electricalgorithm · electricalgorithm · Nov 22, 2025 · Nov 22, 2025 · Dec 11, 2025 · Dec 12, 2025
diff --git a/.gitignore b/.gitignore
@@ -1,9 +1,13 @@
 .venv
 __pycache__
+**/*.pth
+**/*.parquet
+**/*.png
 checkpoints/
 logs/
 .DS_Store
 .env
 .pytest_cache/
 .ruff_cache/
-.cursor/
+.cursor/
+**/*egg-info*
diff --git a/README.md b/README.md
@@ -1,10 +1,79 @@
-# HOLO-PASWIN: In-Line Holographical Physics-Aware SWIN Transformer
+# HoloPASWIN v2
 
-A research-grade deep learning project for eliminating the **twin-image problem in in-line holography** using a **physics-aware Swin-UNet architecture** trained with synthetic holograms generated via the Angular Spectrum Method.
+Physics-Aware Swin Transformer for eliminating twin-image artifacts in in-line holography.
 
-Build a modern, high-performing, physics-aware neural architecture capable of reconstructing complex object fields ( $\hat{O} = \hat{O}_{real} + i\hat{O}_{imag}$ ) from inline holograms while robustly suppressing **twin-image artifacts**, **noise**, and **aberrations**.
+```
+┌─────────────────────────────────────────────────────────────┐
+│                      HOLOPASWIN MODEL                        │
+└─────────────────────────────────────────────────────────────┘
+
+Input: Hologram (B,1,H,W)
+   │
+   ▼
+┌─────────────────┐
+│  Physics (ASM)  │  Back-propagation to object plane
+│  FFT→H*→IFFT    │  Output: Dirty field (with twin image)
+└────────┬────────┘
+         │
+         │ Complex → Real/Imag (B,2,H,W)
+         │
+         ▼
+┌─────────────────┐
+│  Swin Encoder   │  Multi-scale feature extraction
+│  4 stages       │  Scales: 1/4, 1/8, 1/16, 1/32
+│  Channels:      │  Channels: 96, 192, 384, 768
+│  96→192→384→768 │
+└────┬──┬──┬──┬───┘
+     │  │  │  │
+     │  │  │  └──────────────────────┐
+     │  │  └──────────────┐          │
+     │  └──────┐          │          │
+     │         │          │          │
+     │         ▼          ▼          ▼
+┌────┴─────────────────────────────────┐
+│  Swin Decoder                        │  Upsampling with skip connections
+│  4 stages                            │  Scales: 1/32→1/16→1/8→1/4→1/1
+│  Channels: 768→384→192→96→48→2       │
+└────────────────┬─────────────────────┘
+                 │
+                 │ Correction (B,2,H,W)
+                 │
+                 ▼
+              ┌─────┐
+              │ ADD │ ◄─── Dirty Input (residual connection)
+              └──┬──┘
+                 │
+                 ▼
+Output: Clean Reconstruction (B,2,H,W)
+```
+
+## Installation
+
+This project uses [uv](https://github.com/astral-sh/uv) for dependency management.
+
+1.  **Install uv** (if not already installed):
+    ```bash
+    curl -LsSf https://astral.sh/uv/install.sh | sh
+    ```
+
+2.  **Sync Dependencies**:
+    Navigate to the `holopaswin` directory and run:
+    ```bash
+    uv sync
+    ```
+    This will create a virtual environment and install all locked dependencies from `uv.lock`.
+
+3.  **Training**:
+    To start training, run:
+    ```bash
+    uv run src/train.py
+    ```
+
+## Development
 
-## Development Setup
+-   **HOLO-PASWIN v2** builds upon Swin Transformer U-Net architecture.
+-   It accepts Hologram Intensity and outputs Complex Object (Phase & Amplitude).
+-   Dataset is loaded from efficient Parquet files.
 
 ### Installing Git Hooks
 

diff --git a/notebooks/holopaswin_exp1.ipynb b/notebooks/holopaswin_exp1.ipynb
diff --git a/notebooks/holopaswin_exp2.ipynb b/notebooks/holopaswin_exp2.ipynb
diff --git a/pyproject.toml b/pyproject.toml
@@ -38,6 +38,8 @@ dependencies = [
     "scikit-image>=0.20.0",
     "matplotlib>=3.7.0",
     "tqdm>=4.65.0",
+    "pandas>=2.3.3",
+    "pyarrow>=22.0.0",
 ]
 
 [project.optional-dependencies]
@@ -110,3 +112,10 @@ module = [
 ]
 ignore_missing_imports = true
 
+[dependency-groups]
+dev = [
+    "mypy>=1.19.0",
+    "ruff>=0.14.9",
+    "types-tqdm>=4.67.0.20250809",
+]
+
diff --git a/results/ablation_study/ablation_study_20260202_235539.log b/results/ablation_study/ablation_study_20260202_235539.log
@@ -0,0 +1,76 @@
+Starting ablation study at Mon Feb  2 23:55:39 CET 2026
+This will take approximately 20 hours to complete.
+Logging to: ablation_study_20260202_235539.log
+
+[2026-02-02 22:55:40] ================================================================================
+[2026-02-02 22:55:40] SEQUENTIAL ABLATION STUDY EXECUTION
+[2026-02-02 22:55:40] ================================================================================
+[2026-02-02 22:55:40] Output directory: results/ablation_final
+[2026-02-02 22:55:40] Start time: 2026-02-02 22:55:40
+[2026-02-02 22:55:40] 
+[2026-02-02 22:55:40] 
+[2026-02-02 22:55:40] ================================================================================
+[2026-02-02 22:55:40] EXPERIMENT 1: Loss Component Ablation
+[2026-02-02 22:55:40] Expected duration: ~10 hours
+[2026-02-02 22:55:40] ================================================================================
+[2026-02-02 22:55:40] Starting: ablation_loss_components.py
+[2026-02-02 22:55:40]   Args: --output-dir results/ablation_final/loss_components
+[2026-02-02 22:55:40]   Log file: results/ablation_final/loss_components.log
+[2026-02-03 01:01:36] ✓ ablation_loss_components.py completed successfully (125.9 min)
+[2026-02-03 01:01:36] 
+[2026-02-03 01:01:36] ================================================================================
+[2026-02-03 01:01:36] EXPERIMENT 2: Loss Weight Sensitivity
+[2026-02-03 01:01:36] Expected duration: ~3-4 hours
+[2026-02-03 01:01:36] ================================================================================
+[2026-02-03 01:01:36] Starting: ablation_loss_weights_sensitivity.py
+[2026-02-03 01:01:36]   Args: --output-dir results/ablation_final/loss_weights
+[2026-02-03 01:01:36]   Log file: results/ablation_final/loss_weights.log
+[2026-02-03 03:02:05] ✓ ablation_loss_weights_sensitivity.py completed successfully (120.5 min)
+[2026-02-03 03:02:05] 
+[2026-02-03 03:02:05] ================================================================================
+[2026-02-03 03:02:05] EXPERIMENT 3: Architecture Ablation
+[2026-02-03 03:02:05] Expected duration: ~8 hours
+[2026-02-03 03:02:05] ================================================================================
+[2026-02-03 03:02:05] Starting: ablation_architecture.py
+[2026-02-03 03:02:05]   Args: --output-dir results/ablation_final/architecture
+[2026-02-03 03:02:05]   Log file: results/ablation_final/architecture.log
+[2026-02-03 07:05:09] ✓ ablation_architecture.py completed successfully (243.1 min)
+[2026-02-03 07:05:09] 
+[2026-02-03 07:05:09] ================================================================================
+[2026-02-03 07:05:09] EXPERIMENT 4: Robustness Tests
+[2026-02-03 07:05:09] Expected duration: ~1 hour
+[2026-02-03 07:05:09] ================================================================================
+[2026-02-03 07:05:09] Starting: ablation_robustness.py
+[2026-02-03 07:05:09]   Args: --model-path results/ablation_final/loss_components/full_model_best.pth --output-dir results/ablation_final/robustness
+[2026-02-03 07:05:09]   Log file: results/ablation_final/robustness.log
+[2026-02-03 07:05:16] ✗ ablation_robustness.py failed with code 1 (0.1 min)
+[2026-02-03 07:05:16] 
+[2026-02-03 07:05:16] ================================================================================
+[2026-02-03 07:05:16] EXECUTION SUMMARY
+[2026-02-03 07:05:16] ================================================================================
+[2026-02-03 07:05:16] Total time: 8.16 hours (489.6 minutes)
+[2026-02-03 07:05:16] 
+[2026-02-03 07:05:16] ✓ SUCCESS    Loss Components            125.9 min
+[2026-02-03 07:05:16] ✓ SUCCESS    Loss Weights               120.5 min
+[2026-02-03 07:05:16] ✓ SUCCESS    Architecture               243.1 min
+[2026-02-03 07:05:16] ✗ FAILED     Robustness                   0.1 min
+[2026-02-03 07:05:16] 
+[2026-02-03 07:05:16] Results saved to:
+[2026-02-03 07:05:16]   - Loss components: results/ablation_final/loss_components
+[2026-02-03 07:05:16]   - Loss weights: results/ablation_final/loss_weights
+[2026-02-03 07:05:16]   - Architecture: results/ablation_final/architecture
+[2026-02-03 07:05:16]   - Robustness: results/ablation_final/robustness
+[2026-02-03 07:05:16] 
+[2026-02-03 07:05:16] Next steps:
+[2026-02-03 07:05:16] 1. Review CSV files in each results directory
+[2026-02-03 07:05:16] 2. Copy LaTeX tables to article/src/tables/
+[2026-02-03 07:05:16] 3. Copy z_mismatch_plot.png to article/src/figs/
+[2026-02-03 07:05:16] 4. Update article tables with actual values
+[2026-02-03 07:05:16] 5. Compile article with pdflatex
+[2026-02-03 07:05:16] 
+[2026-02-03 07:05:16] End time: 2026-02-03 07:05:16
+[2026-02-03 07:05:16] ================================================================================
+
+Ablation study completed at Tue Feb  3 08:05:16 CET 2026
+Results are in: results/ablation_final/
+Shell log saved to: ablation_study_20260202_235539.log
diff --git a/results/ablation_study/architecture.log b/results/ablation_study/architecture.log
diff --git a/results/ablation_study/architecture/architecture_ablation.csv b/results/ablation_study/architecture/architecture_ablation.csv
@@ -0,0 +1,5 @@
+model,model_id,num_parameters,inference_time_ms,training_time_sec,best_val_loss,amp_ssim,phase_ssim,amp_psnr,phase_psnr,freq_ssim,bs_ratio
+"HoloPASWIN (Swin-Tiny, Pretrained, Residual)",swin_pretrained_residual,30883820,12.147433757781982,1386.3456587791443,0.18496015738873256,0.8970394458833664,0.937056549917808,37.65529804884799,43.58432071069855,0.14945638298392874,0.6815521614325624
+"ResNet-18 U-Net (Pretrained, Residual)",resnet18_pretrained_residual,15559842,9.034426212310791,958.1014878749847,0.4587241330790141,0.47471485622771775,0.7523955224838984,29.249287768628708,41.07767062065633,-0.10465396731275439,0.7102488699038664
+"HoloPASWIN (Swin-Tiny, Pretrained, Direct)",swin_pretrained_direct,30883820,13616.046197414398,1613.4761559963226,0.1556178300626694,0.9704758093909414,0.958927794765349,41.20732633444617,44.604655123042804,0.2502256519645422,0.5659946480225171
+"HoloPASWIN (Swin-Tiny, From Scratch, Residual)",swin_scratch_residual,30883820,39.06041860580444,8441.657612085342,0.14487511510886844,0.9526423536533443,0.9567646939222026,38.84532393151695,42.090110098302475,0.28041522251086287,1.2115895154495393
diff --git a/results/ablation_study/architecture/architecture_table.tex b/results/ablation_study/architecture/architecture_table.tex
@@ -0,0 +1,15 @@
+\begin{table}[h]
+\centering
+\caption{Architecture ablation study comparing Swin Transformer vs ResNet-18, residual vs direct reconstruction, and pretrained vs from-scratch training.}
+\label{tab:architecture_ablation}
+\scriptsize
+\begin{tabular}{@{}lcccccc@{}}
+\toprule
+\textbf{Model} & \textbf{Params (M)} & \textbf{Inf. Time (ms)} & \textbf{Phase PSNR} & \textbf{Phase SSIM} & \textbf{B/S Ratio} \\ \midrule
+HoloPASWIN (Swin-Tiny, Pretrained, Residual) & 30.88 & 12.1 & 43.58 & 0.9371 & 0.6816 \\
+ResNet-18 U-Net (Pretrained, Residual) & 15.56 & 9.0 & 41.08 & 0.7524 & 0.7102 \\
+HoloPASWIN (Swin-Tiny, Pretrained, Direct) & 30.88 & 13616.0 & 44.60 & 0.9589 & 0.5660 \\
+HoloPASWIN (Swin-Tiny, From Scratch, Residual) & 30.88 & 39.1 & 42.09 & 0.9568 & 1.2116 \\
+\bottomrule
+\end{tabular}
+\end{table}
diff --git a/results/ablation_study/loss_components.log b/results/ablation_study/loss_components.log
diff --git a/results/ablation_study/loss_components/ablation_results.csv b/results/ablation_study/loss_components/ablation_results.csv
@@ -0,0 +1,6 @@
+config_name,config_id,training_time_sec,best_val_loss,amp_ssim,phase_ssim,amp_psnr,phase_psnr,freq_ssim,bs_ratio
+Full Model (Baseline),full_model,1396.5777411460876,0.1689160324278332,0.9591731344876756,0.9679168719273492,37.72111893437823,45.57350964001303,0.212460134640332,0.374627482566622
+Full - L_freq,no_freq,1413.2068510055542,0.026857876617993628,0.7303544944024373,0.9655956369427163,34.74318675465086,46.574444171130914,-0.022423141127542526,0.5300978468731046
+Full - L_phy,no_phy,1594.3169980049133,0.1562851546775727,0.8704455189791516,0.961978000050189,37.17501457953428,44.13533834375469,0.2285210327001943,0.7219058168270895
+Only L_complex,only_complex,1569.116487979889,0.021731258029975588,0.6296047793003886,0.9713036532540739,33.87423366601785,46.83406757897833,-0.07376106024671542,0.5267030513214488
+"Pure Spatial (no L_freq, no L_phy)",pure_spatial,1395.5910940170288,0.024972101791747033,0.6252144627465249,0.9525428230300735,33.29202170586924,44.721528451134866,-0.06412226380691187,0.6361389030071516
diff --git a/results/ablation_study/loss_components/ablation_table.tex b/results/ablation_study/loss_components/ablation_table.tex
@@ -0,0 +1,15 @@
+\begin{table}[h]
+\centering
+\caption{Loss function component ablation study. All models trained for 5 epochs on 25,000 samples.}
+\label{tab:loss_component_ablation}
+\begin{tabular}{@{}lcccccc@{}}
+\toprule
+\textbf{Configuration} & \textbf{Amp SSIM} & \textbf{Phase SSIM} & \textbf{Phase PSNR} & \textbf{Freq SSIM} & \textbf{B/S Ratio} \\ \midrule
+Full Model (Baseline) & 0.9592 & 0.9679 & 45.57 & 0.2125 & 0.3746 \\
+Full - L_freq & 0.7304 & 0.9656 & 46.57 & -0.0224 & 0.5301 \\
+Full - L_phy & 0.8704 & 0.9620 & 44.14 & 0.2285 & 0.7219 \\
+Only L_complex & 0.6296 & 0.9713 & 46.83 & -0.0738 & 0.5267 \\
+Pure Spatial (no L_freq, no L_phy) & 0.6252 & 0.9525 & 44.72 & -0.0641 & 0.6361 \\
+\bottomrule
+\end{tabular}
+\end{table}
diff --git a/results/ablation_study/loss_components/full_model_history.csv b/results/ablation_study/loss_components/full_model_history.csv
@@ -0,0 +1,4 @@
+train_loss,val_loss,epoch
+0.3789018812417984,0.23248187369770473,1
+0.23278295228481294,0.18489884052957808,2
+0.1911716034412384,0.1689160324278332,3
diff --git a/results/ablation_study/loss_components/no_freq_history.csv b/results/ablation_study/loss_components/no_freq_history.csv
@@ -0,0 +1,4 @@
+train_loss,val_loss,epoch
+0.0881531357705593,0.03953465228042905,1
+0.03408360904753208,0.035909083272729604,2
+0.0296357734978199,0.026857876617993628,3
diff --git a/results/ablation_study/loss_components/no_phy_history.csv b/results/ablation_study/loss_components/no_phy_history.csv
@@ -0,0 +1,4 @@
+train_loss,val_loss,epoch
+0.3440481254577637,0.21518866741468037,1
+0.18709273314476013,0.1647950524375552,2
+0.16642933180332184,0.1562851546775727,3
diff --git a/results/ablation_study/loss_components/only_complex_history.csv b/results/ablation_study/loss_components/only_complex_history.csv
@@ -0,0 +1,4 @@
+train_loss,val_loss,epoch
+0.06533894625306129,0.030536581067338822,1
+0.024858536159992218,0.021731258029975588,2
+0.020725958310067655,0.023033877036401203,3
diff --git a/results/ablation_study/loss_components/pure_spatial_history.csv b/results/ablation_study/loss_components/pure_spatial_history.csv
@@ -0,0 +1,4 @@
+train_loss,val_loss,epoch
+0.0922015791118145,0.04076199269010907,1
+0.03314999188482761,0.03344916005337995,2
+0.02780363717675209,0.024972101791747033,3
diff --git a/results/ablation_study/loss_weights.log b/results/ablation_study/loss_weights.log
diff --git a/results/ablation_study/loss_weights/lambda_sensitivity.csv b/results/ablation_study/loss_weights/lambda_sensitivity.csv
@@ -0,0 +1,4 @@
+config,lambda_phy,amp_ssim,phase_ssim,amp_psnr,phase_psnr,freq_ssim,bs_ratio
+No Physics (λ=0.0),0.0,0.8362126284857192,0.9485394833479281,36.56277850857484,43.27810880819259,0.19819815184803266,0.8650377859031001
+Current (λ=0.1),0.1,0.9277890965142687,0.9695644435414934,38.426569654699215,45.90852256750022,0.23131856952639637,0.47128404917255523
+High Physics (λ=0.2),0.2,0.9040665044904769,0.9635805658961512,38.69888800668452,45.41432580026462,0.2308427794923376,0.49929088282008327
diff --git a/results/ablation_study/loss_weights/weight_sensitivity.csv b/results/ablation_study/loss_weights/weight_sensitivity.csv
@@ -0,0 +1,4 @@
+config,final_val_loss
+Amp-heavy (0.5/0.2/0.15/0.15),0.1537831380726799
+Balanced (0.4/0.2/0.2/0.2) [Current],0.17710744506782955
+Phase-heavy (0.3/0.3/0.2/0.2),0.18449455830785963
diff --git a/results/ablation_study/main.log b/results/ablation_study/main.log
@@ -0,0 +1,68 @@
+[2026-02-02 22:55:40] ================================================================================
+[2026-02-02 22:55:40] SEQUENTIAL ABLATION STUDY EXECUTION
+[2026-02-02 22:55:40] ================================================================================
+[2026-02-02 22:55:40] Output directory: results/ablation_final
+[2026-02-02 22:55:40] Start time: 2026-02-02 22:55:40
+[2026-02-02 22:55:40] 
+[2026-02-02 22:55:40] 
+[2026-02-02 22:55:40] ================================================================================
+[2026-02-02 22:55:40] EXPERIMENT 1: Loss Component Ablation
+[2026-02-02 22:55:40] Expected duration: ~10 hours
+[2026-02-02 22:55:40] ================================================================================
+[2026-02-02 22:55:40] Starting: ablation_loss_components.py
+[2026-02-02 22:55:40]   Args: --output-dir results/ablation_final/loss_components
+[2026-02-02 22:55:40]   Log file: results/ablation_final/loss_components.log
+[2026-02-03 01:01:36] ✓ ablation_loss_components.py completed successfully (125.9 min)
+[2026-02-03 01:01:36] 
+[2026-02-03 01:01:36] ================================================================================
+[2026-02-03 01:01:36] EXPERIMENT 2: Loss Weight Sensitivity
+[2026-02-03 01:01:36] Expected duration: ~3-4 hours
+[2026-02-03 01:01:36] ================================================================================
+[2026-02-03 01:01:36] Starting: ablation_loss_weights_sensitivity.py
+[2026-02-03 01:01:36]   Args: --output-dir results/ablation_final/loss_weights
+[2026-02-03 01:01:36]   Log file: results/ablation_final/loss_weights.log
+[2026-02-03 03:02:05] ✓ ablation_loss_weights_sensitivity.py completed successfully (120.5 min)
+[2026-02-03 03:02:05] 
+[2026-02-03 03:02:05] ================================================================================
+[2026-02-03 03:02:05] EXPERIMENT 3: Architecture Ablation
+[2026-02-03 03:02:05] Expected duration: ~8 hours
+[2026-02-03 03:02:05] ================================================================================
+[2026-02-03 03:02:05] Starting: ablation_architecture.py
+[2026-02-03 03:02:05]   Args: --output-dir results/ablation_final/architecture
+[2026-02-03 03:02:05]   Log file: results/ablation_final/architecture.log
+[2026-02-03 07:05:09] ✓ ablation_architecture.py completed successfully (243.1 min)
+[2026-02-03 07:05:09] 
+[2026-02-03 07:05:09] ================================================================================
+[2026-02-03 07:05:09] EXPERIMENT 4: Robustness Tests
+[2026-02-03 07:05:09] Expected duration: ~1 hour
+[2026-02-03 07:05:09] ================================================================================
+[2026-02-03 07:05:09] Starting: ablation_robustness.py
+[2026-02-03 07:05:09]   Args: --model-path results/ablation_final/loss_components/full_model_best.pth --output-dir results/ablation_final/robustness
+[2026-02-03 07:05:09]   Log file: results/ablation_final/robustness.log
+[2026-02-03 07:05:16] ✗ ablation_robustness.py failed with code 1 (0.1 min)
+[2026-02-03 07:05:16] 
+[2026-02-03 07:05:16] ================================================================================
+[2026-02-03 07:05:16] EXECUTION SUMMARY
+[2026-02-03 07:05:16] ================================================================================
+[2026-02-03 07:05:16] Total time: 8.16 hours (489.6 minutes)
+[2026-02-03 07:05:16] 
+[2026-02-03 07:05:16] ✓ SUCCESS    Loss Components            125.9 min
+[2026-02-03 07:05:16] ✓ SUCCESS    Loss Weights               120.5 min
+[2026-02-03 07:05:16] ✓ SUCCESS    Architecture               243.1 min
+[2026-02-03 07:05:16] ✗ FAILED     Robustness                   0.1 min
+[2026-02-03 07:05:16] 
+[2026-02-03 07:05:16] Results saved to:
+[2026-02-03 07:05:16]   - Loss components: results/ablation_final/loss_components
+[2026-02-03 07:05:16]   - Loss weights: results/ablation_final/loss_weights
+[2026-02-03 07:05:16]   - Architecture: results/ablation_final/architecture
+[2026-02-03 07:05:16]   - Robustness: results/ablation_final/robustness
+[2026-02-03 07:05:16] 
+[2026-02-03 07:05:16] Next steps:
+[2026-02-03 07:05:16] 1. Review CSV files in each results directory
+[2026-02-03 07:05:16] 2. Copy LaTeX tables to article/src/tables/
+[2026-02-03 07:05:16] 3. Copy z_mismatch_plot.png to article/src/figs/
+[2026-02-03 07:05:16] 4. Update article tables with actual values
+[2026-02-03 07:05:16] 5. Compile article with pdflatex
+[2026-02-03 07:05:16] 
+[2026-02-03 07:05:16] End time: 2026-02-03 07:05:16
+[2026-02-03 07:05:16] ================================================================================