How raiyan22/waste works
This repository is a research workspace, not a conventional application codebase. Its center of gravity is a NuScenes-based pedestrian intent prediction project: a hybrid PyTorch model that fuses trajectory history with environmental context such as weather and time-of-day.
1. Core research objective
The README describes the main architecture: an LSTM branch for temporal pedestrian kinematics, a separate context encoder for scene metadata, and a feature-level fusion stage for intent classification. The intended output has evolved from binary intent to multiclass behavior labels.
2. Main data boundary
The dataset boundary is external and important: NuScenes v1.0-mini is expected to live in Google Drive, then be mounted in Colab, unzipped locally, and read through nuscenes-devkit. That means the runtime architecture is “Drive storage -> Colab local runtime -> dataset parser -> PyTorch tensors.”
3. Model pipeline
The ML pipeline implied by the README is:
NuScenes raw annotations -> pedestrian trajectory extraction -> context feature extraction -> fused latent representation -> classification head.
This is the only clearly defined computation path in the repo.
4. Experiment progression
The project appears to be staged in two phases: first synthetic-data validation, then real-data training on NuScenes. That suggests the notebook is likely the execution layer for proving the architecture works before full dataset integration.
5. Notebook as execution surface
Untitled0.ipynb is likely the main interactive development artifact. In this repo, the notebook is probably where data loading, model definition, training, and debugging happen, while the text files serve as companion notes.
6. Research/process documentation layer
Several .md files (copy-repo-remote.md, docker-cmds.md, gulli-trivy.md, kyc.md, notif-slk.md, oss-ollama-claudecode.md, prank.md) look like operational notes, prompts, or auxiliary experiments rather than part of the model itself. They form a “workflow/documentation” layer around the research project.
7. Prompt and text artifacts
Files like prompt, try-deep.txt, and stat.txt suggest prompt engineering, experiment notes, or status tracking. These are supporting artifacts that help reproduce or communicate the research, but they are not the core inference path.
8. HTML/JS side artifacts
abcd.html, chatpa.html, lio.html, and paste-notif.html, plus chat-data.js, look like standalone browser experiments or small UI snippets. They are isolated from the NuScenes model architecture and should be treated as a separate sidecar layer, not part of the ML pipeline.
9. Binary/archive assets
humanize.zip is a packaged asset bundle, likely unrelated to the primary model flow unless explicitly unpacked into a demo or prompt-related workflow. Its presence reinforces that this repo is a mixed artifact store, not a single-purpose source tree.
10. Repository boundaries to show in an architecture diagram
A correct diagram would show:
• External sources: NuScenes download site, Google Drive, Colab
• Execution layer: notebook-based PyTorch training
• Data layer: raw NuScenes annotations and local extracted samples
• Model layer: LSTM + context encoder + fusion classifier
• Support layer: markdown notes, prompts, HTML/JS experiments, and archives
The important architectural point is that the repository is organized around research preparation and experimentation, with the actual model pipeline living primarily in the notebook and README-defined workflow.
Main entry points: Untitled0.ipynb Training loop · prompt Prompt notes · README.md README spec · copy-repo-remote.md Ops notes
Components 18
External sources
Off-repo inputs and runtime services used by the research workflow.
- NuScenes data · External NuScenes v1.0-mini annotations and scenes.
- Google Drive · Expected remote storage for the dataset bundle.
- Colab runtime · Execution environment where data is mounted and training runs.
Core research
Notebook-centered ML pipeline for pedestrian intent prediction.
- NuScenes devkit · Library used to read and structure NuScenes data.
- Traj extract · Builds pedestrian trajectory histories from raw annotations.
- Context extract · Derives weather, time-of-day, and other scene context features.
- LSTM branch · Encodes pedestrian kinematics over time.
- Context encoder · Encodes environmental and scene metadata.
- Fusion stage · Combines temporal and context embeddings.
- Classifier · Predicts binary or multiclass pedestrian intent labels.
- Training loop · Notebook-driven optimization, validation, and debugging.
Untitled0.ipynb
Support artifacts
Notes, prompts, and operational documents around the project.
- README spec · Defines the intended model design and workflow.
README.md - Ops notes · Repository coordination and environment notes.
copy-repo-remote.md - Prompt notes · Prompting, status, and experiment notes.
prompt - Status text · Loose experiment or progress tracking text.
stat.txt - Archive bundle · Packaged assets or demo content.
humanize.zip
Sidecar experiments
Standalone browser/UI artifacts kept alongside the research work.
- HTML snippets · Standalone browser-side UI or interaction experiments.
abcd.html - Chat script · Client-side data or interaction helper.
chat-data.js
Connections 20
- Google Drive → NuScenes data: stores · Dataset is kept in Drive before use.
- Colab runtime → NuScenes devkit: loads · Notebook runtime imports the dataset parser.
- NuScenes data → NuScenes devkit: parsed by · Raw scenes and annotations are read through the devkit.
- NuScenes devkit → Traj extract: feeds · Parsed annotations are turned into trajectory samples.
- NuScenes devkit → Context extract: feeds · Scene records supply contextual features.
- Traj extract → LSTM branch: input · Trajectory histories are encoded temporally.
- Context extract → Context encoder: input · Context features are encoded separately.
- LSTM branch → Fusion stage: embeddings · Temporal representation enters fusion.
- Context encoder → Fusion stage: embeddings · Context representation enters fusion.
- Fusion stage → Classifier: latent · Fused features drive intent classification.
- Training loop → LSTM branch: trains · Notebook defines and trains the temporal branch.
- Training loop → Context encoder: trains · Notebook defines and trains the context branch.
- Training loop → Fusion stage: orchestrates · Notebook coordinates feature fusion.
- Training loop → Classifier: optimizes · Notebook runs the prediction head and loss.
- README spec → Training loop: describes · README documents the intended workflow executed in the notebook.
- Ops notes → Training loop: supports · Operational notes support setup and execution.
- Prompt notes → Training loop: supports · Prompt and experiment text supplements the workflow.
- Status text → Training loop: tracks · Status notes reflect experiment progress.
- HTML snippets → Chat script: uses · Browser snippets pair with client-side script artifacts.
- Archive bundle → Prompt notes: bundles · Packaged assets may accompany notes or prompts.