Skip to content

Larq Zoo Pretrained Models

Larq Zoo provides reference implementations of deep neural networks with extremely low precision weights and activations that are made available alongside pre-trained weights. These models can be used for prediction, feature extraction, and fine-tuning.

The code for all models including a reproducible training pipeline is available at larq/zoo.

Larq Zoo consists of a literature and a sota submodule.

The literature submodule contains replications from research papers (all current models). These models are intended to provide a stable reference for ideas presented in specific papers. The model implementations will be maintained, but we will not attempt to improve these models over time by applying new training strategies or architecture innovations.

The sota submodule contains top models for various scenarios. These models are intended to be used in a SW 2.0-like fashion. We will do our best to continuously improve the models, which means that their weights and even details of their architectures may change from release to release.

If you have developed or reimplemented a Binarized or other Extremely Quantized Neural Network and want to share it with the community such that future papers can build on top of your work, please add it to Larq Zoo or get in touch with us if you need any help.

Larq Zoo is part of a family of libraries for BNN development; you can also check out Larq for building and training BNNs and Larq Compute Engine for optimized deployment.

Available models

The following models are trained on the ImageNet dataset. The Top-1 and Top-5 accuracy refers to the model's performance on the ImageNet validation dataset, memory refers to the memory after quantization of the weights. Models were benchmarked using Larq Compute Engine on a Pixel 1 phone (2016), single-threaded1.

The model definitions and the train loops are available in the Larq Zoo repository.

The sota submodule contains these models:

Model Top-1 Accuracy Top-5 Accuracy Model size Latency (Pixel 1, single thread)
QuickNet 58.6 % 81.0 % 3.18 MB 18.4 ms
QuickNetLarge 62.7 % 84.0 % 4.49 MB 27.6 ms
QuickNetXL 67.0 % 87.3 % 6.22 MB 47.9 ms

The literature submodule contains the following models:

Model Top-1 Accuracy Top-5 Accuracy Model size Latency (Pixel 1, single thread)
RealToBinaryNet 65.0 % 85.7 % 5.13 MB 51.3 ms
BinaryDenseNet45 64.6 % 85.2 % 7.35 MB 138.5 ms
BinaryDenseNet37Dilated 64.3 % 85.2 % 5.13 MB 182.9 ms
BinaryDenseNet37 62.9 % 84.2 % 5.13 MB 102.2 ms
MeliusNet22 62.4 % 83.9 % 3.88 MB 117.7 ms
BinaryDenseNet28 60.9 % 82.8 % 4.04 MB 90.0 ms
BinaryResNetE18 58.3 % 80.8 % 4.00 MB 43.6 ms
Bi-Real Net 57.5 % 79.8 % 4.00 MB 43.4 ms
DoReFaNet 53.4 % 76.5 % 22.80 MB Unsupported2
XNOR-Net 45.0 % 69.2 % 22.77 MB 34.9 ms
Binary AlexNet 36.3 % 61.5 % 7.45 MB 44.3 ms


Larq Zoo is not included in Larq by default. To start using it, you can install it with Python's pip package manager:

pip install larq-zoo

Weights can be downloaded automatically when instantiating a model. They are stored at ~/.larq/models/.

Training Models from Scratch

Larq Zoo ships with a command-line interface powered by zookeeper, allowing you to reproduce the entire training process. If you want to improve an existing model or implement your own, we recommend to installing Larq Zoo in development mode.

E.g. to reproduce the training of Binary AlexNet run:

lqz TrainBinaryAlexNet dataset=ImageNet

To experiment with different hyperparameters you can either edit the task for this model or overwrite them from the command line, e.g.:

lqz TrainBinaryAlexNet dataset=ImageNet epochs=100 batch_size=64

For all available commands and options run lqz --help or checkout the documentation of zookeeper if you want to implement your model for Larq Zoo.

  1. Benchmarked on July 20th, 2020. 

  2. DoReFaNet uses quantizers for which currently no optimized implemention is available in Larq Compute Engine.