Source reference for the traffic CapsNet and CNN in main_capsnet.py. For research attribution, use the paper citation.
- The University of Sheffield software record names the four paper authors and links to this repository.
- The record labels the project MIT. Obtain the applicable license and copyright text from the rights holder before redistribution; the checkout supplies no license file.
routingcredits Huadong Liao; the corresponding upstream capsLayer.py is distributed under Apache-2.0. Preserve applicable upstream notices when reusing that implementation.- The paper attributes traffic-data collection to SETA; consult the dataset guide for supplied files and provenance. Software and dataset permissions are separate.
The paper gives the research architecture. These source conventions matter when adapting it:
| Component | Source behavior |
|---|---|
| Routing | Both branches use the unsquashed s_J after computing squash(s_J); the paper describes squashed sums. |
| Optimization | Adam minimizes normalized RMSE; the paper states MSE. Slim registers L2 terms separately from the minimized loss. |
| Scaling | One global minimum and maximum are fitted before the chronological split, including evaluation values. |
| Error units | scale_error_inverse multiplies normalized RMSE by the speed range divided by the feature range. Speed predictions use scale_inverse, which also restores the offset. |
| Summary | “Average Loss” is the mean of batch RMSE values; pooled RMSE requires pooling squared errors before taking the root. |
| Windows | The loop processes a subset of constructed windows; exact counts are in the running guide. |
| Data | Bundled files contain 33,504 rows. Defaults select 5,000 rows and a 75% training split; paper experiments use their own configurations. |
The synthetic example checks scaling and provides a TensorFlow network exercise. Run the error-unit regression checks with python -m unittest discover -s tests -v; the fixtures use known speed differences and zero error.