Colab demo : Inference

README.md

@@ -99,6 +99,8 @@ and launching the plugin from the Plugins menu.
 You may use the test volume in the `examples` folder to test the inference and review tools.
 This should run in far less than five minutes on a modern computer.
 
+You may also find a demo Colab notebook in the `notebooks` folder.
+
 ## Issues
 
 **Help us make the code better by reporting issues and adding your feature requests!**

napari_cellseg3d/dev_scripts/remote_inference.py

@@ -0,0 +1,188 @@
+"""Script to perform inference on a single image and run post-processing on the results, withot napari."""
+import logging
+from dataclasses import dataclass
+from pathlib import Path
+from typing import List
+
+import numpy as np
+import torch
+
+from napari_cellseg3d.code_models.instance_segmentation import (
+    clear_large_objects,
+    clear_small_objects,
+    threshold,
+    volume_stats,
+    voronoi_otsu,
+)
+from napari_cellseg3d.code_models.worker_inference import InferenceWorker
+from napari_cellseg3d.config import (
+    InferenceWorkerConfig,
+    InstanceSegConfig,
+    ModelInfo,
+    SlidingWindowConfig,
+)
+from napari_cellseg3d.utils import resize
+
+logger = logging.getLogger(__name__)
+logging.basicConfig(level=logging.INFO)
+
+
+class LogFixture:
+    """Fixture for napari-less logging, replaces napari_cellseg3d.interface.Log in model_workers.
+
+    This allows to redirect the output of the workers to stdout instead of a specialized widget.
+    """
+
+    def __init__(self):
+        """Creates a LogFixture object."""
+        super(LogFixture, self).__init__()
+
+    def print_and_log(self, text, printing=None):
+        """Prints and logs text."""
+        print(text)
+
+    def warn(self, warning):
+        """Logs warning."""
+        logger.warning(warning)
+
+    def error(self, e):
+        """Logs error."""
+        raise (e)
+
+
+WINDOW_SIZE = 64
+
+MODEL_INFO = ModelInfo(
+    name="SwinUNetR",
+    model_input_size=64,
+)
+
+CONFIG = InferenceWorkerConfig(
+    device="cuda" if torch.cuda.is_available() else "cpu",
+    model_info=MODEL_INFO,
+    results_path=str(Path("./results").absolute()),
+    compute_stats=False,
+    sliding_window_config=SlidingWindowConfig(WINDOW_SIZE, 0.25),
+)
+
+
+@dataclass
+class PostProcessConfig:
+    """Config for post-processing."""
+
+    threshold: float = 0.4
+    spot_sigma: float = 0.55
+    outline_sigma: float = 0.55
+    isotropic_spot_sigma: float = 0.2
+    isotropic_outline_sigma: float = 0.2
+    anisotropy_correction: List[
+        float
+    ] = None  # TODO change to actual values, should be a ratio like [1,1/5,1]
+    clear_small_size: int = 5
+    clear_large_objects: int = 500
+
+
+def inference_on_images(
+    image: np.array, config: InferenceWorkerConfig = CONFIG
+):
+    """This function provides inference on an image with minimal config.
+
+    Args:
+        image (np.array): Image to perform inference on.
+        config (InferenceWorkerConfig, optional): Config for InferenceWorker. Defaults to CONFIG, see above.
+    """
+    # instance_method = InstanceSegmentationWrapper(voronoi_otsu, {"spot_sigma": 0.7, "outline_sigma": 0.7})
+
+    config.post_process_config.zoom.enabled = False
+    config.post_process_config.thresholding.enabled = (
+        False  # will need to be enabled and set to 0.5 for the test images
+    )
+    config.post_process_config.instance = InstanceSegConfig(
+        enabled=False,
+    )
+
+    config.layer = image
+
+    log = LogFixture()
+    worker = InferenceWorker(config)
+    logger.debug(f"Worker config: {worker.config}")
+
+    worker.log_signal.connect(log.print_and_log)
+    worker.warn_signal.connect(log.warn)
+    worker.error_signal.connect(log.error)
+
+    worker.log_parameters()
+
+    results = []
+    # append the InferenceResult when yielded by worker to results
+    for result in worker.inference():
+        results.append(result)
+
+    return results
+
+
+def post_processing(semantic_segmentation, config: PostProcessConfig = None):
+    """Run post-processing on inference results."""
+    config = PostProcessConfig() if config is None else config
+    # if config.anisotropy_correction is None:
+    # config.anisotropy_correction = [1, 1, 1 / 5]
+    if config.anisotropy_correction is None:
+        config.anisotropy_correction = [1, 1, 1]
+
+    image = semantic_segmentation
+    # apply threshold to semantic segmentation
+    logger.info(f"Thresholding with {config.threshold}")
+    image = threshold(image, config.threshold)
+    logger.debug(f"Thresholded image shape: {image.shape}")
+    # remove artifacts by clearing large objects
+    logger.info(f"Clearing large objects with {config.clear_large_objects}")
+    image = clear_large_objects(image, config.clear_large_objects)
+    # run instance segmentation
+    logger.info(
+        f"Running instance segmentation with {config.spot_sigma} and {config.outline_sigma}"
+    )
+    labels = voronoi_otsu(
+        image,
+        spot_sigma=config.spot_sigma,
+        outline_sigma=config.outline_sigma,
+    )
+    # clear small objects
+    logger.info(f"Clearing small objects with {config.clear_small_size}")
+    labels = clear_small_objects(labels, config.clear_small_size).astype(
+        np.uint16
+    )
+    logger.debug(f"Labels shape: {labels.shape}")
+    # get volume stats WITH ANISOTROPY
+    logger.debug(f"NUMBER OF OBJECTS: {np.max(np.unique(labels))-1}")
+    stats_not_resized = volume_stats(labels)
+    ######## RUN WITH ANISOTROPY ########
+    result_dict = {}
+    result_dict["Not resized"] = {
+        "labels": labels,
+        "stats": stats_not_resized,
+    }
+
+    if config.anisotropy_correction != [1, 1, 1]:
+        logger.info("Resizing image to correct anisotropy")
+        image = resize(image, config.anisotropy_correction)
+        logger.debug(f"Resized image shape: {image.shape}")
+        logger.info("Running labels without anisotropy")
+        labels_resized = voronoi_otsu(
+            image,
+            spot_sigma=config.isotropic_spot_sigma,
+            outline_sigma=config.isotropic_outline_sigma,
+        )
+        logger.info(
+            f"Clearing small objects with {config.clear_large_objects}"
+        )
+        labels_resized = clear_small_objects(
+            labels_resized, config.clear_small_size
+        ).astype(np.uint16)
+        logger.debug(
+            f"NUMBER OF OBJECTS: {np.max(np.unique(labels_resized))-1}"
+        )
+        logger.info("Getting volume stats without anisotropy")
+        stats_resized = volume_stats(labels_resized)
+        return labels_resized, stats_resized
+
+    return labels, stats_not_resized