How hzxon/deep-learning works
This is a reproducible speech-emotion-recognition research pipeline, with a companion validation workflow. Its main path turns four public corpora into a unified six-class dataset, prepared features and evaluation splits, then trains controlled models and summarizes results across three protocols. The graph also shows the documented pre/post-training gates and research-facing analysis. Some script implementations were not sampled, so their documented roles are represented without inferring internal call sequences beyond the README workflow.
Main entry points: run_experiments.py Experiment runner · train.py Training and evaluation · workflow.py Workflow orchestrator · analysis.py RQ analysis
Components 24
Dataset preparation
- Unified manifest
src/ser/data.py - Manifest preparation
scripts/prepare_manifest.py - Log-mel features
src/ser/features.py - Audio preprocessing
scripts/preprocess.py - Evaluation protocols
src/ser/splits.py - Split generation
scripts/make_splits.py
Experiment execution
- Experiment runner
scripts/run_experiments.py - SER dataset
src/ser/data.py - Feature augmentation
src/ser/features.py - Training and evaluation
src/ser/train.py - CNN model variants
src/ser/models.py - Classification metrics
src/ser/metrics.py - Run artifacts
Results and analysis
- RQ analysis
src/ser/analysis.py - Gap analysis
scripts/gap_analysis.py - Research plots
src/ser/plotting.py - Research report
scripts/make_report.py - Research tables
Validation workflow
- Workflow orchestrator
scripts/workflow.py - Pre-training gates
scripts/pre_training_audit.py - Post-training audit
scripts/post_training_audit.py - Pipeline validation
scripts/validate_pipeline.py
Other components
- Researcher
- Speech corpora
Connections 29
- Speech corpora → Manifest preparation: provides audio
- Researcher → Manifest preparation: starts preparation
- Manifest preparation → Unified manifest: builds manifest
- Researcher → Audio preprocessing: starts preprocessing
- Audio preprocessing → Log-mel features: extracts features
- Researcher → Split generation: chooses seed
- Split generation → Evaluation protocols: builds protocols
- Unified manifest → Audio preprocessing: supplies samples
- Unified manifest → Split generation: supplies samples
- Researcher → Experiment runner: selects experiments
- Experiment runner → Evaluation protocols: uses protocols
- Experiment runner → SER dataset: loads partitions
- SER dataset → Log-mel features: reads features
- SER dataset → Feature augmentation: augments training
- Experiment runner → Training and evaluation: runs training
- Training and evaluation → CNN model variants: builds model
- Training and evaluation → Classification metrics: evaluates predictions
- Training and evaluation → Run artifacts: writes checkpoints and metrics
- Run artifacts → Gap analysis: supplies run results
- Gap analysis → RQ analysis: builds RQ tables
- RQ analysis → Research tables: writes result tables
- Researcher → Research report: requests report
- Research report → RQ analysis: summarizes results
- Research report → Research plots: generates figures
- RQ analysis → Research tables: provides analysis
- Researcher → Workflow orchestrator: runs workflow
- Workflow orchestrator → Pre-training gates: runs pre-training gates
- Workflow orchestrator → Post-training audit: runs post-training audit
- Workflow orchestrator → Pipeline validation: checks pipeline