feat: add possibility to simulate images in batches

src/microsim/schema/detectors/_camera.py

@@ -117,15 +117,15 @@ def simulate(
         # dark current
         avg_dark_e = self.dark_current * exposure_s + self.clock_induced_charge
         if not isinstance(avg_dark_e, float):
-            new_shape = avg_dark_e.shape + (1,) * (detected_photons.ndim - 1)
+            new_shape = (1,) + avg_dark_e.shape + (1,) * (detected_photons.ndim - 2)
             avg_dark_e = np.asarray(avg_dark_e).reshape(new_shape)  # type: ignore [assignment]
         thermal_electrons = xp.poisson_rvs(avg_dark_e, shape=detected_photons.shape)
         total_electrons = detected_photons + thermal_electrons
 
         # cap total electrons to full-well-capacity
         total_electrons = xp.minimum(total_electrons, self.full_well)
 
-        if binning > 1:
+        if binning > 1:  # TODO: this function might not work with batch dim
             total_electrons = self.apply_pre_quantization_binning(
                 total_electrons, binning
             )
@@ -141,7 +141,7 @@ def simulate(
         gray_values = self.quantize_electrons(total_electrons, xp)
 
         # sCMOS binning
-        if binning > 1:
+        if binning > 1:  # TODO: this function might not work with batch dim
             gray_values = self.apply_post_quantization_binning(gray_values, binning)
 
         # ADC saturation

src/microsim/schema/dimensions.py

@@ -41,6 +41,7 @@ class Axis(str, Enum):
     T = "t"  # Time
     F = "f"  # Fluorophore
     W = "w"  # Wavelength
+    S = "s"  # Sample
 
     def __repr__(self) -> str:
         return f"<Axis.{self.name}>"

src/microsim/schema/modality/_simple_psf.py

@@ -42,13 +42,13 @@ def psf(
 
     def render(
         self,
-        truth: xrDataArray,  # (F, Z, Y, X)
+        truth: xrDataArray,  # (S, F, Z, Y, X)
         em_rates: xrDataArray,  # (C, F, W)
         objective_lens: ObjectiveLens,
         settings: Settings,
         xp: NumpyAPI,
     ) -> xrDataArray:
-        """Render a 3D image of the truth for F fluorophores, in C channels."""
+        """Render a batch of 3D images of truth for F fluorophores, in C channels."""
         # for every channel in the emission rates...
         channels = []
         for ch in em_rates.coords[Axis.C].values:
@@ -81,12 +81,13 @@ def render(
             # stack the fluorophores together to create the channel
             channels.append(xp.stack(fluors, axis=0))
 
-        return DataArray(
+        channels = DataArray(
             channels,
-            dims=[Axis.C, Axis.F, Axis.Z, Axis.Y, Axis.X],
+            dims=[Axis.C, Axis.F, Axis.S, Axis.Z, Axis.Y, Axis.X],
             coords={
                 Axis.C: em_rates.coords[Axis.C],
                 Axis.F: truth.coords[Axis.F],
+                Axis.S: truth.coords[Axis.S],
                 Axis.Z: truth.coords[Axis.Z],
                 Axis.Y: truth.coords[Axis.Y],
                 Axis.X: truth.coords[Axis.X],
@@ -95,8 +96,10 @@ def render(
                 "space": truth.attrs["space"],
                 "objective": objective_lens,
                 "units": "photons",
+                "long_name": "Optical Image",
             },
         )
+        return channels.transpose(Axis.S, ...)  # put batch dim first
 
     def _summed_weighted_psf(
         self,
@@ -152,6 +155,8 @@ def _summed_weighted_psf(
                 xp=xp,
             )
             summed_psf += psf * weight
+
+        summed_psf = summed_psf[None, ...]  # add batch dim
         return summed_psf  # type: ignore [no-any-return]
 
 
@@ -200,7 +205,7 @@ class Identity(_PSFModality):
 
     def render(
         self,
-        truth: xrDataArray,  # (F, Z, Y, X)
+        truth: xrDataArray,  # (S, F, Z, Y, X)
         em_rates: xrDataArray,  # (C, F, W)
         *args: Any,
         **kwargs: Any,
@@ -211,23 +216,14 @@ def render(
         already convolved with the PSF. Therefore, we simply compute the emission flux
         for each fluorophore and each channel.
         """
-        em_image = em_rates.sum(Axis.W) * truth
-        return DataArray(
-            em_image,
-            dims=[Axis.C, Axis.F, Axis.Z, Axis.Y, Axis.X],
-            coords={
-                Axis.C: em_rates.coords[Axis.C],
-                Axis.F: truth.coords[Axis.F],
-                Axis.Z: truth.coords[Axis.Z],
-                Axis.Y: truth.coords[Axis.Y],
-                Axis.X: truth.coords[Axis.X],
-            },
-            attrs={
-                "space": truth.attrs["space"],
-                "objective": "",
-                "units": "photons",
-            },
+        em_image = (em_rates.sum(Axis.W) * truth).transpose(Axis.S, ...)
+        em_image.attrs.update(
+            units="photons",
+            objective="",
+            space=truth.attrs["space"],
+            long_name="Optical Image",
         )
+        return em_image
 
 
 def bin_spectrum(

src/microsim/schema/sample/sample.py

@@ -53,6 +53,8 @@ def cache_path(self) -> tuple[str, ...] | None:
 
     def render(self, space: xrDataArray, xp: NumpyAPI | None = None) -> xrDataArray:
         """Render the fluorophore distribution into the given space."""
+        # This would need to change, as with additional batch dim there'd be a mismatch
+        # between distribution and space dims
         dist = self.distribution.render(space, xp)
         if isinstance(self.concentration, float | int):
             return dist * self.concentration

src/microsim/schema/simulation.py

@@ -6,7 +6,7 @@
 import numpy as np
 import pandas as pd
 import xarray as xr
-from pydantic import AfterValidator, Field, model_validator
+from pydantic import AfterValidator, Field, field_validator, model_validator
 
 from microsim._data_array import ArrayProtocol, from_cache, to_cache
 from microsim.util import microsim_cache
@@ -53,7 +53,7 @@ class Simulation(SimBaseModel):
 
     truth_space: Space
     output_space: Space | None = None
-    sample: Sample
+    samples: Sample | list[Sample]
     modality: Modality = Field(default_factory=Widefield)
     objective_lens: ObjectiveLens = Field(default_factory=ObjectiveLens)
     channels: list[OpticalConfig] = Field(default_factory=lambda: [FITC])
@@ -93,6 +93,28 @@ def _resolve_spaces(self) -> "Self":
             self.output_space.reference = self.truth_space
         return self
 
+    @model_validator(mode="after")
+    def _check_fluorophores_equal_in_samples(self) -> "Self":
+        fp_names = [{lbl.fluorophore.name for lbl in s.labels} for s in self.samples]
+        if len({frozenset(s) for s in fp_names}) != 1:
+            raise ValueError(
+                "All samples in the batch must use the same set of fluorophores."
+            )
+        return self
+
+    @field_validator("samples")
+    def _samples_to_list(value: Sample | list[Sample]) -> list[Sample]:
+        return [value] if isinstance(value, Sample) else value
+
+    @property
+    def sample(self) -> Sample:
+        warnings.warn(
+            "The `sample` attribute is deprecated. Use `samples` instead.",
+            DeprecationWarning,
+            stacklevel=2,
+        )
+        return self.samples[0]
+
     @property
     def _xp(self) -> "NumpyAPI":
         return self.settings.backend_module()
@@ -101,7 +123,7 @@ def ground_truth(self) -> xr.DataArray:
         """Return the ground truth data.
 
         Returns position and quantity of fluorophores in the sample.  The return array
-        has dimensions (F, Z, Y, X).  The units are fluorophores counts.
+        has dimensions (S, F, Z, Y, X). The units are fluorophores counts.
 
         Examples
         --------
@@ -112,36 +134,51 @@ def ground_truth(self) -> xr.DataArray:
         """
         if not hasattr(self, "_ground_truth"):
             xp = self._xp
-            # make empty space into which we'll add the ground truth
-            # TODO: this is wasteful... label.render should probably
-            # accept the space object directly
-            truth = self.truth_space.create(array_creator=xp.zeros)
-
-            # render each ground truth
-            label_data = []
-            for label in self.sample.labels:
-                cache_path = self._truth_cache_path(
-                    label, self.truth_space, self.settings.random_seed
-                )
-                if self.settings.cache.read and cache_path and cache_path.exists():
-                    data = from_cache(cache_path, xp=xp).astype(
-                        self.settings.float_dtype
-                    )
-                    logging.info(
-                        f"Loaded ground truth for {label} from cache: {cache_path}"
+
+            # render each ground sample
+            # TODO: iterating over sample is a bottleneck for batched approach... Ideas?
+            truths: list[xr.DataArray] = []
+            for sample in self.samples:
+                # render each label in the sample
+
+                # make empty space into which we'll add the ground truth
+                # TODO: this is wasteful... label.render should probably
+                # accept the space object directly
+                truth = self.truth_space.create(array_creator=xp.zeros)
+
+                label_data = []
+                for label in sample.labels:
+                    # TODO: differentiate caching for each label
+                    cache_path = self._truth_cache_path(
+                        label, self.truth_space, self.settings.random_seed
                     )
-                else:
-                    data = label.render(truth, xp=xp)
-                    if self.settings.cache.write and cache_path:
-                        to_cache(data, cache_path, dtype=np.uint16)
-
-                label_data.append(data)
-
-            # concat along the F axis
-            fluors = [lbl.fluorophore for lbl in self.sample.labels]
-            truth = xr.concat(label_data, dim=pd.Index(fluors, name=Axis.F))
-            truth.attrs.update(units="fluorophores", long_name="Ground Truth")
-            self._ground_truth = truth
+                    if self.settings.cache.read and cache_path and cache_path.exists():
+                        data = from_cache(cache_path, xp=xp).astype(
+                            self.settings.float_dtype
+                        )
+                        logging.info(
+                            f"Loaded ground truth for {label} from cache: {cache_path}"
+                        )
+                    else:
+                        data = label.render(truth, xp=xp)
+                        if self.settings.cache.write and cache_path:
+                            to_cache(data, cache_path, dtype=np.uint16)
+
+                    label_data.append(data)
+
+                # concat along the F axis
+                fluors = [lbl.fluorophore for lbl in sample.labels]
+                truth = xr.concat(label_data, dim=pd.Index(fluors, name=Axis.F))
+                truths.append(truth)
+
+            # concat along B axis
+            self._ground_truth = xr.concat(
+                truths, dim=pd.Index(range(len(truths)), name=Axis.S)
+            )  # TODO: is there a better way to give coords to this axis?
+            self._ground_truth.attrs.update(
+                units="fluorophores", long_name="Ground Truth"
+            )
+
         return self._ground_truth
 
     def filtered_emission_rates(self) -> xr.DataArray:
@@ -183,7 +220,7 @@ def emission_flux(self) -> xr.DataArray:
 
         This multiplies the per-fluorophore emission rates by the ground truth data to
         get the total emission flux for each voxel in the ground truth. The return
-        array has dimensions (C, F, Z, Y, X).  The units are photons/s.
+        array has dimensions (S, C, F, Z, Y, X).  The units are photons/s.
 
         Note, this integrates over all wavelengths. For finer control over the emission
         spectrum, you may wish to directly combine `filtered_emission_rates` with the
@@ -202,10 +239,12 @@ def emission_flux(self) -> xr.DataArray:
             return truth
 
         # total photons/s emitted by each fluorophore in each channel
-        total_flux = self.filtered_emission_rates().sum(Axis.W) * truth
+        total_flux = (self.filtered_emission_rates().sum(Axis.W) * truth).transpose(
+            Axis.S, ...
+        )
         total_flux.attrs.update(units="photon/sec", long_name="Emission Flux")
 
-        # (C, F, Z, Y, X)
+        # (S, C, F, Z, Y, X)
         return total_flux
 
     def optical_image_per_fluor(self) -> xr.DataArray:
@@ -214,11 +253,11 @@ def optical_image_per_fluor(self) -> xr.DataArray:
         This is the emission from each fluorophore in each channel, after filtering by
         the optical configuration and convolution with the PSF.
 
-        The return array has dimensions (C, F, Z, Y, X).  The units are photons/s.
+        The return array has dimensions (S, C, F, Z, Y, X).  The units are photons/s.
         """
-        # (C, F, Z, Y, X)
+        # (S, C, F, Z, Y, X)
         return self.modality.render(
-            self.ground_truth(),  # (F, Z, Y, X)
+            self.ground_truth(),  # (S, F, Z, Y, X)
             self.filtered_emission_rates(),  # (C, F, W)
             objective_lens=self.objective_lens,
             settings=self.settings,
@@ -230,10 +269,10 @@ def optical_image(self) -> xr.DataArray:
 
         This is the same as `optical_image_per_fluor`, but sums the contributions of all
         fluorophores in each channel (which a detector would not know). The return
-        array has dimensions (C, Z, Y, X).  The units are photons/s.
+        array has dimensions (S, C, Z, Y, X).  The units are photons/s.
         """
         oipf = self.optical_image_per_fluor()
-        return oipf.sum(Axis.F)  # (C, Z, Y, X)
+        return oipf.sum(Axis.F)  # (S, C, Z, Y, X)
 
     def digital_image(
         self,
@@ -245,13 +284,16 @@ def digital_image(
         """Return the digital image as captured by the detector.
 
         This down-scales the optical image to the output space, and simulates the
-        detector response.  The return array has dimensions (C, Z, Y, X).  The units
-        are gray values, based on the bit-depth of the detector.  If there is no
+        detector response.  The return array has dimensions (S, C, [F], Z, Y, X).  The
+        units are gray values, based on the bit-depth of the detector.  If there is no
         detector or `with_detector_noise` is False, the units are simply photons.
+
+        NOTE: the input `optical_image` can only contain a fluorophore dimension (e.g.,
+        it can be the output of `optical_image_per_fluor()`).
         """
         if optical_image is None:
             optical_image = self.optical_image()
-        image = optical_image  # (C, Z, Y, X)
+        image = optical_image  # (S, C, [F], Z, Y, X)
 
         # downscale to output space
         # TODO: consider how we would integrate detector pixel size
@@ -280,7 +322,7 @@ def digital_image(
             image = image * (ch_exposures / 1000)
             image.attrs.update(units="photons")
 
-        # (C, Z, Y, X)
+        # (S, C, Z, Y, X)
         return image
 
     def run(self) -> xr.DataArray: