MobileNetV3 TensorFlow

Unofficial implementation of MobileNetV3 architecture described in paper Searching for MobileNetV3. This repository follows mobilenetv3-tensorflow.

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=3,    width_multiplier=1.0,
)

MobileNetV3 Large

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

Train

Printer dataset

python train.py 
    --model_type small 
    --width_multiplier 1.0 
    --height 224 
    --width 224 
    --dataset printer 
    --lr 0.001 
    --optimizer sgd 
    --train_batch_size 10 
    --valid_batch_size 10 
    --num_epoch 30 
    --logdir logdir

Test

printer dataset

python test.py 
    --model_path mobilenetv3_small_printer_30.h5 
    --model_type small 
    --image_path test/1.jpg 
    --featuremap False 
    --heatmap False

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