zeromodels/efficientformer_l3_snap_dist_in1k

See our collection for all versions of EfficientFormer.

Run EfficientFormer with Keras 3: JAX, PyTorch, or TensorFlow

zeromodels/efficientformer_l3_snap_dist_in1k

Paper: EfficientFormer: Vision Transformers at MobileNet Speed (arXiv:2206.01191) · HF Papers

EfficientFormer redesigns transformers for MobileNet-level latency on mobile devices. Classifier or 4-stage backbone.

For more details on the model, please go to the upstream model card.

Pure-Keras 3 conversion of timm/efficientformer_l3.snap_dist_in1k for zeromodels. One implementation runs unmodified on TensorFlow / Torch / JAX.

This is an image-classification / backbone checkpoint (EfficientFormerImageClassify / EfficientFormerModel).

✨ Quick start

import os

os.environ["KERAS_BACKEND"] = "torch"  # or "jax" / "tensorflow"

from PIL import Image
from zeromodels.models.efficientformer import EfficientFormerImageClassify, EfficientFormerModel, EfficientFormerImageProcessor

model = EfficientFormerImageClassify.from_weights("zeromodels/efficientformer_l3_snap_dist_in1k")
processor = EfficientFormerImageProcessor.from_weights("zeromodels/efficientformer_l3_snap_dist_in1k")

image = Image.open("your_image.jpg").convert("RGB")
pixels = processor(image)  # resize + normalize (normalization lives in the processor)
logits = model(pixels, training=False)
print(logits.shape)  # (1, num_classes)

# Feature extraction: the backbone without the classifier head
backbone = EfficientFormerModel.from_weights("zeromodels/efficientformer_l3_snap_dist_in1k", as_backbone=True)
features = backbone(pixels, training=False)

Load any EfficientFormer variant the same way with from_weights("zeromodels/<variant>"):

Variant	Hub
efficientformer_l1_snap_dist_in1k	zeromodels/efficientformer_l1_snap_dist_in1k
efficientformer_l3_snap_dist_in1k	zeromodels/efficientformer_l3_snap_dist_in1k
efficientformer_l7_snap_dist_in1k	zeromodels/efficientformer_l7_snap_dist_in1k

Tips

• Set KERAS_BACKEND before importing Keras / zeromodels.
• EfficientFormerImageClassify returns class logits; EfficientFormerModel returns features (as_backbone=True for multi-scale stages).
• See docs and Loading Weights.
• Upstream / timm checkpoints: EfficientFormerImageClassify.from_weights("hf:timm/efficientformer_l3.snap_dist_in1k").

Special Thanks

A huge thank you to the EfficientFormer authors and the timm / Hub communities for creating and releasing these models.

License: see YAML license (usually matches the upstream checkpoint).