Clothing Type EfficientNetV2-S

This is a five-class clothing-type image classifier fine-tuned from timm/tf_efficientnetv2_s.in21k_ft_in1k. It predicts one of the following labels, in model-output order:

Model details

The pretrained 1,000-class head is replaced by a dropout layer with probability 0.15 and a five-output linear classifier. The published artifacts are full-precision exports, not quantized variants. The ONNX model supports dynamic batch sizes and fixes each image to 384 pixels high by 256 pixels wide.

Files

• clothing_type-tf_efficientnetv2_s.in21k_ft_in1k.onnx is the portable inference graph.
• clothing_type-tf_efficientnetv2_s.in21k_ft_in1k.safetensors contains the PyTorch EfficientNetClassifier state dictionary.
• config.json records the architecture, labels, and timm model settings.
• preprocessor_config.json records the image preprocessing contract.

The safetensors keys include the custom wrapper's model. prefix and custom classifier head. Load them with the EfficientNetClassifier implementation from the source repository, not as an unmodified upstream timm checkpoint.

Preprocessing

For the model tensor itself:

1. Convert the image to RGB.
2. Resize it with preserved aspect ratio to fit within 256 × 384, using bicubic interpolation when enlarging and area interpolation when reducing.
3. Center-pad it to 256 × 384 with RGB value (127, 127, 127).
4. Rescale unsigned 8-bit pixels by 1 / 255.
5. Normalize each channel with mean 0.5 and standard deviation 0.5.
6. Convert from HWC to NCHW layout and add the batch dimension.

The source data pipeline first detects the largest person with YOLO and crops to that bounding box. Inputs that are already tightly person-centered are closest to the training distribution. Person detection is not part of the published classifier and is unnecessary when the input is already cropped.

ONNX inference

Install huggingface_hub, transformers, timm, torchvision, onnxruntime, numpy, and Pillow, then run:

from huggingface_hub import hf_hub_download
import numpy as np
import onnxruntime as ort
from PIL import Image, ImageOps
from transformers import AutoImageProcessor

REPO_ID = "electblake/clothing_type_classifier"
LABELS = ["denim_shorts", "lingerie", "maid", "nurse", "swim_bikini"]
model_path = hf_hub_download(
    REPO_ID,
    "clothing_type-tf_efficientnetv2_s.in21k_ft_in1k.onnx",
)

image = ImageOps.exif_transpose(Image.open("person.jpg")).convert("RGB")
image_processor = AutoImageProcessor.from_pretrained(REPO_ID)
inputs = image_processor(image, return_tensors="pt")

session = ort.InferenceSession(model_path, providers=["CPUExecutionProvider"])
logits = session.run(
    ["logits"],
    {"image": inputs["pixel_values"].numpy()},
)[0][0]
probabilities = np.exp(logits - logits.max())
probabilities /= probabilities.sum()

prediction = int(probabilities.argmax())
print(LABELS[prediction], float(probabilities[prediction]))

Safetensors inference

From a checkout of the source repository:

from huggingface_hub import hf_hub_download
from safetensors.torch import load_file

from src.model import EfficientNetClassifier

weights_path = hf_hub_download(
    "electblake/clothing_type_classifier",
    "clothing_type-tf_efficientnetv2_s.in21k_ft_in1k.safetensors",
)
model = EfficientNetClassifier(
    model_name="tf_efficientnetv2_s.in21k_ft_in1k",
    num_classes=5,
    dropout=0.15,
    pretrained=False,
)
model.load_state_dict(load_file(weights_path))
model.eval()

Training data

The prepared dataset contains 16,496 person crops across the five labels.

Training used weighted cross-entropy, AdamW, MixUp, CutMix, horizontal flips, geometric transforms, brightness and contrast changes, and hue and saturation changes. The backbone was initially frozen and then fully fine-tuned.

Evaluation

The published checkpoint was selected at epoch 31 with 89.73% validation accuracy. This value comes from the checkpoint metadata and is a model-selection result on the project's validation split, not an independent benchmark or test-set result.

Intended use and limitations

The model is intended for research, media organization, and non-critical clothing-type tagging. It is not intended for decisions affecting a person's rights or access to services.

Performance is expected to vary with occlusion, multiple people, unusual framing, small subjects, lighting, image quality, and clothing outside the five published categories. The classifier always selects one of its five known labels and has no other or rejection class. Confidence scores are softmax probabilities and have not been calibrated.