See our collection for all versions of EfficientNet.

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

zeromodels/tf_efficientnet_b2_aa_in1k

Paper: EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks (arXiv:1905.11946) · HF Papers

EfficientNet compound-scales depth/width/resolution for strong accuracy/efficiency. Classifier or multi-scale MBConv backbone.

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

Pure-Keras 3 conversion of timm/tf_efficientnet_b2.aa_in1k for zeromodels. One implementation runs unmodified on TensorFlow / Torch / JAX.

This is an image-classification / backbone checkpoint (EfficientNetImageClassify / EfficientNetModel).

✨ Quick start

import os

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

from PIL import Image
from zeromodels.models.efficientnet import EfficientNetImageClassify, EfficientNetModel, EfficientNetImageProcessor

model = EfficientNetImageClassify.from_weights("zeromodels/tf_efficientnet_b2_aa_in1k")
processor = EfficientNetImageProcessor.from_weights("zeromodels/tf_efficientnet_b2_aa_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 = EfficientNetModel.from_weights("zeromodels/tf_efficientnet_b2_aa_in1k", as_backbone=True)
features = backbone(pixels, training=False)

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

Variant	Hub
`tf_efficientnet_b0_aa_in1k`	`zeromodels/tf_efficientnet_b0_aa_in1k`
`tf_efficientnet_b0_ap_in1k`	`zeromodels/tf_efficientnet_b0_ap_in1k`
`tf_efficientnet_b0_in1k`	`zeromodels/tf_efficientnet_b0_in1k`
`tf_efficientnet_b0_ns_jft_in1k`	`zeromodels/tf_efficientnet_b0_ns_jft_in1k`
`tf_efficientnet_b1_aa_in1k`	`zeromodels/tf_efficientnet_b1_aa_in1k`
`tf_efficientnet_b1_ap_in1k`	`zeromodels/tf_efficientnet_b1_ap_in1k`
`tf_efficientnet_b1_in1k`	`zeromodels/tf_efficientnet_b1_in1k`
`tf_efficientnet_b1_ns_jft_in1k`	`zeromodels/tf_efficientnet_b1_ns_jft_in1k`
`tf_efficientnet_b2_aa_in1k`	`zeromodels/tf_efficientnet_b2_aa_in1k`
`tf_efficientnet_b2_ap_in1k`	`zeromodels/tf_efficientnet_b2_ap_in1k`
`tf_efficientnet_b2_in1k`	`zeromodels/tf_efficientnet_b2_in1k`
`tf_efficientnet_b2_ns_jft_in1k`	`zeromodels/tf_efficientnet_b2_ns_jft_in1k`
`tf_efficientnet_b3_aa_in1k`	`zeromodels/tf_efficientnet_b3_aa_in1k`
`tf_efficientnet_b3_ap_in1k`	`zeromodels/tf_efficientnet_b3_ap_in1k`
`tf_efficientnet_b3_in1k`	`zeromodels/tf_efficientnet_b3_in1k`
`tf_efficientnet_b3_ns_jft_in1k`	`zeromodels/tf_efficientnet_b3_ns_jft_in1k`
`tf_efficientnet_b4_aa_in1k`	`zeromodels/tf_efficientnet_b4_aa_in1k`
`tf_efficientnet_b4_ap_in1k`	`zeromodels/tf_efficientnet_b4_ap_in1k`
`tf_efficientnet_b4_in1k`	`zeromodels/tf_efficientnet_b4_in1k`
`tf_efficientnet_b4_ns_jft_in1k`	`zeromodels/tf_efficientnet_b4_ns_jft_in1k`
`tf_efficientnet_b5_aa_in1k`	`zeromodels/tf_efficientnet_b5_aa_in1k`
`tf_efficientnet_b5_ap_in1k`	`zeromodels/tf_efficientnet_b5_ap_in1k`
`tf_efficientnet_b5_in1k`	`zeromodels/tf_efficientnet_b5_in1k`
`tf_efficientnet_b5_ns_jft_in1k`	`zeromodels/tf_efficientnet_b5_ns_jft_in1k`
`tf_efficientnet_b6_aa_in1k`	`zeromodels/tf_efficientnet_b6_aa_in1k`
`tf_efficientnet_b6_ap_in1k`	`zeromodels/tf_efficientnet_b6_ap_in1k`
`tf_efficientnet_b6_ns_jft_in1k`	`zeromodels/tf_efficientnet_b6_ns_jft_in1k`
`tf_efficientnet_b7_aa_in1k`	`zeromodels/tf_efficientnet_b7_aa_in1k`
`tf_efficientnet_b7_ap_in1k`	`zeromodels/tf_efficientnet_b7_ap_in1k`
`tf_efficientnet_b7_ns_jft_in1k`	`zeromodels/tf_efficientnet_b7_ns_jft_in1k`
`tf_efficientnet_b8_ap_in1k`	`zeromodels/tf_efficientnet_b8_ap_in1k`
`tf_efficientnet_l2_ns_jft_in1k`	`zeromodels/tf_efficientnet_l2_ns_jft_in1k`
`tf_efficientnet_l2_ns_jft_in1k_475`	`zeromodels/tf_efficientnet_l2_ns_jft_in1k_475`

Tips

Set KERAS_BACKEND before importing Keras / zeromodels.
EfficientNetImageClassify returns class logits; EfficientNetModel returns features (as_backbone=True for multi-scale stages).
See docs and Loading Weights.
Upstream / timm checkpoints: EfficientNetImageClassify.from_weights("hf:timm/tf_efficientnet_b2.aa_in1k").