Feature/compound bends

docs/gui.md

@@ -64,3 +64,64 @@ In the manuscript, refer to the interface as a *local graphical user interface*
 The compound latent-space tab is intentionally inference-only. The latent dimension is fixed at the trained 2-D embedding and the batch size is fixed internally; the GUI does not expose training-from-zero controls.
 
 The compound workflow tab now shows a compact legacy-style reach diagnostic inspired by the research scripts: reach-scale CWT energy, the integrated corridor-energy signal, curvature, and the centreline with unit boundaries. The separate model CWT image remains the 64 x 64 grayscale image used as autoencoder input.
+## GUI polish notes for paper/Zenodo release
+
+The GUI separates display diagnostics from model inputs:
+
+- The reach-scale CWT diagnostic is a legacy-style visual check for compound-unit segmentation. The frequency display range can be widened from the GUI without changing the segmentation calculation.
+- The selected-unit model CWT image is the exact 64 x 64 array passed to the trained compound autoencoder. An enhanced preview is shown beside it only for human visibility.
+- The compound latent-space plot uses a simple legend: reference meander cloud, detected meander units, and selected unit. It does not use `n_lobes` as the colour scale by default.
+- The single-bend classifier shows a visible latent-space cluster plot and cluster-size table when the optional single-bend autoencoder is available locally.
+
+
+
+## GUI display notes
+
+The compound latent-space panel can optionally annotate detected units with their unit numbers. The reference cloud is shown only as context, while the selected detected unit is highlighted separately.
+
+The compound CWT image shown in the workflow tab has two display modes. The enhanced preview is intended for visual inspection. The exact 64 x 64 image is the reproducible model input passed to the compound autoencoder; it uses pixel axes because it is a resized array rather than a physical CWT diagnostic figure.
+
+The reach-scale CWT diagnostic is a display-only diagnostic. The maximum displayed frequency can be adjusted in the GUI without changing the underlying compound segmentation.
+
+
+## Compound model CWT image convention
+
+The compound-bend model-input image is different from the reach-scale CWT diagnostic. The reach-scale CWT is a physical/diagnostic plot used to inspect segmentation. The compound model image is a resized 64 x 64 training-style array generated from one extracted unit's curvature signal. It follows the legacy training-image polarity: white background with dark high-energy CWT structures. The GUI enhanced preview is for visual inspection only; the exact 64 x 64 array is the image passed to the compound autoencoder.
+
+## Single-bend Zenodo model download
+
+The single-bend classifier tab can download the public Zenodo autoencoder directly into `models/Autoencoder_Meander_Bend.h5`. This file remains ignored by Git and should not be committed. The GUI verifies the downloaded file with the MD5 checksum stored in `meander_morphology.model`.
+
+
+## Compound model-image axes
+
+The compound model-image preview uses the same 64 x 64 array passed to the autoencoder. The horizontal axis is displayed as `S_bend/S_bend,max` to match the training plots. The vertical axis is a CWT scale index because the resized model input does not preserve the original physical frequency or period coordinate.
+
+
+## Compound CWT preview
+
+The compound workflow displays a training-style CWT preview by default. This preview uses the same convention as the research scripts:
+
+- x-axis: `S_bend / S_bend,max`
+- y-axis: `l = 1 / frequency`
+- plotting style: `contourf(..., cmap="binary", vmin=mean + std)`
+
+The exact 64 x 64 model input is not displayed in the GUI; export `compound_spectra.npy` for reproducibility checks.
+
+## Exact compound autoencoder input
+
+The compound autoencoder input is generated by rasterising the same axis-free
+Matplotlib `contourf(..., cmap="binary", vmin=mean+std)` image used in the
+training scripts and resizing that raster to 64 x 64. The physical/legacy axes
+are shown only in the training-style preview. The hidden exact model-input view
+is axis-free because the autoencoder was trained on images, not on axes.
+
+
+## Compound model-input display policy
+
+The GUI does not display the exact 64 x 64 compound autoencoder input by default.
+That array is a resized/rasterised model artifact and is easy to misinterpret as a
+scientific CWT figure. The GUI instead displays the training-style CWT preview
+with labelled axes (`S_bend / S_bend,max` and `l = 1 / frequency`). The internal
+model arrays can still be exported through `compound_spectra.npy` for
+reproducibility.

src/meander_morphology/compound_pipeline.py

@@ -6,7 +6,7 @@
 import pandas as pd
 
 from .compound import compound_bends_to_metadata_rows, extract_compound_bends
-from .cwt import save_spectrum_image, spectrum_image_from_geometry
+from .cwt import legacy_compound_training_image_from_curvature, save_spectrum_image
 from .io import read_centerline_table, write_bend_summary
 
 
@@ -50,12 +50,7 @@ def extract_compound_bends_and_spectra(
 
     spectra = []
     for unit in units:
-        image = spectrum_image_from_geometry(
-            unit.x,
-            unit.y,
-            image_size=image_size,
-            target_points=unit_points,
-        )
+        image = legacy_compound_training_image_from_curvature(unit.curvature, image_size=image_size)
         spectra.append(image)
         save_spectrum_image(str(spectra_dir / f"compound_unit_{unit.unit_id:04d}.png"), image)