Skip to content

Latest commit

 

History

History
117 lines (95 loc) · 2.74 KB

README.md

File metadata and controls

117 lines (95 loc) · 2.74 KB

MobileNetV3 TensorFlow

Unofficial implementation of MobileNetV3 architecture described in paper Searching for MobileNetV3. This repository contains small and large MobileNetV3 architecture implemented using TensforFlow with tf.keras API.

Google Colab

  • Open In Colab MNIST
  • Open In Colab CIFAR10

Requirements

  • Python 3.6+
  • TensorFlow 1.13+
pip install -r requirements.txt

Build model

MobileNetV3 Small

from mobilenetv3_factory import build_mobilenetv3
model = build_mobilenetv3(
    "small",
    input_shape=(224, 224, 3),
    num_classes=1001,
    width_multiplier=1.0,
)

MobileNetV3 Large

from mobilenetv3_factory import build_mobilenetv3
model = build_mobilenetv3(
    "large",
    input_shape=(224, 224, 3),
    num_classes=1001,
    width_multiplier=1.0,
)

Train

CIFAR10 dataset

python train.py \
    --model_type small \
    --width_multiplier 1.0 \
    --height 128 \
    --width 128 \
    --dataset cifar10 \
    --lr 0.01 \
    --optimizer rmsprop \
    --train_batch_size 256 \
    --valid_batch_size 256 \
    --num_epoch 10 \
    --logdir logdir

MNIST dataset

python train.py \
    --model_type small \
    --width_multiplier 1.0 \
    --height 128 \
    --width 128 \
    --dataset mnist \
    --lr 0.01 \
    --optimizer rmsprop \
    --train_batch_size 256 \
    --valid_batch_size 256 \
    --num_epoch 10 \
    --logdir logdir

Evaluate

CIFAR10 dataset

python evaluate.py \
    --model_type small \
    --width_multiplier 1.0 \
    --height 128 \
    --width 128 \
    --dataset cifar10 \
    --valid_batch_size 256 \
    --model_path mobilenetv3_small_cifar10_10.h5

MNIST dataset

python evaluate.py \
    --model_type small \
    --width_multiplier 1.0 \
    --height 128 \
    --width 128 \
    --dataset mnist \
    --valid_batch_size 256 \
    --model_path mobilenetv3_small_mnist_10.h5

TensorBoard

Graph, training and evaluaion metrics are saved to TensorBoard event file uder directory specified with --logdir` argument during training. You can launch TensorBoard using following command.

tensorboard --logdir logdir

License

Apache License 2.0