refactor: move dataloader creation out of analysis function (ensure consistency due to rand perm in prepare_dataset)

Author: fabioseel

retinal_rl/analysis/channel_analysis.py

@@ -41,17 +41,13 @@ class HistogramAnalysis:
 
 def spectral_analysis(
     device: torch.device,
-    imageset: Imageset,
+    dataloader: DataLoader[tuple[Tensor, Tensor, int]],
     brain: Brain,
-    max_sample_size: int = 0,
 ) -> dict[str, SpectralAnalysis]:
     brain.eval()
     brain.to(device)
     _, cnn_layers = get_cnn_circuit(brain)
 
-    # Prepare dataset
-    dataloader = _prepare_dataset(imageset, max_sample_size)
-
     # Initialize results
     results: dict[str, SpectralAnalysis] = {}
 
@@ -73,17 +69,13 @@ def spectral_analysis(
 
 def histogram_analysis(
     device: torch.device,
-    imageset: Imageset,
+    dataloader: DataLoader[tuple[Tensor, Tensor, int]],
     brain: Brain,
-    max_sample_size: int = 0,
 ) -> dict[str, HistogramAnalysis]:
     brain.eval()
     brain.to(device)
     _, cnn_layers = get_cnn_circuit(brain)
 
-    # Prepare dataset
-    dataloader = _prepare_dataset(imageset, max_sample_size)  # TODO: Move outside?
-
     # Initialize results
     results: dict[str, HistogramAnalysis] = {}
 
@@ -101,7 +93,7 @@ def histogram_analysis(
     return results
 
 
-def _prepare_dataset(
+def prepare_dataset(
     imageset: Imageset, max_sample_size: int = 0
 ) -> DataLoader[tuple[Tensor, Tensor, int]]:
     """Prepare dataset and dataloader for analysis."""
@@ -394,9 +386,8 @@ def _plot_receptive_fields(ax: Axes, rf: FloatArray):
 
 
 def analyze_input(
-    device: torch.device, imageset: Imageset, max_sample_size: int
+    device: torch.device, dataloader: DataLoader[tuple[Tensor, Tensor, int]]
 ) -> tuple[SpectralAnalysis, HistogramAnalysis]:
-    dataloader = _prepare_dataset(imageset, max_sample_size)
     spectral_result = _layer_spectral_analysis(device, dataloader, nn.Identity())
     histogram_result = _layer_pixel_histograms(device, dataloader, nn.Identity())
     return spectral_result, histogram_result

runner/frameworks/classification/analyze.py

@@ -68,12 +68,10 @@ def analyze(
     log.save_dict(cfg.analyses_dir / f"receptive_fields_epoch_{epoch}.json", rf_result)
 
     if cfg.channel_analysis:
-        spectral_result = channel_ana.spectral_analysis(
-            device, test_set, brain, cfg.plot_sample_size
-        )
-        histogram_result = channel_ana.histogram_analysis(
-            device, test_set, brain, cfg.plot_sample_size
-        )
+        # Prepare dataset
+        dataloader = channel_ana.prepare_dataset(test_set, cfg.plot_sample_size)
+        spectral_result = channel_ana.spectral_analysis(device, dataloader, brain)
+        histogram_result = channel_ana.histogram_analysis(device, dataloader, brain)
         channel_ana.plot(
             log,
             rf_result,
@@ -142,8 +140,9 @@ def _extended_initialization_plots(
     if channel_analysis:
         # Input 'rfs' is just the colors
         rf_result = np.eye(input_shape[0])[:, :, np.newaxis, np.newaxis]
+        dataloader = channel_ana.prepare_dataset(train_set, max_sample_size)
         spectral_result, histogram_result = channel_ana.analyze_input(
-            device, train_set, max_sample_size
+            device, dataloader
         )
         channel_ana.input_plot(
             log,