
Path to the dataset (without extension, or with .zarr/.geff).
target_scale : tuple, optional
Resample to this isotropic (Z, Y, X) voxel scale. If None, no resampling.
normalize : bool
Whether to normalize the image intensities.
gamma : float
Gamma correction value for normalization.
device : str
Device to use for GPU processing.
require_tracks : bool
Whether to require a .geff tracks file.
load_image : bool
If False, skip loading image data (metadata + tracks only).
downsample : tuple, optional
Spatial downsample strides (Z, Y, X). Applied via strided indexing.
Returns
-------
Dataset
Dataset with image shape (T, Z, Y, X). When ``load_image=False``,
``image`` is None but ``image_shape`` and ``zarr_path`` are populated.
"""
ds_path = Path(ds_path)
if ds_path.suffix in (".zarr", ".geff"):
ds_path = ds_path.parent / ds_path.stem
image_path = ds_path.parent / f"{ds_path.stem}.zarr"
tracks_path = ds_path.parent / f"{ds_path.stem}.geff"
if not image_path.exists():
raise FileNotFoundError(f"Image file not found: {image_path}")
if require_tracks and not tracks_path.exists():
raise FileNotFoundError(f"Tracks file not found: {tracks_path}")
img_ds = zarr.open_group(image_path, mode="r")
tracks = None
if require_tracks and tracks_path.exists():
result = td.graph.IndexedRXGraph.from_geff(tracks_path)
tracks = result[0] if isinstance(result, tuple) else result
attrs = dict(img_ds.attrs)
scale = _parse_scale(attrs)
quantiles = attrs.get("image_statistics", {}).get("quantiles", {})
raw_shape = tuple(img_ds["0"].shape)
if downsample is not None:
ds_shape = raw_shape[:1] + tuple(-(-s // d) for s, d in zip(raw_shape[1:], downsample))
else:
ds_shape = raw_shape
if not load_image:
return Dataset(
path=ds_path, image=None, tracks=tracks, scale=scale,
image_shape=ds_shape, zarr_path=image_path, quantiles=quantiles,
)
da_arr = da.from_zarr(img_ds["0"]).compute()
resample = target_scale is not None
original_scale = None
if resample or normalize:
if resample:
original_scale = scale
da_arr, tracks, scale = _process_on_gpu(
da_arr, tracks, scale, device,
resample=resample, target_scale=target_scale,
normalize=normalize, gamma=gamma,
precomputed_quantiles=quantiles,
)
assert da_arr.ndim == 4
return Dataset(
path=ds_path, image=da_arr, tracks=tracks, scale=scale,
original_scale=original_scale, image_shape=tuple(da_arr.shape),
zarr_path=image_path, quantiles=quantiles,
)
def _parse_scale(attrs: dict) -> tuple[float, float, float]:
"""Extract (Z, Y, X) voxel scale from OME-NGFF zarr attrs."""
if "multiscales" in attrs:
transform = attrs["multiscales"][0]["datasets"][0]["coordinateTransformations"][0]
if transform["type"] != "scale":
raise ValueError(f"Transform type is not 'scale': {transform}")
return tuple(transform["scale"][-3:])
return DEFAULT_SCALE
def _lookup_precomputed_quantile(
quantiles: dict[str, float] | None, q: float, tol: float = 1e-6,
) -> float | None:
"""Return the stored quantile value for *q* if present in *quantiles*."""
if not quantiles:
return None
for key, val in quantiles.items():
try:
if abs(float(key) - q) <= tol:
return float(val)
except (TypeError, ValueError):
continue
return None
def _process_on_gpu(
image: np.ndarray,
tracks: td.graph.IndexedRXGraph | None,
scale: tuple[float, float, float],
device: str,
resample: bool = False,
target_scale: tuple[float, float, float] | None = None,
normalize: bool = True,
gamma: float = 1.0,
q_min: float = 0.000,
q_max: float = 1.000,
subsample_factor: int = 50,
precomputed_quantiles: dict[str, float] | None = None,
) -> tuple[torch.Tensor, td.graph.IndexedRXGraph | None, tuple[float, float, float]]:
"""Process image on GPU: normalization and isotropic resampling.
All GPU processing is consolidated here to minimize CPU<->GPU transfers.
Expects 4D input: (T, Z, Y, X)
Processing order:
1. Normalize (quantile + gamma) - on CPU/numpy
2. Transfer to GPU
3. Resample to isotropic (if enabled)
4. Transfer back to CPU
"""
torch_device = torch.device(device)
# Convert to float32 for processing (zarr data is typically uint16)
image = image.astype(np.float32)
# 1. Quantile stats for normalization. Prefer precomputed values from zarr
# attrs (instant); fall back to a subsampled np.quantile if missing.
q1, q2 = None, None
if normalize:
q1 = _lookup_precomputed_quantile(precomputed_quantiles, q_min)
q2 = _lookup_precomputed_quantile(precomputed_quantiles, q_max)
if q1 is None or q2 is None:
flat = image.ravel()[::subsample_factor]
q1, q2 = np.quantile(flat, [q_min, q_max]).astype(np.float32)
else:
q1 = np.float32(q1)
q2 = np.float32(q2)
# 2. Transfer raw float32 to GPU once (pinned memory for faster DMA)
tensor = torch.from_numpy(image).pin_memory().to(torch_device, non_blocking=True)
# 3. Apply normalization on GPU
if normalize:
tensor = (tensor - float(q1)) / (float(q2) - float(q1) + 1e-6)
tensor = tensor.clamp(min=0.0)
if gamma != 1.0:
tensor = tensor.pow(gamma)
tensor = tensor.clamp(0.0, 4.0)
# 4. Resample to target scale (if enabled)
if resample:
scale_arr = np.array(scale)
if target_scale is None:
target_scale_val = scale_arr.min()
else:
target_scale_val = np.array(target_scale)
zoom_factors = scale_arr / target_scale_val
# For 4D input (T, Z, Y, X), reshape to (T, 1, Z, Y, X) for interpolate
T = tensor.shape[0]
new_spatial_shape = (np.array(tensor.shape[1:]) * zoom_factors).astype(int).tolist()
tensor = tensor[:, None] # (T, 1, Z, Y, X)
tensor = torch.nn.functional.interpolate(
tensor, size=new_spatial_shape, mode="trilinear", align_corners=False
)
tensor = tensor[:, 0] # (T, Z, Y, X)
# Update track coordinates if tracks are provided
if tracks is not None:
node_attrs = tracks.node_attrs()
orig_dtypes = {col: node_attrs.schema[col] for col in ["z", "y", "x"]}
node_attrs = node_attrs.with_columns(
(pl.col("z") * zoom_factors[0]).round(0).cast(orig_dtypes["z"]),
(pl.col("y") * zoom_factors[1]).round(0).cast(orig_dtypes["y"]),
(pl.col("x") * zoom_factors[2]).round(0).cast(orig_dtypes["x"]),
)
tracks.update_node_attrs(
attrs=node_attrs.select("z", "y", "x").to_dict(),
node_ids=node_attrs[td.DEFAULT_ATTR_KEYS.NODE_ID].to_list(),
)
# Update scale to isotropic
if target_scale is not None:
scale = target_scale
else:
scale = (float(target_scale_val),) * 3
return tensor, tracks, scale
def invert_time_graph(tracks: td.graph.IndexedRXGraph, max_t: int = 100) -> td.graph.IndexedRXGraph:
"""Invert the time axis of the image and tracks."""
# Update node time attributes
node_attrs = tracks.node_attrs()
node_attrs = node_attrs.with_columns(
((max_t - 1) - pl.col("t")).alias("t"),
)
tracks.update_node_attrs(
attrs=node_attrs.select("t").to_dict(),
node_ids=node_attrs[td.DEFAULT_ATTR_KEYS.NODE_ID].to_list(),
)
# Collect all edge info first before modifying the graph
edge_attrs = tracks.edge_attrs()
edges_to_reverse = []
for row in edge_attrs.iter_rows(named=True):
source = row[td.DEFAULT_ATTR_KEYS.EDGE_SOURCE]
target = row[td.DEFAULT_ATTR_KEYS.EDGE_TARGET]
edge_id = row[td.DEFAULT_ATTR_KEYS.EDGE_ID]
attrs = {k: v for k, v in row.items()
if k not in (td.DEFAULT_ATTR_KEYS.EDGE_SOURCE,
td.DEFAULT_ATTR_KEYS.EDGE_TARGET,
td.DEFAULT_ATTR_KEYS.EDGE_ID)}
edges_to_reverse.append((edge_id, source, target, attrs))
# Now modify the graph
for edge_id, source, target, attrs in edges_to_reverse:
tracks.remove_edge(edge_id=edge_id)
tracks.add_edge(source_id=target, target_id=source, attrs=attrs)
return tracks
def rescale_graph_to_original(
graph: td.graph.BaseGraph,
original_scale: tuple[float, float, float],
) -> td.graph.BaseGraph:
"""
Rescale graph coordinates from isotropic space back to original anisotropic space.
This inverts the isotropic resampling transformation applied during dataset loading.
Parameters
----------
graph : td.graph.BaseGraph
Graph with coordinates in isotropic space.
original_scale : tuple[float, float, float]
Original (z, y, x) voxel scale before isotropic resampling.
Returns
-------
td.graph.BaseGraph
Graph with coordinates scaled back to original space.
"""
scale_arr = np.array(original_scale)
target_scale = scale_arr.min()
zoom_factors = scale_arr / target_scale
# Invert the zoom by dividing coordinates
inverse_zoom = 1.0 / zoom_factors
node_attrs = graph.node_attrs()
node_attrs = node_attrs.with_columns(
(pl.col("z") * inverse_zoom[0]),
(pl.col("y") * inverse_zoom[1]),
(pl.col("x") * inverse_zoom[2]),
)
graph.update_node_attrs(
attrs=node_attrs.select("z", "y", "x").to_dict(),
node_ids=node_attrs[td.DEFAULT_ATTR_KEYS.NODE_ID].to_list(),
)
return graph
def save_graph(graph: td.graph.BaseGraph, output_path: Path | str, overwrite: bool = True) -> None:
"""
Save a tracksdata graph to a .geff file.
Parameters
----------
graph : td.graph.BaseGraph
The graph to save.
output_path : Path or str
Output path for the .geff file.
overwrite : bool
Whether to overwrite existing files.
"""
import shutil
output_path = Path(output_path)
output_path.parent.mkdir(parents=True, exist_ok=True)
# Ensure path has .geff extension
if not output_path.suffix == ".geff":
output_path = output_path.with_suffix(".geff")
# Remove existing file/directory if overwrite is enabled
if overwrite and output_path.exists():
if output_path.is_dir():
shutil.rmtree(output_path)
else:
output_path.unlink()
graph.to_geff(output_path)
def list_datasets(data_dir: Path | str, require_geff: bool = True) -> list[Path]:
"""
List all valid datasets in a directory.
A valid dataset has a .zarr file and optionally a corresponding .geff file.
Parameters
----------
data_dir : Path or str
Directory containing datasets.
require_geff : bool
If True, only return datasets with both .zarr and .geff files.
If False, return all .zarr files.
Returns
-------
list[Path]
List of dataset paths (without extension).
"""
data_dir = Path(data_dir)
if not data_dir.exists():
raise FileNotFoundError(f"Directory not found: {data_dir}")
zarr_files = sorted(data_dir.glob("*.zarr"))
datasets = []
for zarr_file in zarr_files:
if require_geff:
geff_file = data_dir / f"{zarr_file.stem}.geff"
if geff_file.exists():
datasets.append(data_dir / zarr_file.stem)
else:
datasets.append(data_dir / zarr_file.stem)
return datasets