--- pipeline_tag: image-classification license: apache-2.0 base_model: timm/convnext_xlarge.fb_in22k library_name: zeromodels tags: - keras - zeromodels - image-classification - convnext - backbone - arxiv:2201.03545 - pytorch - jax - tf --- ## ***See [our collection](https://huggingface.co/collections/zeromodels/convnext-6a8eaf1426a17001e2adc554) for all versions of ConvNeXt.*** # Run ConvNeXt with Keras 3: JAX, PyTorch, or TensorFlow [![GitHub](https://img.shields.io/badge/GitHub-ZeroModels-black?logo=github)](https://github.com/IMvision12/ZeroModels) [![Docs](https://img.shields.io/badge/Docs-ConvNeXt-blue)](https://imvision12.github.io/ZeroModels/classification_backbones/) [![Collection](https://img.shields.io/badge/HF-ConvNeXt%20collection-yellow)](https://huggingface.co/collections/zeromodels/convnext-6a8eaf1426a17001e2adc554) # zeromodels/convnext_xlarge_fb_in22k Paper: [A ConvNet for the 2020s (arXiv:2201.03545)](https://arxiv.org/abs/2201.03545) · [HF Papers](https://huggingface.co/papers/2201.03545) ConvNeXt modernizes a ResNet-style CNN with ViT-inspired design choices. Use as ImageNet classifier or 4-stage backbone. For more details on the model, please go to the upstream [model card](https://huggingface.co/timm/convnext_xlarge.fb_in22k). Pure-**Keras 3** conversion of [`timm/convnext_xlarge.fb_in22k`](https://huggingface.co/timm/convnext_xlarge.fb_in22k) for [zeromodels](https://github.com/IMvision12/ZeroModels). One implementation runs unmodified on **TensorFlow / Torch / JAX**. This is an **image-classification / backbone** checkpoint (`ConvNeXtImageClassify` / `ConvNeXtModel`). ## ✨ Quick start ```python import os os.environ["KERAS_BACKEND"] = "torch" # or "jax" / "tensorflow" from PIL import Image import numpy as np from zeromodels.models.convnext import ConvNeXtImageClassify, ConvNeXtModel model = ConvNeXtImageClassify.from_weights("zeromodels/convnext_xlarge_fb_in22k") backbone = ConvNeXtModel.from_weights( "zeromodels/convnext_xlarge_fb_in22k", as_backbone=True ) image = Image.open("your_image.jpg").convert("RGB") image = image.resize((224, 224)) x = np.asarray(image, dtype="float32")[None] # (1, H, W, 3) print(model(x).shape) # (1, num_classes) feats = backbone(x) print(len(feats), [tuple(f.shape) for f in feats]) ``` Load any ConvNeXt variant the same way with `from_weights("zeromodels/")`: | Variant | Hub | |---|---| | `convnext_atto_d2_in1k` | [`zeromodels/convnext_atto_d2_in1k`](https://huggingface.co/zeromodels/convnext_atto_d2_in1k) | | `convnext_base_fb_in1k` | [`zeromodels/convnext_base_fb_in1k`](https://huggingface.co/zeromodels/convnext_base_fb_in1k) | | `convnext_base_fb_in22k` | [`zeromodels/convnext_base_fb_in22k`](https://huggingface.co/zeromodels/convnext_base_fb_in22k) | | `convnext_base_fb_in22k_ft_in1k` | [`zeromodels/convnext_base_fb_in22k_ft_in1k`](https://huggingface.co/zeromodels/convnext_base_fb_in22k_ft_in1k) | | `convnext_base_fb_in22k_ft_in1k_384` | [`zeromodels/convnext_base_fb_in22k_ft_in1k_384`](https://huggingface.co/zeromodels/convnext_base_fb_in22k_ft_in1k_384) | | `convnext_femto_d1_in1k` | [`zeromodels/convnext_femto_d1_in1k`](https://huggingface.co/zeromodels/convnext_femto_d1_in1k) | | `convnext_large_fb_in1k` | [`zeromodels/convnext_large_fb_in1k`](https://huggingface.co/zeromodels/convnext_large_fb_in1k) | | `convnext_large_fb_in22k` | [`zeromodels/convnext_large_fb_in22k`](https://huggingface.co/zeromodels/convnext_large_fb_in22k) | | `convnext_large_fb_in22k_ft_in1k` | [`zeromodels/convnext_large_fb_in22k_ft_in1k`](https://huggingface.co/zeromodels/convnext_large_fb_in22k_ft_in1k) | | `convnext_large_fb_in22k_ft_in1k_384` | [`zeromodels/convnext_large_fb_in22k_ft_in1k_384`](https://huggingface.co/zeromodels/convnext_large_fb_in22k_ft_in1k_384) | | `convnext_nano_d1h_in1k` | [`zeromodels/convnext_nano_d1h_in1k`](https://huggingface.co/zeromodels/convnext_nano_d1h_in1k) | | `convnext_nano_in12k_ft_in1k` | [`zeromodels/convnext_nano_in12k_ft_in1k`](https://huggingface.co/zeromodels/convnext_nano_in12k_ft_in1k) | | `convnext_pico_d1_in1k` | [`zeromodels/convnext_pico_d1_in1k`](https://huggingface.co/zeromodels/convnext_pico_d1_in1k) | | `convnext_small_fb_in1k` | [`zeromodels/convnext_small_fb_in1k`](https://huggingface.co/zeromodels/convnext_small_fb_in1k) | | `convnext_small_fb_in22k` | [`zeromodels/convnext_small_fb_in22k`](https://huggingface.co/zeromodels/convnext_small_fb_in22k) | | `convnext_small_fb_in22k_ft_in1k` | [`zeromodels/convnext_small_fb_in22k_ft_in1k`](https://huggingface.co/zeromodels/convnext_small_fb_in22k_ft_in1k) | | `convnext_small_fb_in22k_ft_in1k_384` | [`zeromodels/convnext_small_fb_in22k_ft_in1k_384`](https://huggingface.co/zeromodels/convnext_small_fb_in22k_ft_in1k_384) | | `convnext_tiny_fb_in1k` | [`zeromodels/convnext_tiny_fb_in1k`](https://huggingface.co/zeromodels/convnext_tiny_fb_in1k) | | `convnext_tiny_fb_in22k` | [`zeromodels/convnext_tiny_fb_in22k`](https://huggingface.co/zeromodels/convnext_tiny_fb_in22k) | | `convnext_tiny_fb_in22k_ft_in1k` | [`zeromodels/convnext_tiny_fb_in22k_ft_in1k`](https://huggingface.co/zeromodels/convnext_tiny_fb_in22k_ft_in1k) | | `convnext_tiny_fb_in22k_ft_in1k_384` | [`zeromodels/convnext_tiny_fb_in22k_ft_in1k_384`](https://huggingface.co/zeromodels/convnext_tiny_fb_in22k_ft_in1k_384) | | `convnext_xlarge_fb_in22k` | [`zeromodels/convnext_xlarge_fb_in22k`](https://huggingface.co/zeromodels/convnext_xlarge_fb_in22k) | | `convnext_xlarge_fb_in22k_ft_in1k` | [`zeromodels/convnext_xlarge_fb_in22k_ft_in1k`](https://huggingface.co/zeromodels/convnext_xlarge_fb_in22k_ft_in1k) | | `convnext_xlarge_fb_in22k_ft_in1k_384` | [`zeromodels/convnext_xlarge_fb_in22k_ft_in1k_384`](https://huggingface.co/zeromodels/convnext_xlarge_fb_in22k_ft_in1k_384) | ## Tips - Set `KERAS_BACKEND` **before** importing Keras / zeromodels. - `ConvNeXtImageClassify` returns class logits; `ConvNeXtModel` returns features (`as_backbone=True` for multi-scale stages). - See [docs](https://imvision12.github.io/ZeroModels/classification_backbones/) and [Loading Weights](https://imvision12.github.io/ZeroModels/loading_weights/). - Upstream / timm checkpoints: `ConvNeXtImageClassify.from_weights("hf:timm/convnext_xlarge.fb_in22k")`. ## Special Thanks A huge thank you to the ConvNeXt authors and the timm / Hub communities for creating and releasing these models. License: see YAML `license` (usually matches the upstream checkpoint).