Goonable, Model Card

An experimental binary image classifier that flags lewd anime images as goonable, otherwise normal. Runs fully client-side in the browser via ONNX Runtime Web.

MobileNetV3-Small ONNX · opset 17 2 classes 224×224 input ~5.9 MB val acc 83.3%

Overview

This model takes a single RGB image and predicts whether it is goonable (lewd/suggestive anime content) or normal. It is intended for lightweight, on-device content filtering and is experimental, not a moderation authority.

Architecture

The graph is a MobileNetV3-Small (torchvision) convolutional network: an efficient mobile architecture using depthwise-separable convolutions, Squeeze-and-Excitation blocks, and HardSwish/HardSigmoid activations. The final classifier layer was replaced with a 2-output linear head.

Backbone	torchvision `mobilenet_v3_small`
Graph ops	52× Conv, 19× HardSwish, 14× Relu, 10× GlobalAveragePool, 9× HardSigmoid, 2× Gemm (122 nodes total)
Classifier head	`classifier[3]` replaced with `Linear(in, 2)` → 2 logits
Parameters	108 weight tensors · ~5.9 MB on disk (fp32)

Inputs & Outputs

Input name	`input`
Input shape	`[batch, 3, 224, 224]`, NCHW, float32
Output name	`logits`
Output shape	`[batch, 2]`, raw logits (apply softmax)
Class order	`0 = goonable`, `1 = normal` (alphabetical, torchvision `ImageFolder` default)

Preprocessing

Each image is transformed to match the training pipeline (standard torchvision / ImageNet):

Resize so the shorter side is 256 px (aspect-preserving)
Center-crop to 224 × 224
Convert to RGB, scale to [0, 1]
Normalize with mean [0.485, 0.456, 0.406], std [0.229, 0.224, 0.225]
Layout as NCHW float32 tensor

Inference output logits are passed through softmax to produce class probabilities.

Training

Fine-tuned in PyTorch on Google Colab. Transfer learning from ImageNet-pretrained MobileNetV3-Small, trained iteratively over 6 versions, each run resumed from the previous checkpoint on a freshly reviewed batch of images, accumulating ~16k images total across all versions.

Base weights	torchvision `MobileNet_V3_Small_Weights.DEFAULT` (ImageNet), fine-tuned end-to-end
Data source	~16k anime images scraped from the author's Twitter/X liked posts, hand-reviewed and labeled `goonable` / `normal`
Training scheme	6 resumed versions, each fine-tuning on a reviewed batch
Train / val split	80 / 20 per run (`split-folders`, seed 42)
Augmentations	Train: `RandomResizedCrop(224, scale 0.8–1.0)` + `RandomHorizontalFlip`. Val: `Resize(256)` → `CenterCrop(224)`. Both: ImageNet normalize.
Optimizer / LR	Adam, lr = 1e-4
Loss	CrossEntropyLoss
Epochs / batch size	10 epochs (this resumed run) · batch size 32
Hardware	Google Colab, T4 GPU
Export	`torch.onnx.export` → ONNX (opset 17), input `input`, output `logits`

Evaluation

Measured on the 20% held-out validation split during training.

Best val accuracy	83.3%
Final train accuracy	~93–94% (train accuracy outran val → mild overfitting)

Intended Use & Limitations

Intended: lightweight, private, on-device filtering of anime-style images.
Domain: trained on anime imagery, accuracy on photos or other art styles is unknown and likely poor.
Not a definitive content-moderation tool; "lewd" is subjective and label noise is expected.
Binary only, no severity grading or localization.
Predictions can be flipped by adversarial or out-of-distribution inputs.

How to Run

Inference happens entirely in your browser; no image is uploaded. Serve the folder over HTTP and open localhost:

# for example here using python
python -m http.server 8000
# open http://localhost:8000