Pocket Wgan Lab

Verified project card

# Pocket WGAN-GP

Pocket WGAN-GP is a compact class-conditional adversarial image generator for the
8x8 handwritten-digits benchmark. It combines a projection critic, Wasserstein
training with gradient penalty, and an auxiliary class objective. The interactive
Space exposes class, noise seed, and latent temperature.

## Why this experiment exists

GANs can report strong-looking averages while collapsing to one prototype. This
project therefore treats class fidelity, within-class variance, quantized sample
uniqueness, nearest-training-example distance, and exact-copy rate as first-class
evaluation outputs. A frozen 2,198-parameter classifier trained in Tiny Vision
Foundry supplies the label-fidelity measurement.

## Verified local result

The 31,712-parameter generator reached 97.0% class fidelity across 1,000 held-out
generations. Its mean within-class pixel variance was 98.24% of the real training
data baseline. Every quantized sample was unique, the exact training-copy rate was
zero, and mean nearest-training-image MSE was 0.03026. The critic contains 17,931
parameters and received 7,200 updates.

## Reproduce

```bash
uv run python projects/pocket-wgan/train.py
uv run pytest tests/test_pocket_wgan.py
```

The published bundle contains the generator and critic in SafeTensors format,
the complete evaluation report, a 100-sample contact sheet, and 1,000 generated
evaluation samples. Reported numbers are local measurements, not claims about a
currently deployed Hub Space.

Evaluation snapshot

{
  "model": "Pocket WGAN-GP",
  "method": "Projection-conditioned WGAN-GP with auxiliary class supervision",
  "generator_parameters": 31712,
  "critic_parameters": 17931,
  "training_examples": 1257,
  "generator_steps": 2400,
  "critic_updates": 7200,
  "best_generator_step": 2400,
  "generation": {
    "judge_accuracy": 0.9700000286102295,
    "judge_accuracy_by_class": {
      "0": 0.9900000095367432,
      "1": 0.949999988079071,
      "2": 0.9900000095367432,
      "3": 0.9700000286102295,
      "4": 0.9599999785423279,
      "5": 0.9700000286102295,
      "6": 0.9900000095367432,
      "7": 1.0,
      "8": 0.9599999785423279,
      "9": 0.9200000166893005
    },
    "mean_pixel_variance_by_class": {
      "0": 0.0305124931037426,
      "1": 0.04492313414812088,
      "2": 0.03864439204335213,
      "3": 0.04265550151467323,
      "4": 0.0401928536593914,
      "5": 0.04807004705071449,
      "6": 0.032149579375982285,
      "7": 0.04377623647451401,
      "8": 0.04335470125079155,
      "9": 0.04571780934929848
    },
    "real_mean_pixel_variance_by_class": {
      "0": 0.025008603930473328,
      "1": 0.05708667263388634,
      "2": 0.04482439160346985,
      "3": 0.03898561745882034,
      "4": 0.04554782807826996,
      "5": 0.04596441984176636,
      "6": 0.03248969465494156,
      "7": 0.046020835638046265,
      "8": 0.044565171003341675,
      "9": 0.044857367873191833
    },
    "diversity_ratio_by_class": {
      "0": 1.2200798248703002,
      "1": 0.7869285785182504,
      "2": 0.8621286460553024,
      "3": 1.0941343063176903,
      "4": 0.8824318382497515,
      "5": 1.0458099376908663,
      "6": 0.9895315950927983,
      "7": 0.9512264579203641,
      "8": 0.9728382114261526,
      "9": 1.0191817201254216
    },
    "mean_diversity_ratio": 0.9824291116266899,
    "samples": 1000,
    "collapse_and_memorization_checks": {
      "mean_nearest_training_mse": 0.030259807055350394,
      "median_nearest_training_mse": 0.02845568861812353,
      "exact_training_copy_fraction": 0.0,
      "quantized_unique_fraction_by_class": {
        "0": 1.0,
        "1": 1.0,
        "2": 1.0,
        "3": 1.0,
        "4": 1.0,
        "5": 1.0,
        "6": 1.0,
        "7": 1.0,
        "8": 1.0,
        "9": 1.0
      },
      "mean_quantized_unique_fraction": 1.0
    }
  },
  "judge": "Frozen Tiny Vision student, 98.52% real-image test accuracy",
  "training_history": [
    {
      "generator_step": 1,
      "generator_loss": 1.832496166229248,
      "critic_loss": 7.296276569366455,
      "gradient_penalty": 0.6319434642791748,
      "judge_fidelity": 0.08749999850988388,
      "mean_diversity_ratio": 0.09041444920003414,
      "selection_score": 0.101062165889889
    },
    {
      "generator_step": 200,
      "generator_loss": 2.1555261611938477,
      "critic_loss": -0.743652880191803,
      "gradient_penalty": 0.00692558940500021,
      "judge_fidelity": 0.24250000715255737,
      "mean_diversity_ratio": 0.03933937028050423,
      "selection_score": 0.24840091269463302
    },
    {
      "generator_step": 400,
      "generator_loss": 1.0727102756500244,
      "critic_loss": -0.5879160165786743,
      "gradient_penalty": 0.005078175105154514,
      "judge_fidelity": 1.0,
      "mean_diversity_ratio": 0.03597373897209764,
      "selection_score": 1.0053960608458146
    },
    {
      "generator_step": 600,
      "generator_loss": 1.01450514793396,
      "critic_loss": -0.5324597358703613,
      "gradient_penalty": 0.006901979446411133,
      "judge_fidelity": 0.9950000047683716,
      "mean_diversity_ratio": 0.10493496991693974,
      "selection_score": 1.0107402502559126
    },
    {
      "generator_step": 800,
      "generator_loss": 1.043497920036316,
      "critic_loss": -0.49861836433410645,
      "gradient_penalty": 0.008488744497299194,
      "judge_fidelity": 0.9775000214576721,
      "mean_diversity_ratio": 0.3373389258980751,
      "selection_score": 1.0281008603423833
    },
    {
      "generator_step": 1000,
      "generator_loss": 0.9358897805213928,
      "critic_loss": -0.4203786253929138,
      "gradient_penalty": 0.009127671830356121,
      "judge_fidelity": 0.9725000262260437,
      "mean_diversity_ratio": 0.43268701434135437,
      "selection_score": 1.0374030783772468
    },
    {
      "generator_step": 1200,
      "generator_loss": 1.0532721281051636,
      "critic_loss": -0.30880528688430786,
      "gradient_penalty": 0.006839464418590069,
      "judge_fidelity": 0.9725000262260437,
      "mean_diversity_ratio": 0.58423230946064,
      "selection_score": 1.0601348726451396
    },
    {
      "generator_step": 1400,
      "generator_loss": 1.1917768716812134,
      "critic_loss": -0.25530850887298584,
      "gradient_penalty": 0.0061252727173268795,
      "judge_fidelity": 0.9674999713897705,
      "mean_diversity_ratio": 0.6854087769985199,
      "selection_score": 1.0703112879395484
    },
    {
      "generator_step": 1600,
      "generator_loss": 1.0977249145507812,
      "critic_loss": -0.2501622140407562,
      "gradient_penalty": 0.0051786331459879875,
      "judge_fidelity": 0.9524999856948853,
      "mean_diversity_ratio": 0.7668859779834747,
      "selection_score": 1.0675328823924064
    },
    {
      "generator_step": 1800,
      "generator_loss": 1.106730580329895,
      "critic_loss": -0.24072407186031342,
      "gradient_penalty": 0.0037523529026657343,
      "judge_fidelity": 0.9649999737739563,
      "mean_diversity_ratio": 0.8528995037078857,
      "selection_score": 1.092934899330139
    },
    {
      "generator_step": 2000,
      "generator_loss": 1.3110945224761963,
      "critic_loss": -0.1876629739999771,
      "gradient_penalty": 0.004364368040114641,
      "judge_fidelity": 0.9549999833106995,
      "mean_diversity_ratio": 0.9044061362743377,
      "selection_score": 1.0906609037518502
    },
    {
      "generator_step": 2200,
      "generator_loss": 1.3018580675125122,
      "critic_loss": -0.1839527189731598,
      "gradient_penalty": 0.00520103657618165,
      "judge_fidelity": 0.9624999761581421,
      "mean_diversity_ratio": 0.899994432926178,
      "selection_score": 1.0974991410970687
    },
    {
      "generator_step": 2400,
      "generator_loss": 1.5436924695968628,
      "critic_loss": -0.16877740621566772,
      "gradient_penalty": 0.0042749010026454926,
      "judge_fidelity": 0.9599999785423279,
      "mean_diversity_ratio": 1.0144544422626496,
      "selection_score": 1.1099999785423278
    }
  ]
}

Backed-up artifact tree

• README.md
• __pycache__/app.cpython-311.pyc
• __pycache__/model.cpython-311.pyc
• app.py
• artifacts/pocket-wgan/critic.safetensors
• artifacts/pocket-wgan/evaluation.json
• artifacts/pocket-wgan/generated_samples.npz
• artifacts/pocket-wgan/generator.safetensors
• artifacts/pocket-wgan/samples.png
• data/evaluation_samples.parquet
• model.py
• requirements.txt
• train.py