| Version range | Main idea | Data scale | Notes |
|---|---|---|---|
| v1 | Context last-frame decoder | 14 files | Initial production-style ConvLSTM baseline |
| v3 | Flatten ReLU baseline | 14 files | Lightweight smoke/baseline model |
| v4-v9 | Multi-head attention variants | 14 files | BCE, Tversky, CBAM, stacked ConvLSTM trials |
| v10-v13 | Spatial UNet family | 120 files | Streaming input, holdout sample, mask/time channels |
| v14-v16 | 365-file experiments | 341 matched files | Larger data regime, gated UNet, production v16 |
- Version:
v16 - Variant:
multihead_spatial_unet_gated - Parameters:
335,741 - Input shape:
12 x 66 x 66 x 5 - Added channels: inflow, outflow, occupancy mask, time sin, time cos
- Training monitor:
val_ssim, maximize - Best epoch:
24 - Epochs trained:
38
Production v16 improved over the weakest observed model by:
- R2:
+873.8% - SSIM:
+39.6% - Occupancy F1:
+231.7% - Outflow MAE:
+60.5%reduction from worst - Validation MAE:
+66.5%reduction from worst
The best F1 model was v15 at 0.5608, while production v16 reached 0.5485 and achieved the best R2 and SSIM.
These improvements are based on actual-value evaluation: predictions and labels were inverse-transformed to the original AIS traffic scale before computing R2, SSIM, F1, and MAE. See actual_value_evaluation.md.
v16 is a strong production candidate because it combines:
- Larger 365-file training setup
- Gated spatial UNet architecture
- Two ConvLSTM layers
- Route occupancy prior
- Time-of-day channels
- Monitoring based on SSIM rather than only loss
This improved structural similarity and global fit while keeping F1 close to the best model.