dtd-convnext-xxl

ConvNeXt-XXL Fine-Tuned on DTD

This repository provides a ConvNeXt-XXL convolutional neural network fine-tuned on the Describable Textures Dataset (DTD) for image classification.

Citation

If you use this model in your work, please cite it as:

@misc{vishnu_chityala_2025,
  author       = {Vishnu Chityala},
  title        = {dtd-convnext-xxl (Revision b9ec966)},
  year         = 2025,
  url          = {https://huggingface.co/vishnurchityala/dtd-convnext-xxl},
  doi          = {10.57967/hf/6417},
  publisher    = {Hugging Face}
}

---

Training Results

Accuracy Curves

• ConvNeXt-XXL (fine-tuned, best accuracy = 80.59%)
  

Final Evaluation Metrics

• Accuracy: 80.6%
• Loss: ~0.687

---

Files

• convnext_dtd_acc_80.59.pth — fine-tuned model checkpoint (PyTorch)
• training_plot_acc_80.59.png — accuracy curves for ConvNeXt fine-tuning
• README.md — model card and documentation

---

Usage

Load the model checkpoint with PyTorch and run inference:

import torch
from torch import nn
from PIL import Image
from torchvision import transforms

# Define your model class
class ConvNeXtClassifier(nn.Module):
    def __init__(self, num_classes=47, pretrained=True, mlp_hidden=None, 
                 dropout=0.35, freeze_backbone=True):
        super().__init__()
        import timm
        if mlp_hidden is None:
            mlp_hidden = [1024, 512, 256]
        self.backbone = timm.create_model("convnext_xxlarge", pretrained=pretrained)
        in_features = self.backbone.head.fc.in_features
        self.backbone.head.fc = nn.Identity()
        layers = []
        input_dim = in_features
        for h in mlp_hidden:
            layers += [nn.Linear(input_dim, h), nn.ReLU(inplace=True), nn.Dropout(dropout)]
            input_dim = h
        layers.append(nn.Linear(input_dim, num_classes))
        self.classifier = nn.Sequential(*layers)
        if freeze_backbone:
            for param in self.backbone.stem.parameters():
                param.requires_grad = False
            for param in self.backbone.stages.parameters():
                param.requires_grad = False

    def forward(self, x):
        x = self.backbone(x)
        x = self.classifier(x)
        return x

# Load model
model = ConvNeXtClassifier(pretrained=False)
model.load_state_dict(torch.load("convnext_dtd_acc_80.59.pth", map_location="cpu"))
model.eval()

# Example inference
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
])

image = Image.open("example_image.jpg").convert("RGB")
input_tensor = transform(image).unsqueeze(0)  # Add batch dimension

with torch.no_grad():
    output = model(input_tensor)
    predicted_class = output.argmax(dim=1).item()

print(f"Predicted class: {predicted_class}")