Code for "A Generative Indexing Framework for Robust Particle Identification in CR-39 Detectors via Latent Diffusion and Retrieval"

This file defines the common network architecture (encoder and decoder) used exclusively by the autoencoder in the AE-diffusion pipeline (i.e., the three variants trained by FINAL_PROJECT_RANDOM_SEED_TRAIN_CHAOCANSHU).FINAL_PROJECT_BASE_AE_NET_V2：Network definition file for the Pre-Inpainting VAE. This file defines the network architecture used exclusively by the Pre-Inpainting VAE (trained by FINAL_PAPER_VAE_inpainting_train_size_v2).FINAL_PROJECT_DIFFUSION_train：Latent-space diffusion training script.

This script trains the Denoising Diffusion Implicit Model (DDIM) on the 64×1 latent codes. It requires the trained autoencoder from FINAL_PROJECT_RANDOM_SEED_TRAIN_CHAOCANSHU as the encoder/decoder and the latent codes from the encoding script as training data.FINAL_PAPER_VAE_inpainting_train_size_v2：Training script for the Pre-Inpainting VAE. This script trains a separate VAE with a composite loss (including reconstruction, size regression, and noise-injection terms) to perform coarse gap-filling on occluded images.

It utilizes the network defined in FINAL_PROJECT_BASE_AE_NET_V2.FINAL_PAPER_VAE_inpainting_v2_eval：Pre-Inpainting VAE inference script. This script applies the trained Pre-Inpainting VAE model (the output of FINAL_PAPER_VAE_inpainting_train_size_v2) to perform coarse inpainting on occluded particle images, filling missing regions to shorten the effective diffusion path.FINAL_PROJECT_ID_decoder：Training script for the Identity Decoder.

This script trains a small fully-connected network that maps latent codes to a probability distribution over database IDs. It is trained with cross-entropy loss using clean-image latent codes as input.FINAL_PAPER_latent_EVAL：Latent space evaluation script. This script evaluates the quality of the learned latent manifold and the retrieval performance on the test sets.FINAL_PAPER_REAL_MASK_DIFFUSION_ID_EVAL_CIRCLE：Complete pipeline evaluation script.

This script integrates the entire generative indexing pipeline—pre-inpainting, latent encoding, diffusion denoising, and identity decoding—to evaluate the end-to-end performance of the framework on the three test sets. It reproduces the mixed loss curves and etching time classification results reported in the paper. For actual use, please manually modify the model loading path variables inside the script to point to the corresponding trained model weight files./utils：Directory containing helper functions for data loading, preprocessing, and model component definitions.requirements.txt：File listing all required Python packages and their versions.Instructions:Ensure Python 3.8+ and PyTorch 1.12+ are installed.Install dependencies: pip install -r requirements.txtPlace the preprocessed dataset (separately deposited in Science Data Bank) in the ./data directory.Train the autoencoder: python FINAL_PROJECT_RANDOM_SEED_TRAIN_CHAOCANSHU --model stable_ae --seed 42Train the Pre-Inpainting VAE: python FINAL_PAPER_VAE_inpainting_train_size_v2Apply Pre-Inpainting VAE: python FINAL_PAPER_VAE_inpainting_v2_eval --model_path /path/to/pre_vae.pthTrain the diffusion model: python FINAL_PROJECT_DIFFUSION_train --ae_path /path/to/trained_ae.pth --latent_path /path/to/latent_codes.ptTrain the Identity Decoder: python FINAL_PROJECT_ID_decoder --latent_path /path/to/latent_codes.ptEvaluate the full pipeline: modify the model path variables in FINAL_PAPER_REAL_MASK_DIFFUSION_ID_EVAL_CIRCLE and run: python FINAL_PAPER_REAL_MASK_DIFFUSION_ID_EVAL_CIRCLENote:The dataset required by this code (raw CR-39 images, annotations, and preprocessed tensors) is deposited separately in Science Data Bank.

Please refer to the Data Availability Statement in the paper for access information.Disclaimer: All file paths in the code (including model loading paths, data storage paths, and output save paths) are provided as examples. Users are responsible for modifying the relevant path variables according to their actual software environments and file storage locations to ensure successful execution.