Data Transformers¶
The dnanet.data.transformer package provides callable transformers that
convert HIDImage or ExtractedPeak objects into PyTorch tensor inputs
and targets for different training tasks.
Base Class¶
Abstract base class for all transformers. Implements __call__ (image →
(inputs, target) tuple) and a collate_fn for batch assembly.
Segmentation Transformer¶
Produces (data, target) where both are (D, L) tensors. The target is a
scanpoint-indexed label mask.
Allele Metadata Transformer¶
Wraps another transformer and appends allele metadata (panel, scaler, signal, allele annotation) to each output. Enables allele-level metric computation during evaluation.
Output: (inputs, target, metadata_dict)
Combined Transformer¶
Transformer for the combined PeakNet model. Produces a tuple of:
(full_image, peak_windows, marker_idxs, peak_centers, n_peaks) and a
(D, L) target mask.
Constructor args:
- threshold — Minimum RFU for peak detection (default: 25)
- window_size — Peak extraction window in scan points (default: 120)
- include_max_pool_dyes — Include max-pooled other-dye channel
- autoencoder_log_scale — Apply log scaling to full image
- autoencoder_max_rfu — Max RFU for normalization
Reconstruction Transformer¶
For autoencoder reconstruction tasks. Input is preprocessed (log-scaled), target is raw data so the loss is computed in non-preprocessed space.
Constructor args:
- n_dyes — Number of dye channels to include (default: 5)
- autoencoder_log_scale — Apply log scaling (default: True)
- autoencoder_max_rfu — Max RFU for normalization
Peak Classification Transformer¶
For the PeakNet classifier. Converts ExtractedPeak objects into
(peak_tensor, marker_tensor) inputs and class-index targets.
Constructor args:
- include_marker — Include marker index tensor (default: True)