Skip to content

02_quickstart.md

Latest commit

 

History

History
101 lines (70 loc) · 2.17 KB

02_quickstart.md

File metadata and controls

101 lines (70 loc) · 2.17 KB

Quick Start

Train your first model in 5 minutes.

1. Generate Test Data

DeepFense includes a helper to create dummy data for testing:

python tests/create_samples.py --output-dir ./tests

This creates tests/test.parquet with 20 samples pointing to a dummy WAV file.

2. Check the Config

The example config at deepfense/config/train.yaml is already set up to use this test data. Open it and check:

  • Frontend: Wav2Vec2 from HuggingFace (downloads automatically)
  • Backend: Nes2Net
  • Loss: CrossEntropy + AMSoftmax
  • Data: Points to ./tests/test.parquet

3. Train

python train.py --config deepfense/config/train.yaml

You'll see:

[INFO] Experiment directory: outputs/Wav2Vec2_Nes2Net_Example_20250401_120000
[INFO] Trainable parameters: 1,234,567
Epoch 1/10: 100%|██████████| 2/2 [00:05<00:00]
[INFO] Average Metrics: loss: 0.6931, EER: 50.00

(EER = 50% is expected on random dummy data.)

4. Test

python test.py \
    --config deepfense/config/train.yaml \
    --checkpoint outputs/Wav2Vec2_Nes2Net_Example_*/best_model.pth

5. Use Your Own Data

Replace the dummy data with your real dataset:

import pandas as pd

df = pd.DataFrame({
    "ID": ["utt_001", "utt_002", "utt_003"],
    "path": ["/data/audio1.flac", "/data/audio2.flac", "/data/audio3.flac"],
    "label": ["bonafide", "spoof", "bonafide"],
})
df.to_parquet("my_train.parquet")

Then update the config:

data:
  train:
    parquet_files: ["my_train.parquet"]
  val:
    parquet_files: ["my_val.parquet"]

Output Structure

After training, check outputs/YOUR_EXP_NAME/:

outputs/Wav2Vec2_Nes2Net_Example_20250401_120000/
├── config.yaml          # Saved config (for reproducibility)
├── train.log            # Full training log
├── best_model.pth       # Best checkpoint (by monitor_metric)
├── ckpts/               # All checkpoints
├── results/             # Metrics JSON files
└── plots/               # Training curves (loss, EER, etc.)

Next: Full Tutorial -- complete walkthrough of every config option